PROCESSING OF

REMOTE SENSING DATA

Processing of Remote
Sensing Data

MICHEL-CLAUDE GIRARD
Professor of Soil Science at the Paris-Grignon INA

COLETTE GIRARD
Professor emeritus of Geobotany at the Paris-Grignon INA

In collaboration with
Dominique Courault
Jean-Marc Gilliot
Lionel Loubersac
Jean Meyer-Roux
Jean-Marie Monget
Bernard Seguin

Translated by
N. Venkat Rao
Professor of Geophysics
Osmania University
Hyderabad, India

A.A. BALKEMA PUBLISHERS L isse / A b in g d o n / E xto n (PA) / T o k y o

Published by arrangement with Dunod, Paris

Ouvrage publié avec le concours du Ministère français chargé de la Culture—Centre

national du livre
Published with the support of the French Ministry of Culture

Translation of: Traitement des données de télédétection

ISBN 2 10 004185 1 © 1999 Dunod, Paris

Copyright Reserved © 2003

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or
transmitted In any form or by any means, electronic, mechanical, photocopying, recording, or other­
wise, without the prior written permission of the publisher.

Published by: A.A. Balkema, a member of Swets & Zeitlinger Publishers

www.balkema.nl and www.szp.swets.nl

ISBN 90 5809 232 1

 Preface
‘Who sees from a height, sees well; who sees from far, sees correctly’
Victor Hugo, The art of being a grandfather, satisfied exile, 1877

‘I can make a circle around the Earth in twenty minutes’

William Shakespeare, A midsummer night’s dream, 1595

After the ‘Course on Photo-interpretation’, published by ENSA of Grignon in 1970, and the books
‘Applications of remote sensing in the biosphere (1975) and ‘Applied remote sensing—Temperate and
intertropical zones (1989)’ published by Masson, this book is one of the most important to be released
by Dunod this year. It covers a part of the preceding books, which are out of print; however, its novelty
lies in the fact that it describes all the methods of digital data processing of remote sensing images.
In fact, today, image processing can be carried out on personal computers, which are accessible
to all professionals, researchers and students, and the large number of existing software programs
provides a wide choice. We have opted for one of the commercial softwares. Although it is not the most
widely known, it is mathematically most explicit and gives many details about the quality of classifications.
It also enables analyses of reflectance and digital computations, and is less expensive. Only a part of
the possibilities of the TeraVue program is given in the CD supplied with this book.
Starting with a simplified version of this program, examples of SPOT and LANDSATTM images
are given, showing various phases of the methods, which constitute the justification for a new book.
. For ail the methods described In this book, corresponding images are given in the CD, which can be
analysed on personal computers.
For simplification, the given examples of data processing deal with one SPOT image of the Brienne
region between the humid and chalky zones of Champagne. We have had the occasion to work with
our friend F. Dumas, Director of the Chamber of Agriculture of Aube for studying the erosion-prone
zones, among others.
A number of illustrations related to the various chapters, which can be analysed by the same
program, are also given in the CD. Thus, the conventional colour illustrations of the remote sensing
books are replaced by digital images In this book.
Obviously, it Is assumed that the reader has a PC at his disposal, as it is a common practice to
have one these days. If not, it is expected that all readers will have access to one while studying this
book.
Remote sensing is a multispectral tool; for studying any problem, it is often very useful to have
data In different bands of the electromagnetic spectrum, viz., visible, near and middle reflective infrared,
thermal infrared and microwave.
This book is mainly concerned with the processing of data acquired In visible and near- and
middle-infrared bands. In fact, these domains are the most often used, since the corresponding satellite
images can be readily procured. Also, the methods of visual interpretation of Images and aerial
vi Processing of Remote Sensing Data

photographs have been updated, as they are still very useful.They are, perhaps, forgotten now, although
they constituted the basis for any interpretation about twenty years ago. It was good to recollect them
and adapt them to new remote sensing data.
As the data cannot be processed by various models underlying each method without validating
the model, the theoretical and practical bases of estimating the quality of the results obtained have
been reviewed.
In order to safeguard the main objective of image processing, viz., answering a relevant thematic
question, a number of concrete and real examples of data processing of remote sensing are included
at the end of the book.
As we did not wish to make this book a bibliographic collection on remote sensing and listing ail
the references would have required ten pages at the end of every chapter, these are restricted to the
absolute minimum.The references cited constitute only guides to those that can be found In periodicals
whose names are given in the appendix. The fact that a book is not cited does not Indicate lack of
Interest. Such books include those authored by MM. Flouzat, Wllmet and a number of others.
This book is also addressed to teachers of secondary schools who have introduced In the first,
second and final programs multidisciplinary topics on remote sensing. It ought to help them in presenting
specific cases and in preparing the students for applied studies in mathematics, physics, geography
or biology.
The Images given In the CD are due to kind courtesy as acknowledged below:
— ^The two Thematic Mapper scenes of LANDSAT-5 acquired on 1 April 1990 (scene 197/26) and
15 May 1992 (scene 198/26), by ESA: @ ESA (1990 and 1992), acquired by Fuclno Station,
Distributed by Eurimage-Geosys, by courtesy of Eurimage);
— ^The image of Aube taken on 16 September 1996, by SPOT-lmage:©CNES 1996, Distribution
SPOT-lmage.
 Acknowledgements
We sincerely thank various colleagues with whom we worked and who have provided information on
different aspects of remote sensing:
M.J.-M. Gilliot (INA PG) for studies on filters and geometric corrections (Chapters 12 and 13);
M J. Meyer-Roux (CCR of ISPRA, European Union) for estimation of yields (Chapter 22);
Mrs. D. Courault, Ch. King, E. Vaudour, MM. R. Escadafal, B. Mougenot, C.Yongchalermchai, for
the numerous studies on soils conducted in our laboratory of INA PG, most of which were borrowed
for this book (Chapter 23);
M.J.-M. Monget (National Higher School of Mines, Paris) for applications in mineral prospecting
Chapter 24);
M .L Loubersac (IFREMER) for coastal-zone applications (Chapter 25);
Mrs. D Courault (INRA) and M. B. Seguin (INRA) for thermal-imagery applications (Chapter 26)
and Mrs. Ch. King for her data on microwave remote sensing.
We thank P. Bertrand, F. Burlot, T. Francoual and S. Mollet who, during the preparation of their
theses, have contributed to experimentations on various methods developed in our laboratory as part
of their Master’s program in ‘Localised Information Systems for land-use management’.
Several parts of this book have been taken from our own investigations, published or not, as well
as those of our research unit on ‘Dynamics of Environment and Spatial Organisation’ (Department of
Agronomy-Environment) of the Paris-Grignon National Institute of Agronomy: Mmes. Ch. King, D.
Courault and D. Orth, E. Vaudour and MM. J.-R Rogala, D. King, R Escadafal, B. Mougenot, J.-M.
Gilliot, C.Yongchalermchai, G. Belluzo who, although working in various organisations (INA PG, BRGM,
ORSTOM, INRA, APIC-System), have constituted a scientific team for the last twenty years studying
applications of remote sensing to soil and vegetation. We also thank P. Boissard, M. Bornand, J.-P.
Lagouarde, B. Seguin (INRA) who since long have been associated with our studies on remote sensing.
We have taken several figures from MM. D’Allemand, Becker, Guyot, Lliboutry, Perrin de
Brichambaut, Vossen and others, which we have modified to our requirement. Thanks are also due to
P. Guillore (t) and J.-C. Carle for a significant number of drawings used in the book. We are thankful to
D. Lepoutre and GeoSYS who have permitted us to use a digital elevation model for Brienne region.
We cannot forget that remote sensing studies were initiated by M.R. Chevalller, M. Guy and M.
Jean Boulaine, who have provided the results of their research on the use of aerial photos in the
Agricultural Academy of France. Our acknowledgements are also due to M. G. Brachet for his
encouragement to continue our studies; to M.B. Cabrieres, the incharge of SPOT-4 system at CNES,
for providing data on SPOT systems (a major part of which is given in the CD-ROM) and for carefully
reading the material on this subject; to Mme. Lecochennec of SPOT-lmage for reading the part
concerning SPOT; M. Egels of IGN who has read Chapter 14 on fundamentals of photogrammetry; M.
P. Fasquel for extending wise counsel on geometric quality of data (Chapter 17).
Practice of remote sensing techniques in administrative and consultancy agencies and various
professional sectors has enabled us to apply our methods in real situations. We thank all those who
have shared this challenge with us: M. V. Le Dolley who, as the district director of agriculture, encouraged
viii Processing of Remote Sensing Data

processing of images of Yonne district; MM. F. Dumas and V. Ellisseeff of the Aube Chamber of
Agriculture, M. F. Limaux of Lorraine Regional Chamber of Agriculture, M.J.-M. Vinatier and Mrs I.
Boutefoy of the Soil Info Rhone Alps (SIRA), M. P. Juillet de Saint-Lager, Regional Director of
Environment of Champagne-Ardenne, M. L. Lurton of the Inter-professional Committee on Vineyards
of Cotes-du-Rhone and Rhone Valley.
In this book, we also address the secondary school teachers who have introduced multidisciplinary
courses on remote sensing in their primary, secondary and final school programs. We had the occasion
to work In this experiment due to some teachers, viz., Ch. Guisti, M. Dupuis and A. Herpe of La Queue
les-yvelines high school and professors Monchamp and VIgneron of Plalsir high school. We have
often thought of them and their students while preparing this book: it should provide concrete base for
preparing to applied studies In mathematics, physics, geography or biology. We also wish that this
book forms a useful tool for all other secondary teachers who want to engage themselves in this
marvellous multidisciplinary adventure extending across processing and interpretation of images.
We are especially grateful to Lucien Bugeat who nominated one of us to the Remote Sensing
Committee of the Agricultural Ministry and to Jean Dunglass and Maurice Dehegere who subsequently
presided over it. We were thus informed of the several attempts in the agricultural domain and the
numerous applications of remote sensing that were already operational.
Our thanks are also due to Mrs. S. Meriaux, secretary of section VII, our colleagues of section VII
and the permanent secretary A. Cauderon of Agricultural Academy of France for organising meetings
on remote sensing. Wish several other such meetings followed.
We are thankful to the national program of spatial remote sensing and CNES (stimulated action
specific to use of SPOT Images, headed by F. Blasco), Ministry of Agriculture and Fisheries (directorate
of rural areas and forest, bureau of soils, crop inventory program, management and conservation of
soils), IBM and INRA, who enabled progress of these studies through their scientific and financial
support.
Lastly, we wish to acknowledge the person associated with the pedagogic aspect of this book,
i.e., introduction of a CD-ROM as the heart of book. M. J.-M. Monget, our colleague at the National
Higher School of Mines, developed the concepts and wrote the TeraVue programs. We can thank him
only by assuring that our joint endeavour will continue through the proposed exciting program of
establishing information exchange between the European Union and the United States.
Our thanks to J.-M. Monget cannot be complete without simultaneously mentioning Mrs. Druel,
the incharge of La Boyere publications, who was responsible for the publication of the TeraVue image
processing software and the production of CD-ROM.
 Contents

Data S ources

1. Physical Basis 3
1.1 Radiation 3
1.2 Source and Sensor Parameters 5
1.3 Reflectance Factor 8
1.4 Solar Radiation and Atmospheric Perturbations 9
1.4.1 Atmospheric absorption 9
1.4.2 Atmospheric scattering 11
1.4.3 Atmospheric radiation 11
1.5 Thermal Infrared Remote Sensing 12
1.5.1 Physical basis 12
1.5.2 Correction for atmospheric radiation 13
1.5.3 Relationship between radiant temperature and aerodynamic temperature 16
1.5.4 Relationship between surface temperature and évapotranspiration 18
1.6 Principles of Microwave Remote Sensing 19
1.6.1 Special laws at microwave frequencies 20
1.6.2 Frequency bands 21
1.6.3 Polarisation 21
1.6.4 Doppler effect 21
1.6.5 Backscatter signal 22
1.6.6 Radar equation 22
1.6.7 Logistic parameters 22
1.6.8 Analysis of physical processes of backscatter 25
1.7 General Conclusion 31
References 31

2. Sensors and Platforms 33

2.1 Sensors 33
2.1.1 General scheme 34
2.1.2 Radiation receiving systems 35
2.1.3 Spectral bands 37
2.1.4 Airborne systems 37
2.2 Platforms 38
2.2.1 General principles of orbital motion 38
2.2.2 Orbits 40
 Processing of Remote Sensing Data

2.2.3 SPOT system 42

2.3 Other Systems 45
2.3.1 Meteosat 45
2.3.2 NOAA 45
2.3.3 Thermal sensor systems 47
2.3.4 LANDSAT 48
2.3.5 ERS-1 and 2 50
2.3.6 RADARSAT 50
2.3.7 JERS 50
2.3.8 Evolution of sensors 50
2.3.9 Satellite photography 53
2.4 Conclusion 53
References 54

B
Physical interpretation of Data

C om position o f Coiours 57
3.1 The Human Eye and Colour 57
3.1.1 Vision 57
3.1.2 Sensitivity 58
3.1.3 Contrasts 58
3.2 Red-Green-Blue System 58
3.3 Cubic Representation 60
3.3.1 Additive system 60
3.3.2 Subtractive system 60
3.4 Triangle of the International Commission on Illumination 61
3.5 Munsell System 62
3.6 Metamerism 63
3.7 Colour In Photography 65
3.7.1 Principle of emulsions 65
3.7.2 Panchromatic and infrared 65
3.7.3 Colour 66
3.7.4 Colour infrared 67
3.8 Treatment of Colours on Colour Screen 68
3.8.1 Colours on screen 68
3.8.2 Colour display 68
3.9 Use of Colours in Image Processing 69
3.9.1 Colour code of an Image (8 bits) 69
3.9.2 Interpretation of a 3-by-8 bits code 69
3.9.3 Choice of colours 70
3.9.4 Colour printing of documents 71
3.10 Conclusion 71
References 71

4. Spectral C haracteristics 72
4.1 Vegetation 72
4.1.1 Laboratory measurements 73
 Contents xi

4.1.2 Field measurements 79

4.2 Soils 82
4.3 Vegetation Indices 84
4.3.1 Identification of vegetation cover 84
4.3.2 Estimation of vegetation parameters 85
4.3.3 Use of ‘soil clusters’ 86
4.3.4 Various vegetation indices 86
4.4 Water 88
4.4.1 Response in visible and near- and middle-infrared bands 88
4.4.2 Approximate method of bathymetry in clear water 88
4.4.3 Measurement of water colour 90
4.5 Snow and Ice 91
References 91

Processing and Interpretation

Visual Interpretation o f Photographs and Images 95

5.1 Visual Interpretation 95
5.1.1 The eye 95
5.1.2 The brain 95
5.1.3 Interpretation procedure 97
5.2 Photo-Interpretation 97
5.2.1 Lines and points 97
5.2.2 Closed areas 103
5.2.3 Identification of thematic objects 105
5.2.4 Vinicultural “terroirs” 106
5.2.5 Conclusion 107
5.3 Visual Interpretation of Satellite Images 109
5.3.1 Interpretation of photographic prints 109
5.3.2 Interpretation on computer monitor 109
5.3.3 Stereoscopic vision with satellite images 110
5.4 Conclusion 111
References 112

6. Image Processing— General Features 113

6.1 Introduction 113
6.2 Image processing methods 114
6.2.1 Texture 114
6.2.2 Structure 115
6.3 Classification 115
6.3.1 Multiple languages 115
6.3.2 Segmentation and classification 115
6.3.3 Classification and grading 116
6.3.4 Ascendant and descendant methods 116
6.3.5 Concept of mathematical distance 117
6.3.6 Utilisation of distances for classification 118
xii Processing of Remote Sensing Data

6.4 Conclusion 120

7. Preliminary Processing 121

7.1 Processing of Single-band Data 121
7.1.1 Histogram analysis 121
7.1.2 Transformations of digital numbers 123
7.2 Processing of Multispectral Data 127
7.2.1 Comparison between colours and spectral characteristics of objects 127
7.2.2 Choice of band combinations 128
7.2.3 Two-dimensional or three-dimensional histograms 130
7.2.4 Segmentation of an histogram 133
7.2.5 Arithmetic combination of bands 133
7.2.6 Statistical analysis of bands 136
7.2.7 Principal component analysis 137
7.3 Masks 141
7.3.1 Radiometric masking 141
7.3.2 Geographic masking 142
7.3.3 Logical masking 142
7.4 Conclusion 143

8. Unsupervised Classification 144

8.1 Ascendant Hierarchic Classification 144
8.1.1 Principles 144
8.1.2 Groups and legend 145
8.2 Example: Image of Brienne Region 148
8.2.1 Statistical interpretation 148
8.2.2 Digital interpretation 150
8.2.3 Spatial interpretation 153
8.2.4 General Interpretation 156
8.2.5 Classification quality assessment 159
8.2.6 Response to identified objectives 159
8.2.7 Conclusion 161
8.3 Classification by Mobile Centres (or “K-means”) 162
8.3.1 Principle 162
8.3.2 Method of interpretation 163
8.3.3 Comparison with ascendant hierarchic classification 165
8.3.4 Conclusion 166

9. Supervised Classification 167

9.1 Parallelepiped Classification 167
9.1.1 Segmentation of radiometric scatter diagram 167
9.1.2 Single-band segmentation 168
9.1.3 Multispectral segmentation 170
9.1.4 Chorological segmentation 172
9.1.5 Spectral characteristics of various groups 172
9.1.6 Statistics of various groups 174
9.1.7 Visual interpretation 174
9.1.8 Quality of parallelepiped classification 175
9.1.9 Conclusion 175
 Contents xiii

9.2 Maximum-likelihood Classification 176

9.2.1 Probabilistic spectral behaviour 176
9.2.2 Principles of classification 176
9.2.3 Rejection threshold 177
9.2.4 Classification operations 177
9.2.5 Iterations: a heuristic approach 185
9.2.6 Classification quality assessment 185
9.2.7 Conclusion 186

Image Processing M ethodology 187

10.1 Objectives 187

10.2 Method 187
10.2.1 Input 187
10.2.2 Processing 188
10.2.3 Output 188
10.3 Procedure 189
10.3.1 Initialisation 189
10.3.2 Correlation 191
10.3.3 Verification 191
10.3.4 Modelling 191
10.4 Interpretation of Processing 191
10.4.1 Radiometric approach 191
10.4.2 Chorological approach 191
References 192

Structural Processing o f Sateiiite Images 193

11.1 Introduction 193
11.1.1 Boundaries and mixels 193
11.1.2 Classification and mapping 193
11.1.3 Complex map units 194
11.2 VOISIN 194
11.2.1 Neighbourhood in window: composition vector 194
11.2.2 Method 195
11.2.3 Example 196
11.3 OASIS 198
11.3.1 Method 198
11.3.2 Example 199
References 206

12. Digital Filtering o f Images 208

12.1 Linear Filtering 208

12.1.1 Image as a two-dimensional signal 208
12.1.2 Signal-processing systems: convolution product 212
12.1.3 Spatial filtering 214
12.2 Non-linear Filtering 216
12.2.1 Order Filtering: median 216
12.2.2 Morphological filtering 217
12.2.3 Homomorphic filtering 219
12.2.4 Adaptive filtering 220
xiv Processing of Remote Sensing Data

12.3 Noise Reduction in Image Preprocessing 221

12.3.1 Noise in images 221
12.3.2 Linear and non-linear filters for noise reduction 221
12.4 Edge Detection 222
12.4.1 Model of an edge 223
12.4.2 Edge detection by differential filters 224
12.4.3 Edge detection by other filters 226
12.4.4 Closure of edges 228
12.4.5 Other edge detection methods 228
12.4.6 Evaluation of edge detection operators 231
12.5 Conclusion 233
References 233

Geometric Transformation of Remote-sensing Images 234

13.1 Methods of Geometric Correction 234

13.1.1 Causes of geometric deformation 234
13.1.2 Parametric and interpolation methods 234
13.1.3 Global or local modelling of Interpolation methods 235
13.1.4 Direct and inverse transformations 235
13.1.5 Interpolation In geometric transformation 235
13.2 Interpolation Methods 236
13.2.1 Radiometric Interpolation 236
13.2.2 Geometric interpolation 237
13.2.3 Conclusion 245
13.3 Image Rectification 245
13.3.1 General problem 245
13.3.2 Systems and data for rectification 246
13.3.3 Formulation of rectification model 249
13.3.4 Conclusion 253
References 253

Fundamentals of Aerial Photography 255

14.1 Photo-acquisition 255

14.1.1 Photographic sensors 255
14.1.2 Types of aerial photos 258
14.1.3 Regular missions 261
14.1.4 Organisation and overlap of aerial photos 262
14.1.5 Flight plan (example) 264
14.2 Stereoscopy 264
14.2.1 General 264
14.2.2 Stereoscopic vision 266
14.2.3 Stereoscopes 268
14.2.4 Anaglyphs and vertographs 269
14.3 Photogrammetry 269
14.3.1 Orientation of photographs 270
14.3.2 Use of a stereo pair 270
14.3.3 Scale of a photograph 271
14.3.4 Aerial photo mosaics 273
14.3.5 Radial distortions 274
 Contents xv

14.3.6 Parallax and altitude determination 275

14.3.7 Stereo restitution— orthophotos 276
14.3.8 Historical background of aerial photography 277
14.4 Conclusion 278
References 279

Quality Assessment

15. Scale Changes 283

15.1 Introduction 283
15.2 Scale and Organisational Level of Media 283
15.3 Specific Aspects of Remote Sensing 285
15.3.1 Resolution 285
15.3.2 Adequacy between level of observation and level of organisation:
Image segmentation 286
15.4 Resolution of Pixel into Constituents or Descendant Approach 287
15.5 Aggregation of Pixels or Ascendant Approach 291
References 292

16. Criteria of Choice for the User 293

16.1 What Data to Choose? 293
16.1.1 What Is the question posed and the problem to be solved? 293
16.1.2 Are remote-sensing data apt to answer the problem? 293
16.1.3 What are the data acquisition conditions most likely to provide
an answer? 293
16.1.4 Are remote-sensing data responding to the conditions defined and
problem posed available? 294
16.2 Criteria for Choice of Remote-sensing data 294
16.2.1 Choice of spectral and geometric resolutions 294
16.2.2 Choice of date of acquisition 294
16.2.3 Choice of bands 295
16.3 Choice of Type of Processing 297
16.3.1 Visual or computer interpretation of data 297
16.3.2 Supervision of digital classification methods 297
16.4 Conclusion 299

17. Quality of Interpretation 300

17.1 Introduction 300
17.2 Geometric Accuracy 301
17.2.1 Precision of position 301
17.2.2 Precision of shape 302
17.2.3 Reliability 302
17.3 Semantic Accuracy 302
17.3.1 Definition 302
17.3.2 Establishment of typology 303
17.3.3 Establishment of error matrix 304
xvi Processing of Remote Sensing Data

17.3.4 Discussion of error matrix 306

17.3.5 Limitations of conventional methods of accuracy assessment 308
17.4 Conclusion 309
References 309

Applications

18. Agrolandscapes 313

18.1 Landscape Interpretation 313
18.1.1 Concept of landscape 313
18.1.2 Landscape: synthetic descriptor in remote sensing 313
18.1.3 Agrolandscape 314
18.2 Landscape Analysis 314
18.2.1 Different points of vision 315
18.2.2 Principal components of landscape 316
18.2.3 Descriptive format and statistics 320
18.3 Example: Agrolandscape of Yonne District 321
18.3.1 Method of description 321
18.3.2 Interpretation procedure: delineationand description of map zones 325
18.3.3 Assessment of visual Interpretation 326
18.3.4 Classification of map zones: agrolandscape units 327
18.3.5 Verification 327
18.4 Other Examples 328
18.4.1 Auge region 328
18.4.2 Rhone district 328
18.4.3 Aube district 328
18.4.4 Champagne-Ardenne region 329
18.4.5 Agrolandscapes and “small agricultural zones” or “agricultural
reference units” 329
18.5 Conclusion 329
References 330

19. CORINE Land Cover 331

19.1 Introduction 331
19.1.1 CORINE Land-Cover mapping 331
19.1.2 Automatic mapping 333
19.2 Data Processing and Discussion of Results 334
19.2.1 Supervised classification of land cover 334
19.2.2 Spatial integration 336
19.2.3 Assessment of results 336
19.3 Conclusion 343
References 343

20. Herbaceous Formations and Permanent Grasslands 345

20.1 Terrestrial Herbaceous Formations 345

20.2 Problems 346
20.3 Applications 348
 Contents xvìi

20.3.1 Distribution of herbaceous formations and permanent grasslands 348

20.3.2 Identification of vegetation species and groups 349
20.3.3 Mapping of units and estimation of areas 351
20.3.4 Aerial phytomass quantity estimation 353
20.3.5 Phytomass quality assessment 355
20.3.6 Monitoring and forecasting of changes In vegetation groups 356
20.4 Conclusion 357
References 358

21. Wetlands 359

21.1 Nature and Importance of Wetlands 359

21.2 Delineation of Wetlands by Remote Sensing 360
21.2.1 Marshes 360
21.2.2 Flooded and wet grasslands 361
21.3 Classification and Mapping of Wetlands 364
21.3.1 Mapping at national or regional level 365
21.3.2 Mapping at local level 366
21.4 Study of Plant Community Dynamics 368
21.5 Conclusion 369
References 370

22. Crop Inventory 371

22.1 Economic Aspects 371
22.2 Operational Use of Spatial Remote Sensing 372
22.3 Areal Estimation In Europe by Remote Sensing 372
22.4 Agrometeorological Models of Yield Estimation 374
22.4.1 Yield estimation methods 374
22.4.2 MARS method or CGMS model 375
22.4.3 Remote sensing as a complement to models 375
22.4.4 Perspectives 376
22.5 Conclusion 377
References 377

23. Soil Mapping 378

23.1 Remote Sensing and Soils 378
23.2 Directly Sensed parameters 378
23.2.1 Spectral characteristics of soils in visible and near infrared bands 378
23.2.2 Surface states of soil cover 379
23.2.3 Colour 379
23.2.4 Roughness 379
23.2.5 Carbonate 383
23.2.6 Organic matter 384
23.2.7 Iron 385
23.2.8 Moisture 386
23.2.9 Grain size 388
23.2.10 Salts 389
23.2.11 Multifactorial analysis 390
23.2.12 Linear clusters of soils 391
23.3 Model of Image Interpretation 392
xviii Processing of Remote Sensing Data

23.4 Remote Sensing and Soil Mapping 394

23.4.1 Soil landscape and remote sensing 394
23.4.2 Method of soil landscape elaboration 395
23.4.3 Example 396
23.5 Conclusions 399
References 399

24. Mining Geoiogy 401

24.1 Metallogeny and Types of Deposits 401
24.1.1 Models of mineral deposits 401
24.1.2 An example: uranium deposits in France 402
24.1.3 Regional metallogeny and metalliferous provinces 402
24.1.4 Metallotect: an operational concept 403
24.2 Multiscale Approach in Mineral Prospecting 404
24.2.1 Multlscale structural model in remote sensing of vein deposits 404
24.2.2 Example of regional application of structural interpretation 405
24.3 Remote Sensing: an indirect method 405
24.3.1 Reflectance properties 405
24.3.2 Use of thermal band 409
24.4 Conclusion 412
References 413

25. Remote Sensing and Coastal-zone Management 414

25.1 Introduction: General Problem 414
25.2 Spatial Observation of Coastal Environments 415
25.2.1 Littoral and coastal-zone objects 415
25.2.2 Littoral objects and specifications of aerospatial observation systems 415
25.3 Spectral Characteristics of Littoral Objects 418
25.3.1 Intertidal and subtidal littoral environments 418
25.3.2 Mineral targets of intertidal zone (optical domain) 418
25.3.3 Vegetal targets of intertidal zone (optical domain) 418
25.3.4 Subtidal zone: hydrocarbon pollution 421
25.4 Examples of Application to Littoral Management 422
25.4.1 Mapping of coral environments 423
25.4.2 Aquaculture management (raising tropical shrimps) 423
25.4.3 Thematic mapping of seaweeds 424
25.4.4 Detection of hydrocarbon pollution in sea 424
25.4.5 Monitoring surface state of sea by radar Imaging 424
25.4.6 Detection and monitoring of sea-surface temperature variations in
littoral environment 426
25.5 Conclusions 427
References 427

26. Applications of Thermal-infrared and Microwave Data 430

26.1 Thermal-infrared Data 430
26.1.1 Detection of drought in France 430
26.1.2 Estimation of exchanges between soil, vegetation and atmosphere 432
26.1.3 Mapping rainfall distribution in Sahel 432
26.1.4 Characterisation of frost zones 433
 Contents xix

26.1.5 Analysis of topoclimates from surface temperatures 433

26.1.6 Use of surface temperature and vegetation index 435
26.1.7 Conclusion 435
26.2 Microwave Data: Use of Radar images 436
26.2.1 Characteristics of radars 437
26.2.2 Image quality 440
26.2.3 Applications of radar data 442
26.2.4 Conclusion 447

References 449
Glossary 451
General References 457
Useful Internet Sites 461
Alphabetical Index 465
CD Rom Images Index 487
 A
DATA SOURCES
 ________ 1
Physical Basis
The subject presented in this chapter provides the fundamentals necessary for understanding the
techniques and applications of remote sensing. For more detailed information, the reader may refer to
specialised books, in particular the one by Guyot (1997), from which some of the concepts presented
here have been taken.

1.1 RADIATION
Remote sensing is based on the utilisation of emission and reflection properties of electromagnetic
radiation, which carries energy and propagates without attenuation in vacuum but is more or less
absorbed in different media. The solar radiation constitutes the external source of energy for the Earth.
This energy is fixed by chlorophyll vegetation for producing living matter (primary producer). It will be
seen later how this characteristic is used in remote sensing to study vegetation (Chap. 4).
Any object at a temperature higher than 0° K acts as a source of electromagnetic radiation by
converting a part of the thermal energy into radiative energy. In addition, the object also receives
energy from its environment, partly absorbing it and transforming it into thermal energy. The fraction of
energy absorbed modifies internal energy of the object, which is represented by emission in a different
wavelength. The blackbody is an ideal source. Its radiant exitance M q = dOg / dS (radiant energy flux
Og emitted by an extended source per unit area in a hemisphere) is independent of the angle of
emission (a) and depends only on its temperature according to Planck’s radiation law:

dM/(A-T) / dX = 2nhc / [expihc / XkT) - 1] in W • • ijm-'' (1)

where T is the absolute temperature In Kelvin, X the wavelength in pm, h Planck’s constant:
6.63 X J-s, cthe velocity of light: 3 x 10^ m-s“ ^ and/cthe Boltzmann constant: 1.38 x
The exitance of a blackbody Is related to the radiance by the equation:

M q — 71 X Lg

Integration of equation (1) over the entire spectral region gives the total exitance emitted by a
blackbody (Stefan-Boltzmann law):

M = aT 4

where a is the Stefan-Boltzmann constant: 5.67 x 10“ ®W m '^ ’ K” "^.

According to Planck’s law, radiance decreases with temperature (7).The maximum wavelength of
emission of a body is given by Wien’s displacement law:

= 2 8 9 7 /7
4 Processing of Remote Sensing Data

For the Sun, whose temperature is approximately 6000° K, Xmax = energy

flux occurs in the wavelength range of 0.15 to 4 pm (CD 1.1; see the CD enclosed with the book). For
the Earth, whose temperature is about 300° K, = 9.6 pm and the energy flux lies in the wavelength
region of 3 to 100 pm.
A comparison of the flux reflected by the Earth (assuming its reflectivity as 1) and the flux directly
emitted by the Earth (assuming its emissivity as 1) is shown in Fig. 1.1 .The atmospheric windows are
shaded (Table 1.1 and section 1.4).

Infrared
I , ■ 1 Atmospheric windows p = Reflectivity

Fig. 1.1: Comparison between the fluxes reflected and emitted by the Earth (after Becker, 1978).

The Earth’s surface behaves like a grey body. Spectral emissivity of a natural surface (e) In the
direction 0 is defined as the ratio of the radiance of the surface LxiQ) to that of the blackbody Bx:

Sy^{Q) = L^{T,Q)/B^{T)

For the entire spectrum, we get;

R — 80 Tg + (1 £)^atm

is the atmospheric radiation, which will be discussed In connection with thermal infrared remote
sensing (Chap. 26 and Table 1.3).
Any radiation comprises a wide range of wavelengths;

X = cN

where c is the velocity of light; 3x10® m-s""* and v the frequency in s“ ^

Energy carried by a photon Is; e = h v , where v = clx and h Planck’s constant. The total energy of
a radiation is;

E = he Z{nx / X)6X
 Data Sources 5

where is the photon density: a;^) A/the number of photons per unit time and AX the
spectral window or band.
The various domains of the electromagnetic spectrum and the main types of sensors used in
remote sensing are shown in Table 1.1.

Table 1.1: Electromagnetic spectrum and remote sensing systems

SPECTRAL Remote sensing systems

DOMAIN

290 nm UV scanner
ULTRAVIOLET
380 nm
Photography

400 nm
VISIBLE Colour
700 nm panchromatic
500 mn
Colour
infrared
700 nm
800 nm 900 mn
Black and
white infrared
1500 nm 900 nm

1.6 pm

2.2 pm

3 pm

5 pm

Infrared radiometers
8 pm

14 pm

MICROWAVE Radar, backscatter sensors,

1.6 pm
radiometers
136 cm

1.2 SOURCE AND SENSOR PARAMETERS

The radiant intensity {1^) is the energy flux emitted by a point source per unit solid angle in a given
direction:
6 Processing of Remote Sensing Data

9 ) = clOg(0, (p) / dQ in W •sr"^

where 0 is the angle of zenith, cp the azimuth and Q the solid angle. If dZ is the surface situated at a
distance rfrom the source and corresponding to the solid angle di2, then dQ = dZ !r^ (Fig. 1.2).

Fig. 1.2: Representation of the characteristics of a source.

If a point source is isotropic, it emits a radiant flux whose intensity is independent of the
orientation under consideration:

/e = / 4 ti in W -sr
The concept of radiant intensity is valid only for a point source. In the case where the dimension
of the source is small relative to the distance at which measurement is carried out (Sun for
measurements on the Earth), the source can be considered a point.
Radiance (L^) of a source corresponds to the total energy radiated by it In a given direction per
solid angle per unit area of its apparent surface:

L^(0,cp) =: d^Og(0, (p)/dQ x dS X COS0 in W m ^ -sr I

Radiance can also be defined, with reference to a point on the surface in a given direction, as the
ratio of the radiant intensity of an infinitely small surface element dS around the point and the area of
orthogonal projection of this element on a plane perpendicular to this direction:

9) = (p)/dSX COS0 in W -m •sr

The radiant exitance (M^) is the energy radiated by an extended source per unit area and in a
hemisphere. At a given point, It is equal to the ratio of the flux emitted by an Infinitely small element of
the surface surrounding the point to the area dS of this element: Mq = dOg / dS in W •m“ ^ •sr~^ •
For a radiation sensor, irradlance (Eg) is defined as the total energy received by a unit surface
area of the sensor. At a point, this irradlance corresponds to the ratio of the radiant flux received d<D er
by an infinitely small element centred on the point to the area dS of this element: Eq = d<l>e/.

dS in W-m"^-sr"^.
The Bouguer law represents the relationship between a source element and a receiver element.
Let us assume two elementary surfaces, of areas dS and dS ', separated by a distance r, whose
normals make angles 0 and 0' respectively, with the axis joining their centres. The radiant flux d^ e
that reaches dS' from the source dS of radiance Lg, is equal to:

d^Og = [(/-g xd S x co s 0 )/r^] (d S 'x c o s 0 ).

If dQ Is the solid angle under which dS is seen from dS' and dQ' the solid angle under which dS'
is seen from dS,

dQ = (d S x c o s 0 )/r^

dQ' = ( d S 'x c o s 0 ')/r^

 Data Sources 7

It follows from the above that:

= Z_g X dS X COS0 X dQ' = L^x dS' x cos0 x dQ.

In remote sensing the fluxdOg,., reaching a receiver from a source, is measured and the radiance
is determined from it. This enables understanding the properties of surfaces.
Albedo corresponds to the fraction of solar radiation (direct and scattered radiations) reflected by
a unit surface area into a hemisphere. It represents the mean hemispherical spectral reflectance in the
range 0.3 to 3.0 pm.
The blackbody emits radiation In all directions and as seen earlier, radiance Is Independent of the
angle of emission. Contrarily, other bodies radiate in preferred directions and their radiance depends
on the angle of emission. According to Lamberfs law:
/ = / g / d S = Lg X cos 0
where / is the radiant intensity emitted, the radiant intensity emitted in a given direction and the
radiance.

Radiant intensity of a Radiance Characteristics of Lambertian,

Lambertian Surface specular and real surfaces

Fig. 1.3: Characteristics of radiant intensity and radiance of different surfaces.

Near-perfect specular reflector

Perfect diffuse Lambertian surface

Fig. 1.4: Examples of various radiant-intensity characteristics

8 Processing of Remote Sensing Data

The characteristic representation of angular variation of radiance (or radiant intensity) at a point
on a surface is shown in Fig. 1.3. In polar co-ordinates, the geometric characteristic of radiant intensity
of a Lambertian surface is a sphere tangential to the surface, of a diameter equal to /^, while the
characteristic of radiance is a hemisphere centred on the surface and of radius L^. Natural surfaces
generally do not follow Lambert’s law and their radiance or radiant intensity varies with the angle of
view. In the case of a perfect reflector, the angle of incidence (0^) and the angle of reflection (0;.) are
equal and situated in a plane perpendicular to the surface. Such a reflection is known as specular
reflection (Fig. 1.4).

1.3 REFLECTANCE FACTOR

Spectral reflectance factor {R^) Is the ratio of the radiant flux reflected by a surface element into a
cone, with its apex on the surface element, to the radiant flux reflected into the same solid angle by a
perfect diffuse reflector (white Lambertian surface) receiving the same irradiance (Fig. 1.5).

R ^ = JL^(0^, cp^)xcos0;. x d Q ;./L n , \cosQfx6Clf . = Lr(Q.r)^L^

Qr Qr

where is the solid angle in which the radiant flux is measured and the radiance of a white
Lambertian reflector.
 Data Sources 9

In conclusion, the optical properties of natural surfaces are very different depending on the spectral
domain under consideration. For example, fresh snow reflects 95% of solar radiation whereas it behaves
like a blackbody in the thermal infrared band (0.90 < s < 0.99). Similarly, albedo of bare soils may vary
significantly with moisture content while their emissivity is less sensitive to water content.

1.4 SOLAR RADIATION AND ATMOSPHERIC

PERTURBATIONS
This section presents the definitions and a review of concepts required for understanding certain data.
More detailed information concerning the Influence of atmosphere on radiation will be given in the
section dealing with thermal infrared remote sensing.
True solar time is the dihedral angle made by the meridian in the Sun’s direction with the observer’s
meridian. It is used to determine the position of the Sun and indicates solar dials. It is zero during the
passage of the Sun in the meridian (solar noon).
Mean solar time is the hour angle observed for an apparent regular motion of the Sun.
Civilian time is the mean solar time enhanced by 12 hours.
Universal time (UT) is the civilian time of the Greenwich meridian chosen as origin.
Legal time or local standard time Is constant on an hour’s axis. It is obtained by adding or subtracting
a certain number of hours to or from the universal time.
Irradiance of a plane surface, normal to the solar radiation and situated at the mean distance Dq
from the Earth to the Sun, is almost constant (with variations before solar activity). This is known as
the solar constant E^. The irradiance (E^^) of a horizontal surface situated at the boundary of the
atmosphere depends on the distance from the Sun and the zenith angle 0. The value of varies
between winter solstice and summer solstice, and the zenith angle varies with the latitude, hour and
season. The greater the zenith angle 0, the smaller the irradiance (Fig. 1.6):

E ,M = E JD ./D fxcose

1.4.1 Atmospheric absorption

The influence of atmosphere on solar radiation is shown in Fig. 1.7. The top curve represents the solar
Irradiance received at the boundary of the atmosphere, whereas the bottom curve the solar radiation

Fig. 1.6: Influence of zenith angle on the irradiance of a horizontal surface (after Guyot, 1997).
10 Processing of Remote Sensing Data

W-m'^

Fig. 1.7: Influence of atmosphere on direct solar radiation (after Perrin de Brichambaut, 1985).

received at sea level. The intermediate curves correspond to absorption by various gases. Ozone
absorbs wavelengths below 290 nm and causes a little attenuation at 600 nm. Oxygen has an intense
but very narrow absorption band at 760 nm. In the range from near-infrared to thermal Infrared, water
vapour, carbon dioxide and methane are mainly responsible for atmospheric absorption.
The larger the thickness of the atmosphere traversed by solar radiation, the weaker the direct
irradiance. The spectral distribution is also modified (Fig. 1.8) with a greater decrease of visible radiation
in the violet and blue spectral bands than in the orange and red bands.
The atmosphere transmits solar radiation only in a limited number of spectral bands known as
atmospheric windows. The atmosphere is relatively transparent In the visible band and has a wide
atmospheric window in the infrared (8 to 14 pm). It is opaque in the range 22 pm to 1 mm and hence
this part of the remote sensing spectrum is not used. In the microwave region, the atmosphere Is
transparent beyond 3 cm but becomes opaque for wavelengths greater than 30 mm due to interaction
with the ionosphere.

W-m'^

Fig. 1.8: Influence of Sun’s height on the spectral distribution of direct solar irradiance on the ground
(after Perrin de Brichambaut, 1985).
 Data Sources 11

1.4.2 A tm ospheric scattering

Scattering in the atmosphere corresponds to the action of molecules and particles (water droplets,
dust, fumes etc.) on the wavelength (/i).The particle diameter dvaries from:
1 to 500 nm for charcoal dust
0.5 to 50 pm for industrial fumes, fog, dust
10 to 100 pm for pollens, ash
20 to 300 pm for clouds, mist
0.5 to 5 mm for raindrops.
The following cases may be distinguished:
• ^ > of. In this case, the Rayleigh scattering is predominant due to interaction of photons with
molecules. This scattering is proportional to . It plays an important role for short wavelengths
{X < 550 nm) where it Is associated with the action of oxygen and nitrogen molecules. This
process of interaction explains the blue colour of the sky and the red colour of the rising or setting
Sun. In fact, blue light =450 nm) is scattered six times more than red light {X = 700 nm). When
the Sun is low on the horizon (Fig. 1.8), the radiation traverses a thicker atmospheric layer than
that when the Sun is at the zenith. Small wavelength radiation is hence eliminated and the Sun
looks red.
• X^^0~‘^ <d<X^10^.\n this case, the main type of scattering is the Mie scattering due to suspended
aerosols in air (dust, microdroplets of water). This scattering Is roughly proportional to X~^ and
manifests in the entire solar spectrum but has very small Influence in the thermal infrared band. It
is especially significant near industrial sites.

1.4.3 Atmospheric radiation

The atmosphere re-emits part of the radiation received, which contributes to Increase the value of the
flux measured by the sensor (Fig. 1.9). This atmospheric radiance adds to the flux reflected or emitted
by the Earth’s surface, which constitutes the object of study in remote sensing. It can be measured
Contribution to the irradiance of the target

Contribution to measured rediance

Target Environment
Fig. 1.9: Influence of the atmosphere and the environment of the target on the signal measured by a satellite.
12 Processing of Remote Sensing Data

from vertical profiles of temperature and humidity obtained by radio-soundings andTOVS sensors of
NOAA satellites. It can also be modelled using different formulae (see infra) since the atmospheric
effects are particularly significant on satellite-measured fluxes.
Atmospheric scattering enhances the reflectance observed at the top of the atmosphere (p')
relative to that observed on the ground (p). An example of reflectance for a green wheat and a ripe
wheat Is shown in Fig. 1.10.

Fig. 1.10: Effect of atmospheric scattering on reflectance observed at the top of the atmosphere ( p ') and on the
ground (p) for a green wheat and a ripe wheat (after Deschampe et al., 1984 in INRA Publications).

1.5 THERMAL INFRARED REMOTE SENSING

The thermal infrared band corresponds to a range of wavelengths varying from a few pm to several
hundred pm, situated between middle infrared and microwave. Some Instruments operate around
3 pm but the widest range used lies between 8 to 14 pm. Measurements in this band using airborne or
satellite sensors enable determination of surface temperatures and furnish Information on the aqueous
and energy state of surfaces. Such measurements are hence significant in the agricultural sector for
monitoring soils and crops.

1.5.1 Physical basis

It was Indicated at the beginning of this chapter that any body with temperature different from absolute
zero emits a radiation depending on Its temperature. Emissivity is always less than 1 and generally
greater than 0.9 for most natural objects. This parameter depends on the wavelength and the direction
0, as well as on the moisture content of surfaces. Its measurement still remains difficult and Is not
representative in natural conditions except very locally (Stoll, 1988). In practice, values found in
publications are used. Its precise knowledge is however essential since a relative error of 1% in emissivity
may result In an error of 0.8°C in surface temperature. Emissivity values for some surfaces are given
in Table 1.2.
Brightness temperature (or apparent radiant temperature T^) of a surface corresponds to the
temperature of a blackbody which delivers the same radiance as that of the surface under investigation.
It can be defined for a spectral band by the relation:

) B^(Tb ) = ) s A i T ) d X (2)
X, X1
 Data Sources 13
Table 1.2: Emissivity values of some important types of surfaces (after Guyot, 1997)

Type of surface Emissivity

Ice 0.92-0.97
Water 0.99
Bare soil (dry sand, wet sand) 0.94 (0.84-0.90; 0.91-0.94)
Grassland 0.98
Wheat 0.97
Maize 0.96
Grapevines, orchards 0.95
Conifers 0.98
Hardwood 0.97

or over the entire spectrum by:

:Ea V (3)
Radiant (brightness) temperature is a directional parameter which varies with the direction 0.
Hence, it varies according to the spectral bands of the satellites used. Conversion of this temperature
to radiometric temperature requires a correction for atmospheric radiation, as given below.
In these wavelengths absorption of radiation by snow is very high and only the surface temperature
affects the radiance. Emissivity of snow is close to 0.99 in this spectral domain.
The surface of the sea can be considered Lambertian in the thermal infrared band and the general
equation for conversion of the radiance received by the sensor to sea surface temperature Is based on
Planck’s law. The emissivity of water is very close to 1, and water surfaces are good accumulators of
heat, exhibiting very small thermal amplitudes on the surface. This is due to convection motions existing
between water masses of different temperatures situated at depth.The observed temperature is related
only to the surficlal layer of water over a thickness equivalent to the observation wavelength. For more
details of software for computing temperature with AVHRR sensors of NOAA satellites, refer to Barton
(1989).

1.5.2 Correction for atmospheric radiation

The radiance received at the point of the receiver depends on the emission by the surface as well
as from the atmosphere (Fig. 1.11). The long wavelength atmospheric radiance (L^) results from
radiation emitted by particles in the atmosphere (see supra). Radiance contributions from water vapour,
carbon dioxide and ozone are the largest. Suspended aerosols also participate in this thermal emission.
The atmospheric radiance comprises different components:
— Its own radiation (thermal emission by particles present in It): On the ground, this radiation
results from all the particles present on the entire profile and having varying temperatures according
to their altitude. This flux comes down in the direction of the surface being irradiated by the atmosphere

— At the receiver, the resultant of contributions from particles, which is a flux directed towards the
sk y(i-a xt)-
Concentration of gases, aerosols and temperature can be determined by means of radio soundings.
They can also be obtained from climatic atlases. The atmosphere attenuates the surface radiance by
an absorption factor t; as a result, the surface emission becomes reflected
14 Processing of Remote Sensing Data

Fig. 1.11: Components of radiance received by the thermal infrared sensor,

atmospheric radiation (1-s;^)La;^.Thus the final equation for radiance at the receiver is obtained as:

k = ^[sA(Ts) + (1-s,)/-ax^] + i-aX Î (4)

where x is the atmospheric absorption coefficient (Fig. 1.11).
Knowledge of emissivity of the surface and atmospheric radiance is hence necessary In order to
determine the surface temperature. Several possibilities exist to determine the emissivity. It can be
measured in the field although such measurements are highly local and cumbersome. Alternatively,
emissivity data can be accessed from published literature. One can also use relations between emissivity
and vegetation index (see Chap. 4) computed from satellite images In the visible and infrared bands
(actes de télédétection IRT, 1993). To estimate atmospheric radiation and corrections for obtaining
surface temperature, various methods are proposed depending on the type of measurement: single or
multichannel thermal infrared sensors, radio soundings etc.

■ Atmospheric corrections for single-channel sensors

Satellites such as LANDSAT or METEOSAT have only one channel in the thermal Infrared band.
Flence atmospheric correction models such as LOWTRAN (Knelsys et al., 1983) or MODTRAN (1995)
have to be used for them. These models are based on dividing the atmosphere into a multilayer
system and solving the transport equation for each layer. The necessary data characterising the
atmosphere (thermal profiles, humidity and pressure) are obtained either from radio soundings or
from indirect measurements, from outputs of meteorological models or forecasts. Use of meteorological
forecasts poses many problems, which have been discussed by Kerr (1991). Among these problems,
we cite those concerned with the hypothesis of surface homogeneity, the cumbersome method of
computations and errors introduced by these models.
Direct measurements by radio soundings or use of the TOVS sensor of NOAA, which provides
vertical profiles of temperature and humidity, are recommended for single-channel sensors (Ottlé and
Vidal-Madjar, 1992). Combining these data with a radiative transport model such as MODTRAN, It is
possible to accurately correct for atmospheric effects on surface-temperature measurements. The
results are generally satisfactory provided a radio sounding is available close to the region of study
and to the time of acquisition of image. However, interpolation of radio soundings is often problematic.
In France, only seven sites exist where radio soundings are carried out routinely every day at 12 and
24 hour (sites in Important airports). Density of radio sounding sites is much smaller in other regions
 Data Sources 15

such as West Africa. This constitutes a major constraint in employing these correction methods. Data
issued by TOVS soundings do not yet overcome the problems of representativeness since
determinations of the characteristics of the lower atmospheric layers are still difficult and too few. In
fact, as this probe does not provide access to direct measurements that are equivalent to radio
soundings, It Is necessary to reconstruct data acquired with low resolution (Ottle and Vidal-Madjar,
1992).

HAtmospheric corrections for two-channel sensors

Methods based on differential absorption can be used for atmospheric corrections in two neighbouring
channels (such as Channels 4 and 5 of AVHRR-2). It is assumed that the spectral emissivity of the
surface in the two spectral bands is identical. The apparent radiative temperature in such a case
results from a linear summation of temperatures obtained in the two channels of the sensor:
T s- Bq + 4(10.8) ^2 "^c 5(11.9) (5)
This method, known as Split-Window, was initially used to compute sea-surface temperatures for
which it gives good results. However, it poses more problems for land surfaces because of differences
in spectral emissivity between bare soils and vegetation covers (Kerr, 1991). The different values of
coefficients are found to be more or less suitable or unsuitable for exposed ground. Kerr (1991) proposed
a semlempirical method based on computing two temperatures from different values of coefficients:
one corresponding to bare soils (7^,^) and the other for dense vegetation (7^^). The two temperatures
and are summed linearly as a function of a coefficient CV, representative of the percentage of
soil cover, and divided by vegetation index (VI) to obtain the final surface temperature.

(7^,) = 3 .1 + 3 .2 7,c4(10.8 ) *^c5(11.2)

i'^ d v ) - 2.4+ 3.6 7^4(.|Qgj-2.6 T ^ s { u .2 )

CV= a /V

a=(C V,-C V,)/{IV^^^-IV^J

T s = C V T ,^ -,(U C V )T ,, (6)

where a is a coefficient varying between 0 and 1 , corresponds to bare soils or poorly covered
surfaces, corresponds to dense vegetation cover or cultivated land. This method gives a precision
of about 1 .5 K.
Another method of atmospheric correction is based on using the data acquired under varying
conditions of observation (e.g. ATSR sensor on ERS-1 which acquires two views) (Ottlé and François,
1993).

■ Sources of errors in measurement by satellite

Let us now review various factors that disturb measurement of 7^. These factors may affect only
certain sensors (e.g. saturation phenomenon) or certain satellites (drift acquisition in time).

□ Orbital drift
The local sun time on each pass of a satellite changes with time. Data issued from the NOAA-9
satellite launched in 1984 and coming to an end in 1988 exhibited a difference of 2 hours in acquisition
time in 1988 relative to the initial time at beginning of launching (Privette etal., 1995). Hence, It Is often
necessary to apply corrections to bring temperatures to the same time in order to compare and to
16 Processing of Remote Sensing Data

monitor temporal variations. The relationships between surface temperature acquired at different times
can be used for this purpose. Based on field data obtained using a thermal radiometer, it is in fact
shown that the relationships between these variables are linear. An example of such relationships is
given by the following equation:

7,14, = (1 -03 ± 0.01 ) (7 ,1 5 - W + (3-3 ± 1 .7 )+ 7„ (7)

where is the temperature of the surface under investigation, the temperature measured by
NOAA-9 in 15 h and the maximum temperature of air measured over a meteorological site
situated in the image.
This relationship was used to correct the NOAA-1988 surface temperatures In order to compare
them (CD 26.1) with those of 1990 and 1991 (Courault et al., 1994).

□ Signal saturation
This factor also needs to be corrected due to the fact that the NOAA sensor gets saturated above 320
K (Kerr, 1991); however, it affects only some regions of the globe.

□ Angular effects
Several types of angular effects are identified:
— atmosphere-related effects (difference in optical trajectory),
— effects related to directional emissivlty, not understood at present (Becker, 1987),
— effects associated with structure of plant canopies.
Models have been developed to ascertain the effects of vegetation structure, combining a
description of thermal profiles inside vegetation with the structure of the canopy (angle of leaves, foliar
Index etc.; actes de télédétection IRT, 1993). Values of radiative temperatures acquired with thermal
radiometers along different angles showed that differences in temperatures between a vertical view
and an inclined view may reach up to 4 K over maize and 1.5 K over grasslands (actes de télédétection
IRT, 1993). Temperature values of vertical measurement are generally higher due to the contribution
of soil, which is Inversely proportional to the density of vegetation.

1.5.3 Relationship between radiant temperature and

aerodynamic temperature
A whole nomenclature exists in thermal infrared technique wherein the surface temperature (general
term) may Indicate various physical parameters. Norman and Becker (1995) discussed these differences
In terminology of surface temperature and emissivlty. In this section, only the relationships between
the radiative temperature measured by a thermal infrared sensor and the aerodynamic temperature of
plant canopies Is reviewed.

H Definitions
The aerodynamic temperature of a surface (T^^) corresponds to the extrapolation of the temperature
profile of air to a height of the apparent canopy given by the height of displacement (d) plus the length
of roughness (Zq). It occurs in the estimation of sensitive heat flux (equation 3). It is combined with air
temperature and aerodynamic resistance (ra) which Is calculated according to the Monin-Obukhov
theory. Variables related to the wind and temperature profiles such as thermal (Zq) and mechanical
(Zq^) roughness are used (Guyot, 1997).
The radiative surface temperature T^ps defined from the radiation emitted by the surface (see
above). It integrates the surface temperatures of all the elements of a canopy, such as leaves in shade
and leaves in sunlight, viewed by the sensor. It depends on the structure of vegetation and wind
 Data Sources 17

profiles in the canopy. It is a directional parameter which corresponds to a given wavelength range
depending on the receiver. If Is the radiance measured by the sensor and the emissivity of the
surface, we get:

^ 'I'

^^X^xC^sr) (8)

If the surface temperature measured by satellite is to be used to compute the detectable heat flux
for vegetation of strong vertical development, a correction has to be applied. It is generally considered
that aerodynamic resistance ra = + r2, where:

= (V ku) [ln {(z - d/z^J} - w ^{(z-d)/L}]

= (V ku) In (z ^jzo ) = (V ku) KB~^

where KB~^ represents the ratio between mechanical and thermal roughness, expressed in logarithmic
form, L is the Monin-Obukhov length and u is called friction velocity.
Becker and Li (1995) gave the relationship between these two temperatures as:

~^sa ~ ~^sr
1+ Ò

b = KB-y kura (9)

Differences between aerodynamic and radiative surface temperatures may vary from -8 K to 4 K
depending on the values of the foliar index (actes de télédétection IRT, 1993).

■ Accuracy of data
The accuracy of surface temperatures thus determined from satellite data is difficult to evaluate without
choosing homogeneous zones of sufficient size so as to include many pixels. It can be estimated only
locally from measurements with thermal radiometers (which are often of wide band, 8-13 pm) and
without covering exactly the same spectral region as that of the sensor. As thermal radiometers view
surfaces of limited dimensions, the measurements are affected by interfield variability. Spatial sampling
of measurements has to be large enough to be representative of the variability of the zone under
study. When the segments are small and varied in land cover, the pixels are heterogeneous. An overall
measurement at the pixel level is obtained from the composite fields. Different methods are thus
possible depending on the objective of investigation: work at the level of the pixel and adjust the
models to low resolution or employ methods of resolving the pixels to arrive at the level of the field (see
Chap. 15). At present, little work is reported on spatial integration or disintegration of 7^. Moreover,
tables of emissivity covering the range of possible surfaces and records of variation of the characteristics
of a surface with time are required.
Various studies have reported comparisons of accuracy In surface temperature obtained by different
methods for correction of thermal data (Kerr, 1991 ; Ottlé and Vidal-Madjar, 1992). Among these, the
tables of Chanzy (1991), which are useful In evaluating the variability of measurements depending on
the values of atmospheric radiance and emissivity employed, are given in Table 1.3. As values of
emissivity are often unknown, these tables show that approximation in assuming 7^ = 7^, Introduces
errors which are larger for smaller values of emissivity and atmospheric radiance.
18 Processing of Remote Sensing Data

Table 1.3: a. Errors in determination of surface temperature when = Tg for different values of
atmospheric radiance and emissivity.
b. Errors in determination of surface temperature when emissivity is assumed constant at 0.95, for different
values of atmospheric radiance and emissivity of soil (after Chanzy, 1991).

Emissivities

^atm a b

(W-m2) 0.85 0.90 0.92 0.98 0.85 0.90 0.92 0.98

100 9.21 6.04 4.81 1.18 6.37 3.14 1.87 -1 .8 3

200 6.56 4.32 3.44 0.85 4.55 2.25 1.34 -1 .3 2
300 3.97 2.63 2.10 0.52 2.77 1.38 0.82 -0.81
400 1.46 0.97 0.77 0.19 1.02 0.51 0.30 -0 .3 0

The accuracy reported in most studies is of the order of 1 to 2 K, which in absolute value is
significant for point measurements but acceptable only when cumulative values over a year are analysed.
Hence, one has to be cautious in using thermal infrared data.
The principal use of data of this spectral band is to evaluate the aqueous and energy states of
surfaces. Before describing some examples of application in agriculture, it is necessary to review the
relationships between surface temperature and évapotranspiration.

1.5ARelationship between surface temperature and

évapotranspiration
ffl Energy balance
Surface temperature is indirectly related to évapotranspiration of plant canopies through energy balance.
In the case of a thin surface (bare soil or low canopy) which can be considered to be in instantaneous
equilibrium, the energy balance is described as:

R ^= LE +H + G ( 10)

where is the net radiance representing the energy balance of various radiative fluxes of short and
long wavelengths at the ground level. It is expressed as:

Fin = i^-^)Fig + e ^atm “ sc? T f ( 11)

where Rg is the total radiance, part of which is reflected by the surface (a being the albedo), the
atmospheric radiance and so the radiance emitted by the surface.
The various equations for each of the fluxes are not discussed here.
The conduction flux G of the ground can be neglected in the scale of a day for dense canopies.
However, Its instantaneous value can be significant in the case of bare and dry soils. For plant canopies,
a conventional order of magnitude is 0.1 to 0.2 in diurnal conditions. G can be expressed in the
form:

G=(K^/A^I{T^-T^) (12)

where K is the thermal conductivity and Tq the ground temperature at depth Az

 Data Sources 19

The sensible heat flux H, related to convective exchanges in air, depends on the difference of
temperatures between the surface and air:

H = pCpy^
(Ts^ ^
ra
(13)

where p and Cp are the density and specific heat of air, the aerodynamic temperature of the
surface and ra the aerodynamic resistance, which varies depending on the wind velocity, surface
roughness and stability conditions of the atmosphere (Guyot, 1997; Courault et al., 1996).
The latent heat flux LE (energy equivalent of évapotranspiration E expressed as mass flux) is
written as:

LE: :pC, {Ts)-es

y(ra + rs)
(14)

where (T^) is the saturated vapour pressure at the temperature T^, y the psychrometric constant
(0.66 mb-K"“'), the vapour pressure of air at the reference level (2 m) and rs the total resistance of
the canopy. The parameter rs depends on climatic stresses to which plants are subjected and the
moisture state of the soil.
Each of the fluxes can hence be expressed as a function of surface temperature. Evapotranspiration
thus seems to be linearly related to and, if flux G is neglected, can take the following form for a time
interval of a day:

LE,(T,) = R, p (T^)-H,(T^) (15)

Tg is the solution of the energy balance equation and attests exchanges between the surface and the
atmosphere. In particular. It indicates the evaporation level of surfaces.

■ Simplified models
Various authors have proposed semiempirical models derived from energy balance on a daily scale.
The most widely used model shows a linear relationship between the Instantaneous temperature
difference (T ^- 7^)^measured in daytime and the daily values of (LE-R^)p, expressed in mm of water
per day (Jackson et al., 1977). Physical justification of this relationship is mainly based on conversion
of instantaneous values into diurnal values (Seguin and Itier, 1983).

L E ^- R^^^ A-i- B ( 7 g - Tgjj.)^ (16)

The coefficients A and B vary with the type of surface, in particular the roughness of the cover, as
well as the wind velocity and conditions of thermal stability (Lagouarde and McAneney, 1992). They
are obtained from ground measurements. Measurements by Seguin et al. (1982), for example, gave
these values as: >4 = 1 mm-d""' and B = - 0.25 mm d~‘‘. They can also be determined from digital
modelling (Lagouarde, 1991).
This relationship has been verified for various surface conditions and gives satisfactory results for
clear days with an accuracy of about ±1 mm d"'' for évapotranspiration. Examples of application are
given in Chapter 26.

1.6 PRINCIPLES OF MICROWAVE REMOTE SENSING

Active sensors, viz., radars (acronym for Radio Detection And Ranging), are characterised by:
20 Processing of Remote Sensing Data

— a working range in which the atmosphere is transparent for most of the wavelengths used (from
3 mm to a little more than 30 cm);
— Independence of the emission-detection system vis-à-vis the solar energy (‘active’ system);
— sensitivity for wavelengths at which the radiation-matter interactions correspond to totally different
phenomena from those recorded by ‘passive’ systems.
Only the fundamentals required for understanding and utilising the microwave data are given In
this book. The theory of radars can be found from more specialised books (for example, Paquet, 1997
and Polidori, 1997). Moreover, only the active systems are discussed, for which a large number of
programmes of study and data for observation of the Earth are available.

1.6.1 Special laws at microwave frequencies

Radiant energies emitted by terrestrial surfaces In the wavelength range of about a cm (see Table 1.4),
the most widely used, are very low compared to the thermal infrared band. A simplified Planck’s
relation (Rayleigh-Jeans law) can be employed in this case. The latter describes a linear relation
between the temperature and the radiance of a blackbody at these wavelengths:

L^^2kTn?

where Is the microwave radiance of the blackbody in W m“ ^ Hz” ‘^ s r “ ‘', kth e Boltzmann constant
(1.38 X 10"2^ J K"''), Tthe temperature of the blackbody in K and X the emission wavelength in cm.
The apparent radiant temperature (T^pp) of an object Is proportional to the energy emitted and
hence to the emissivity of the object:

^app = '^(^Tq + p Ti^) + ( 1 - x)Ts

where t is the transmittance of the atmosphere between the object and the sensor (0 < t < 1 ) , e the
emissivity of the object (0 < s < 1), Tq the absolute temperature of the object in K, p the reflectance of

Table 1.4: Frequency bands (after Paquet, 1997).

Wavelength OTAN nomenclature Radar bands

100 m A 0 to 250 MHz

10m HF 3 to 30 MHz
B 250 to 500 MHz VHF 30 to 300 MHz
1m C 0.5 to 1 GHz UHF 0.3 to 1 GHz
30 cm D 1 to 2 GHz L 1 to 2 GHz
E 2 to 3 GHz S 2 to
10 cm F 3 to 6 GHz 4 GHz
5 cm G 4 to 6 GHz C 4 to
H 6 to 8 GHz 8 GHz
3 cm 1 8 to 10 GHz X 8 to 12 GHz
J 10 to Ku 12 to 18 GHz
2 cm 20 GHz K 18 to 27 GHz
1 cm K 20 to 40 GHz Ka 27 to 40 GHz
L 40 to 60 GHz V 40 to 70 GHz
M 60 to 100 GHz W 70 to 100 GHz
 Data Sources 21

(Number of cycles per second passing through a fixed point)

Fig. 1.12: C haracteristics of a wave (after Bonn and Rochon, 1992, p. 26, reproduced with the
permission of the editor. Extracted from: R e m o te s e n s in g a b s tra c ts , voi. 1, P rin c ip le s a n d m etho ds,
University of Quebec Press, Quebec).

the object (0 < p < 1), T^the equivalent temperature of the radiation incident on the object (in K) and
the absolute temperature of the atmospheric layer situated between the object and the sensor (in K).
The value of emissivity 8 depends on a number of factors such as surface roughness, polarisation
and wavelength of the radiation, dielectric constant and temperature of the object etc. For example,
the emissivity of water is close to 1 in the thermal infrared band and about 0.35 in the microwave band.

1.6.2 Frequency bands

The frequency bands employed in active and passive remote sensing are shown in Table 1.4. While
the OTAN nomenclature applies to radar and radio bands, a radar nomenclature (unofficial) is also
given because very commonly used in active remote sensing, especially in the frequencies between
1 and 40 GHz.

1.6.3 Polarisation
An electromagnetic wave can be represented in space by elei^trlc (E) and magnetic (H) field vectors.
The vector E, which is perpendicular to the X-axis can rotate around this axis (Fig. 1.12).
Phase (p determines the manner in which E behaves In a plane parallel to yOz. If cp varies with
time randomly, the wave is non-polarised and if (p remains constant, the wave is polarised. In the case
of reflection of electromagnetic waves, two types of rectilinear polarisation can be distinguished:
— vertical polarisation, when the vector Elies in the plane of incidence;
— horizontal polarisation, when the vector E is perpendicular to the plane of incidence.
The polarisation phenomenon exists in various spectral domains used in remote sensing. However,
it is particularly significant in microwave regions in which it carries information about the object under
investigation, its surface roughness, for example.

1.6.4 Doppler effect

The Doppler effect produces a shift in frequency between the wave emitted by a moving source and a
receiver, due to relative motion between the source and the receiver. It has important applications in
microwave frequencies for the synthetic-antenna radar.
22 Processing of Remote Sensing Data

1.6.5 Backscatter signal

In a radar system, the source and the receiver are almost at the same place. An electromagnetic wave
of frequency /and power is emitted by the source in the direction of the surface under observation.
On reaching the surface, the wave undergoes an interaction and wave is scattered in all directions in
space. A part of it is thus returned in the direction of the radar: this is the backscatter signal ( oq ) with
a power (Fig. 1.13). From the thematic point of view, it is necessary to understand how the surface
under investigation affects the backscatter signal.

Receiver

Source

Surface

Fig, 1.13: Backscatter signal.

1.6.6 Radar equation

For a given wavelength and polarisation, the radar equation describes a relationship between Pg and
^bs’ separating the effect of logistic parameters, that characterise the instruments used and the source,
from the effect of the surface on backscattering of the signal:

(Gs-Gr)
(4n)^

Characteristic Backscatter
of instrument coefficient

where P^,g is the energy received, P^the energy emitted, X the wavelength, P the distance between
the surface and the antenna, G^the gain of the source antenna, G^the gain of the receiver antenna,
gq the reflecting power per unit area of the surface irradiated and S the area of the trace on the
ground.
It is therefore necessary to thoroughly understand the logistic parameters before identifying and
interpreting gq and its relationship with the properties of the surface under investigation.

1.6.7 Logistic parameters

The main features of logistics are very similar for all active systems whether they are point scatterometers
(with fixed posts) or Imaging radars (airborne or satellite mounted) and are divided Into three groups.

H Source-receiver system
The source-receiver system of a radar, by design, has its own characteristics defined by the following
parameters:
— power of the source P^,
— gain of the source antenna G^,
 Data Sources 23

— gain of the receiver antenna

— loss of energy p.
These parameters undergo variations which can be controlled and corrected or compensated by
calibration coefficients depending on the system used.

B Geometric parameters
The principal parameters in measuring Oq are (Fig. 1.14):
— Angle of incidence of the beam, determined by altitude and inclination of the antennas and the
distance between the receiver and the surface viewed, on the ground. The depression angle corresponds
to the inclination of the antenna on the platform relative to the horizontal as well as to the complement
of the Incidence angle of the wave relative to the surface when the latter is horizontal;
— Cone in which the energy backscattered by the irradiated surface (imprint) is measured, defined
by the field of view Gq of the receiver antenna.
The angle of incidence is important in measurements since it affects detection of the roughness
of the irradiated surface (see the section below). Depending on the type of receiver, the choice of
incidence angles is limited due to constraints imposed by the platform, among others.
The field of view of the antenna and the geometry of its positioning directly influence the geometric
resolution of the system. In the case of an airborne or satellite system, the concept of ground resolution
is complicated due to movement of the platform (longitudinal resolution) (see Chap. 26).

/ : Irradiation beam
a : Depression angle
P : Incidence angle

Fig. 1.14: Characteristics of an antenna.

■ Source characteristics
The characteristics of the source signal are the wavelength (X), signal modulation, Its trajectory and
its coherence.
Signal modulation is a common characteristic of all active systems. The objective of modulation is
to provide a means of comparing the source signal with the backscatter. This comparison ensures
precise measurement of the distance covered by the backscatter signal and the ratio of the energies
of the source signal and the backscatter signal. The latter gives Information about the properties of the
Irradiated surface.
24 Processing of Remote Sensing Data

For observing surfaces situated at large distances, time variations in signal amplitude are used.
An example is the pulse radar, which has an emission duration of a few microseconds (Fig. 1.15).
This type of modulation is not possible in small range (0 to 250 m) observations for fixed-post
scatterometers, since the shift between the source signal and the backscatter signal is too small to be
measurable. In such cases, another system of reference known as frequency modulation is employed:
the source energy remains constant and the frequency varies from one cycle to another (Fig. 1.16).
In this case, variation of frequency with time takes place in a saw-toothed pattern (Fig. 1.17), with
a frequency shift of a few megahertz, around the central frequency Fq.
Whatever be the type of modulation used, the path of the signal in the source-receiver system
can be schematically represented as follows (Fig. 1.18).
A microwave signal Is produced by the source (1). Through a series of amplifiers and multipliers,
the signal is modified to the chosen frequency. A fraction of the source signal (6) is directly sent to the
mixer (5) and acts as the local oscillator. Another part of the signal (2) is routed through the transmission
system (3) with a given polarisation. The transmission antenna irradiates the target surface. A part of
this signal is backscattered towards the receiver antenna (4) and combines In the mixer with the
fraction (6) of the source signal. This mixed signal is sent to the analyser (7) to determine the shift

. Signal amplitude

w w f — f
— f —
T
Time

Fig. 1.15: Pulse modulation.

Fig. 1.16: Frequency modulation.

A Frequency

Fig. 1.17: Triangular modulation.

 Data Sources 25

To analysing
system

Fig. 1.18: Signal path.

between the two signals. This shift corresponds to the time taken by the signal to travel the distance
2F?+ Pi + P2 + P3. The value of Oq is obtained in this way.
The source wave, characterised by its wavelength, polarisation and modulation, is a coherent
monochromatic wave. This wave has totally different physical effects compared to those of incoherent
light waves used in passive systems.
In fact, interaction of coherent monochromatic radiation with the surface produces interferences
during scattering in all directions in space. Very strong local variations In energy are observed while
recording the backscatter signal Pbs- this is referred to as fading. To compensate for this major
disadvantage of radars, averages of several measurements need to be taken:
— at the same frequency but over uncorrelated surfaces (i.e., independent samplings),
— at several frequencies but over the same surface.
Choosing the optimum between the time of integration of various measurements and multiplicity
of less accurate but independent determinations of oq , in other words, choosing between enhancement
In the accuracy of measurement (in decibels: dB) and improvement In the spatial resolution of the
system, is a difficult task.

■ Conclusion
Logistic parameters affect the energy of the backscatter signal. These parameters should be available
to the user so that he can concentrate on the information derivable from Oq about the object or medium
under study. That is why research teams are at present engaged in two different aspects of Investigation:
— analysis of elementary processes involved In the backscatter of the microwave signal. This
refers to the study of Interactions of radiation with matter on a large scale, either in the laboratory or in
the field with scatterometers.
— use of systematically acquired Image data.
This assumes knowledge of the manner in which the surfaces react to pulses received from a
radar depending on the physical properties which per se depend on two types of parameters: geometric
(angle of Incidence, roughness) and dielectric parameters corresponding to the nature of the material
constituting the surface.

1.6.8 Analysis of physical processes of backscatter

■ Geometric parameters
□ Angle of incidence
The angle of Incidence in which Oq is measured has a very important role. As shown in Fig. 1.19, the
backscatter Intensity, which Is very high at near vertical incidences (10^-20°), gradually decreases at
steep incidence angles (60° and above).
26 Processing of Remote Sensing Data

dB

□ Roughness
Natural terrestrial objects, In particular plant canopies, often represent Inhomogeneous media. The
incident wave is transmitted into such media and scattered by various components. The backscatter
signal is related to the characteristics of the volume concerned. However, the volume scattering is
very low compared to surface scattering. Most often, surface scattering is studied In remote sensing.
Surface scattering depends on two parameters: angle of incidence of waves and roughness of
the surface under observation.
It should be remembered that the geometric parameters of the objects affect the measured signal
only if their dimensions are not too different from the wavelengths used. The same surface behaves
differently depending on the condition of Its surface and the wavelength of the incident radiation (Fig.
1. 20 ).
Surface roughness originates from several sources. Its significance varies depending on the type
of surface: bare soil (cropping activity), vegetation cover (roughness related to growth pattern of plants
and their spatial arrangement) or water body (roughness of waves more or less related to the direction
and force of wind). Surface roughness is usually expressed by two parameters, viz.. Root Mean Square
deviation {h) of surface irregularities and a measure of horizontal dimension of roughness. Such a
formulation of surface roughness, widely used for bare-soil and water bodies, poses some problems
in the case of vegetation. In fact, height and distribution of vegetation, orientation and dimensions of
leaves, alignment of trees etc. are difficult to understand and more so when they are natural canopies.
Theoretical models have been proposed for cultivated plants, approximating them to moist cylinders.
However, morphological modifications undergone by plants during their growth and development should
be Integrated into such models. Further, ‘mixed’ surfaces, very common. Introduce additional difficulties.
Even a feeble coverage of bare soil by crop may suppress the effects due to soil. On the other hand,
difficulty is often experienced in observing plants in the course of germination since the soil completely
masks their contribution to the backscatter signal. This phenomenon is produced in the case of wheat
until it covers at least 5% of the total surface.
Roughness is reported relative to the wavelength (X) of the source radiation by the formula h <
X/8 sin Y, where y is the depression angle.
— A smooth surface, relative to the wavelength used {h < X/8 sin y ), reflects the incident radiation
as a specular surface (see Figs. 1.3 and 1.4), and Gq is zero or very low.
— On a more or less rough surface {h > X/S sin y ), the reflection properties of the surface gradually
change from a specular reflection to diffuse reflection, which follows Lambert’s law (see Figs. 1.3 and
1.4), and gq increases with increasing roughness. The specular component proportionately decreases
and only diffusive reflection remains.
 Data Sources 27

State of the surface Behaviour to microwave

signal
RMS
X Band L Band

1 cm

5 cm

17 cm

30 cm

Fig. 1.20: Reflection behaviour of a surface depending on the wavelength of the incident radiation
(after King, 1979).

The following example illustrates this effect. Measurements of gq were carried out on a cultivated
soil, successively subjected to three types of agricultural operations, viz., tilling, clawing and harrowing,
using a scatterometer at wavelengths of 25 cm (1.5 GHz) and 6.6 cm (4.5 GHz). At 1.5 GHz, the effect
of roughness was observed irrespective of the value of the angle of incidence and the tilled surfaces
Indicated gq higher than the harrowed surfaces (Fig. 1.21). At 4.5 GHz, the effect of roughness Is
negligible, in particular between the incidence angles of 20° and 60° (Fig. 1.22).
Measurement of gq gives information on roughness but the choice of wavelength relative to the
degree of roughness studied is very important.
In littoral environments, on exposed tidal flats, sedimentary facies with large gradients in grain
size and hence roughness (generally coarse on high zones and fine on low zones) and moisture
content (low In high zones and high in low zones) are observed depending on hydrodynamic activity.
Studies of tidal flats observed by SAR of ERS-1 showed that radar measurements are sensitive
to the grain size of the sediment. In fact, application of the formula {h < X/8 sin y ) to the SAR-ERS-1
data (;^ = 5.6 cm, 0 = 23 at the centre of the scene) showed that roughness measurement is significant
for h = 0.8 cm. This result indicates the possibility of discriminating sedimentary facies such as sands,
gravel and pebbles. On the other hand, during very calm periods (sea with no wind), in the presence
of vast open tidal flats comprising mud and sandy-mud facies, the land-sea contact cannot be clearly
differentiated, as the two targets have almost equal coefficients of roughness.

■ Dielectric parameters
Dielectric constant (s) and permittivity are the two components of dielectric properties of matter.
They are related to the volume of a body and are dependent on its water content and the wavelength
28 Processing of Remote Sensing Data

Fig. 1.21: Variation of a 1.5-GHz signal backscattered by a bare soil (moisture content by weight 12-16%) with
surface roughness (after King, 1979).

Angle of
incidence

Fig. 1.22: Variation of a 4.5-GHz signal backscattered by a bare soil (moisture content by weight 12-16%) with
surface roughness (after King, 1979).

used (Table 1.5). That is why, even small quantities of free water In any material, air and soil or vegetation
and soil, produce considerable variations in its dielectric constant. The mineralogical and chemical
properties of the objects observed also influence the dielectric constant, but their effects are masked
by the predominant effect of water.

Table 1.5: Relative dielectric constant of soil and water for a frequency of 10 GHz ( / = / I T )
(after Ulaby et al., 1981).

Material Relative dielectric constant

Dry loamy soil E = 3.5 - 0.4 j

Humid loamy soil E = 17.9-7 .2 y
Water E = 54.4 - 36.8 ]
 Data Sources 29

Influence of humidity of a soil on Oq was demonstrated by fixed-post experiments using a

scatterometer. For example, at 1.5 GHz for a clawed soil, the backscatter signal increases with surface
humidity (H = water content, by weight, of the top 2 cm) irrespective of the angle of incidence (Fig.
1.23).

Fig. 1.23: Variation of backscatter signal (1.5 GHz) with surface moisture content for a clawed soil
(after King, 1979).

To establish a general relationship between Gq and moisture content of a soil or plant canopy, the
nature of moisture needs to be defined:
— ^for a soil, be it gravitational water, absorption water or bound water, and the soil thickness
affecting the gq measurement;
— for a vegetation, its water content varies with a number of parameters: vegetation growth stage,
soil water content etc.
Having established the influence of moisture and roughness on gq , investigations were undertaken
to determine the favourable range of frequencies for studying moisture alone. In the case of bare soils,
correlation between moisture and gq was found to be better than 70% for frequencies in the range of
4.5 GHZ to 6.0 GHz, with a maximum at 4.75 GHz. Although moisture and roughness are related, it is
the water content, less prone to possible errors, which is estimated. Measurements should be carried
out with an incidence angle of 10° for the results to uniquely correspond to the water content. This is
possible for fixed-post scatterometer measurements but not for radars, since only a small fraction of
images corresponds to this angle.
In the case of soils covered by vegetation, effect of water content in the underlying soil on the
signal backscattered by plants was observed at 4.3 and 7.5 GHz In VV polarisation. Vegetation on a
humid soil produces a higher gq than the same vegetation on a dry soil. Plant cover increases the
surface roughness but decreases the contribution of the soil to the backscatter signal. It is necessary
to evaluate the effect of vegetation cover on scattering at microwave frequencies. In view of the difficulty
In quantifying the geometry of a plant, Ulaby and Bush showed that knowledge of moisture content
suffices to determine its stage of growth. They thus developed a model describing the backscatter
effect as a function of only this parameter and considering the vegetation as a cloud of scatter elements
having an intrinsic moisture.
Many studies were carried out on the effect of snow on microwave signals. Two types of snow are
distinguished: dry (powdery) snow, which contains no water in liquid form, and wet snow containing a
certain quantity of free water. Surface roughness little Influences gq in the case of dry snow and
30 Processing of Remote Sensing Data

significantly affects erg of wet snow. The backscatter effect of snow increases with increase in its
thickness. For a given type of snow, it increases with frequency of waves, while the influence of the
angle of incidence becomes negligible for 35.5 GHz (Stiles et al., 1981).This is the reason for studies
on snow at frequencies close to the latter.
In the case of sea-water, the radar wave penetrates only a few millimetres from the surface due to
the high dielectric constant of water. Hence, one of the predominant factors Influencing backscatter Is
surface roughness.The roughness directly related to agitation of the sea surface can be observed and
measured. It is the resultant of several phenomena, viz.:
— local action of wind on the surface (capillary waves);
— sea waves, whose direction and wavelength are modified by coastal morphology and bathymetric
level;
— currents, whose directions and velocities are influenced by bathymetric level, especially in
zones where the latter is small, i.e., close to the coast;
— presence of a surface pollutant (see section 25.3.4, the case of hydrocarbons).
These phenomena, Interdependent, represent complex models, making interpretation of SAR
images of coastal marine environment difficult. Nevertheless, several factors, in particular those related
to waves and sea rollers, can be extracted from the marine data obtained by synthetic aperture radar
such as those mounted on the satellites SEASAT, ERS-1 and ERS-2. Such measurements are
independent of meteorological conditions (cloudiness, rainfall etc.) and accessible during the day as
well as night.
Radar measurements can be used to determine a surface effect that may be related to the sea­
floor topography. Thus, hydraulic dunes, which characterise the floor of the Calais Strait, were detected
by radar image. The method requires a shallow zone under the effect of strong hydrodynamic activity.
It may be emphasised that this approach is applicable only in special conditions.
Sea-ice is one of the Important materials studied by microwave remote sensing in Arctic and
Antarctic regions. The ice in such regions contains salt, unlike pure ice, but less than that of normal
sea-water. Consequently, ions are distributed across the Ice, and variations in their concentrations are
represented as variations in the backscatter signal. The properties of sea-ice are dependent on the
salinity of the original sea-water, and temperature, pressure and porosity of Ice. The structure of ice
varies according to the rate of its formation, its age and history. Melting of ice in spring and summer
produces cavities inside the ice mass through which water may escape. An ice of first year (less than
1 year of age) melts faster than an older and less saline ice. The salinity and temperature of an ice vary
with depth and its properties also change in the course of time, leading to a modification of its dielectric
properties. The direction and velocity of wind and currents, number of freeze-defrost cycles, extent of
snow cover on the surface etc. cause variations In thickness and surface roughness of ice and lead to
variations in gq .

■ Penetration properties
Penetration properties of microwave signals were indicated in the preceding examples. These signals
effectively penetrate a certain depth of the dielectric (soil or vegetation) before being completely
absorbed. Depth of penetration a is defined as the depth at which the source wave is attenuated to
about 1/3. In practice, this depth varies from a few cm to 1 m. Very few cases of large penetration,
greater than a metre, were observed for highly homogeneous and dry desert sands.

■ Conclusion
In addition to understanding the elementary processes, field measurement campaigns have witnessed
two kinds of development:
— defining the most suitable airborne or satellite sensors (with minimum influence of their
operational parameters on Oq in the ratio Ps^Pbs)^
 Data Sources 31

— formulation of predictive models or simulation of responses of objects according to their state

and characteristics. These models ought to be gradually generalised to larger areas than the test
zones.
These Investigations have thus brought together a number of signal physicists and thematic
specialists engaged in description of the object of their study. In their respective domains, they adjust
themselves to the two languages of the model and the reality.

1.7 GENERAL CONCLUSION

The physical basis described above enables us to understand the characteristics of different data
acquisition systems In remote sensing as well as the interactions between radiation and natural surfaces.
The latter, in the case of visible and reflective near- and middle-infrared, are illustrated in Chapter 4
dealing with spectral characteristics.

References
Actes de télédétection infrarouge thermique, 1993. Des échanges énergétiques et hydriques de la végétation en
combinaison avec d’autres capteurs. La Londe les Maures, 20-23 sept. 1993. Cemagref, CETP, Pennstate,
330 pp.
Barton J. 1989. Comparison and optimisation of AVHRR Sea Surface Temperature algorithms. J. Atmospheric and
Oceanic Technology, 6-89.
Becker F. 1978. Physique fondamentale de la télédétection, In: École d’été de Physique spatiale; Principes physiques
et mathématiques de la télédétection. CNES, pp. 1-108.
Becker F. 1987. The impact of spectral emissivity on the measurement of land surface temperature from satellite.
Int. J. Remote Sens., 8:1509-1522.
Becker F, Li Z.L. 1995. Surface temperature and emissivity at various scales: definition, measurement and related
problems. Remote Sensing Reviews, 12; 225-253.
Bonn F, Rochon G. 1992. Précis de télédétection, 1: Principes et méthodes. Presses de l’Université du Québec/
AUPELF, 485 pp.
Chanzy A. 1991. Modélisation simplifiée de l’évaporation d’un sol nu en utilisant l’humidité et la température de
surface accessibles par télédétection. Thèse de Doc. Ingénieur INAPG, 210 pp. et annexes.
Courault D, Clastre P, Guinot J-P, Seguin B. 1994. Analyse des sécheresses de 1988 à 1990 en France à partir de
l’analyse combinée de données satellitaires NOAA-AVHRR et d”un modèle agrométéorologique, Agronomie,
14:41-56.
Deschamps P-Y, Duhaut P, Rouquet M-C, Tanre D. 1984. Mise en évidence, analyse et correction des effets
atmosphériques, sur les données multispectrales de Landsatou de SPOT In: II®Coll. Int. Signatures spectrales
d’objets en télédétection. Bordeaux, 12-16 sept. 1983. Éd. Inra PubI, Les colloques de l’INRA, n°23, pp.709-
722.
Guyot G. 1997. Climatologie de l’environnement. De la plante aux écosystèmes. Masson, Paris, 505 pp.
Jackson RD, Reginato RJ, Idso SB. 1977. Wheat canopy temperature: a practical tool for evaluating water
requirements. W ater Resour. Research, 13 (3): 651-665.
Kerr Y. 1991. Corrections atmosphériques dans l’infrarouge thermique. Cas de l’AVHRR. 5® Coll. int. Mesures
physiques et signatures en télédétection. Courchevel, ESA SP 319, pp. 29-34.
King Ch. 1979. Contribution à l’utilisation des micro-ondes dans l’étude des sols. Thèse INAPG, 122 pp.
Kneizys FX et al. 1983. Atmospheric transmittance/radiance; computer code Lowtran 6. Technical report AFGL-TR-
83-0187, Optical Physics Division, US AIR Force geophysics laboratory, Hanscom airforce base, MA, États-
Unis.
Lagouarde J-P. 1991. Use of NOAA AVHRR data combined with an agrometeorological model for evaporation
mapping. Internat. J. Remote Sensing, 12 (19): 1853-1864.
Lagouarde J-P, MacAneney KJ. 1992. Daily sensible heat flux estimation from a single measurement of surface
temperature and maximum air temperature. Boundary, Meteorology, 59 (4): 341-362.
32 Processing of Remote Sensing Data

MODTRAN 3,1995. User Instructions and Comments, version 1.3. Etats-Unis.

Norman JM, Becker F. 1995.Terminology in infrared remote sensing of natural surfaces. Remote Sensing Reviews,
12:159-173.
Ottlé C, Vidal-Madjar D. 1992. Estimation of land surface temperature with NOAA9 data. Remote Sensing of
Environment, 49 (1 ): 27-41.
Ottlé C, François C. 1993. Atmospheric corrections of ATSR-IR data. Actes de télédétection IRT. La Londe les
Maures, 2-23 septembre CEMAGREF. CETP, Pennstate, Carlson et al. (eds), pp. 79-82.
Paquet G. 1997. Detection électromagnétique: fondements théoriques et applications radar. Masson, Paris, 320
PP-
Perrin de Brimchambaut C. 1985. Bilan thermique de la Terre. Encyclopedia universalis, vol. 3, pp. 612-613.
Polidori L. 1997. Cartographie radar. Gordon and Breach Science Publishers, Canada, 287 pp.
Privette J-L, Fowler C, Wick GA, Baldwin D, Emery WJ. 1995. Effects of orbital drift on advanced very high resolution
radiometer products: normalized difference vegetation index and surface temperature. Remote Sensing of
Environment, 53:164-171.
Seguin B, Itier B. 1983. Using midday surface temperature to estimate daily evaporation from satellite thermal IR
data. Internat. J. Remote Sensing, 4 (2): 371-383.
Seguin B, Baelz S, Monget J-M, Petit V. 1982. Utilisation de la thermographie IR pour l’estimation de l’évaporation
régionale. Il: Résultats obtenus à partir de données satellites. Agronomie, 2 (2): 113-118.
Stiles WH, Ulaby FT, Fung AK, Asiam A. 1981. Radar spectral observation of snow. IGARSS’81 Digest, Washington,
pp. 654-668.
Stoll MP. 1988. Mesures de la température et de l’émissivité de surface par télédétection: modèles et méthodes.
Télédétection spatiale, aspects physiques et modélisation. CNES, école d’été, pp. 845-904.
Ulaby FT, Moore RK, Fung AK. 1981-1982. Microwave Remote Sensing Active and Passive, vols. 1-2. Addison-
Wesley Cy. 1064 pp.
 2
Sensors and Platforms

2.1 SENSORS
In remote sensing, various sensors are used to measure in a given wavelength band the radiance of
objects under study. Conventionally two types of sensors, viz., active and passive, are employed.
Active sensors transmit a signal and receive a part of it returned by the objects. Examples of such
sensors are RADAR (Radio Detection And Ranging), LIDAR (Light Detection And Ranging), LASER
(Light Amplification by Stimulated Emitted Radiation), fluorometers etc. Passive sensors receive the
energy emitted or reflected by the objects (radiometers, cameras, spectroradiometers); the main source
of energy in this case is the Sun.
The principle of the sensors Is as follows: radiance is measured within the field of view of a sensor
over a given surface, i.e., over the pixels whose dimensions are determined by the solid angle of the
sensor and its altitude. The altitude evidently depends on the platform employed, while the field of
view, solid angle and wavelength bands in which measurements are made depend on the receiver.
Let us consider, as an example, a radiometer (Isco) (Fig. 2.1) mounted on a support at a height of
1.4 m with a solid angle of 90°. At this height, the area covered on the ground by the detector is a little
more than 6 m^, which represents the area of a pixel. As the receiver in this case is on a fixed mount
that does not move, we define only a pixel (short form of picture element) but not a field of view. This
radiometer has twenty wavelength bands. The EXOTECH radiometer has only four bands corresponding
to the MSS sensor of the LANDSAT satellite. The CIMEL radiometer consists of three bands
corresponding to the HRV sensor of SPOT satellite.
The photographic systems will not be specially discussed in this book even though they continue
to be widely used. It has become common to digitise photographic data and process it as an image.

Fig. 2.1 : Schematic diagram of the Isco radiometer.

34 Processing of Remote Sensing Data

However, care has to be taken to digitise the data in coherence with the grain of the film. The special
problems posed by the geometry of measurement are different for aerial photographs and satellite
photographs. The latter is easier to rectify and restore (Chapters 13 and 14).

2«1.1 General scheme

Various types of sensors are constructed based on a similar scheme. They comprise several components
ranging from signal reception to signal storage (Fig. 2.2), viz.,
— system to receive radiation from the pixel and a telescope (objective),
— calibration source and spectrometer,
— amplifier and recording system.
The sensor thus represents an imaging radiometer, often called scanner, essentially consisting of
a radiometer supported by a system of image acquisition, pixel by pixel.
The electromagnetic radiation from a pixel received on a primary mirror and transmitted to a
secondary mirror, reaches the recording system through an optical fibre after being amplified and
calibrated (Fig. 2.3). The area of a pixel is determined by the receiver system and the sensitivity of the
detector. The angles of instantaneous field of view are very small and consequently the energy reaching
the sensor is very low, about 10,000 times lower in a satellite system than in an airborne system.
Another limiting factor is the duration of exposure, which varies inversely with the square of instantaneous
field of view. The exposure times are of the order of a nanosecond to a millisecond depending on the
receiver system.

Radiation coming Calibration source Detectors

from the surface
under investigation

Fig. 2.2: General scheme of a digital sensor.

Total field of view (11.6°)

Fig. 2.3: Configuration of a multispectral scanner (LANDSAT MSS).

 Data Sources 35

The amplifier is essential for enhancing the signal, which most often is very weak. The energy
received is converted into a digital signal (with discrete values expressed as number of bits: between
6 and 12) by a recording system, which varies according to the detectors.

2.1.2 Radiation receiving systems

Two types of radiation receiving systems are presently In vogue: whiskbroom scanners and pushbroom
scanners.

■ Whiskbroom scanners
The whiskbroom or across-track scanner system was first used in LANDSAT-1 satellite in 1972, as
well as in several other scanners employed In airborne units (Fig. 2.4).

Direction of

a; Angie of Instantaneous Field of View (IFOV) p: Look angle

y; Total field of view

Fig. 2.4: Principle of whiskbroom scanning.

Scanning is carried out by means of a rotating or oscillating mirror, inclined at 45° on the vertical
and situated perpendicular to the direction of motion of the platform. The energy coming from the
Earth’s surface Is received in a solid angle, which determines the instantaneous field of view (IFOV).
The IFOV is about 10"^ to 10“ ^ radian. This angle, which is constant for a given scanner, determines
the area of the ground pixel for a given altitude of the platform that carries the scanner. This constitutes
the geometric resolution of the sensor.
The image received consists of a series of strips limited in width by the field of view of the receiver
and partly by the movements of the mirror. Successive scan rows are covered as the receiver and the
platform on which it is installed move forward. In order to obtain a complete coverage without gaps, the
velocity of the satellite, its altitude and the rate of rotation of the mirror need to be perfectly synchronised.
When a rotating mirror is used (for example, on an aircraft, the Daedalus scanner with a field of
view of 120°), the ground signals are received during a part of its rotation and no signal Is received
during the other part. This time is utilised to calibrate the receivers by means of standard targets; in
particular, a blackbody is employed to record information in thermal bands. When the FOV is small, as
in the case of satellite-mounted sensors (for example, 11.6° for LANDSAT-MSS or 14.8° forTM ),
oscillating mirrors, whose standards are situated on either side of the zone of oscillation, are used.
36 Processing of Remote Sensing Data

The geometry of whiskbroom scanners should be fixed precisely since it affects the restoration
and rectification of images. Depending on whether the pixel is situated at the nadir or laterally on a
scan line, the value of the side of the square, i.e., the geometric resolution, varies (Fig. 2.4). This gives
rise to image distortions; however, such distortions are quite small for satellite-borne sensors, which
have very small solid angles. As for the radiometric aspect, the atmospheric effects are greater since
the atmospheric thickness traversed is larger for an inclined view than for a vertical view (Fig. 1.6).
Geometric rectification of images eliminates such systematic errors. However, before undertaking
geometric rectification of images, it is necessary to evaluate the relative magnitude of errors produced
by geometric distortions and the errors that could be introduced by various processing techniques or
by visual interpretation of images. For example, the Interpretation accuracy for delineation of landscape
units is less than 100 m.Thematic errors for scales between 1 / 1 ,000,000 and 1 / 1 00,000 are generally
greater than geometric errors. Evidently, if thematic information Is to be incorporated into a geographic
information system (GIS), It Is essential that the image be accurate enough to be superposed on the
GIS map and geometric corrections need to be applied (Chapter 13).

a Pushbroom scanners
The pushbroom or along-track scanner system is more recent and uses no scanning mirror. It consists
of a linear array of 1728 to 12,000 CCD (charge-coupled device) detectors, which simultaneously
receive information from 1728 or 6000 pixels aligned In a single row (Fig. 2.5). Each detector converts
the detected radiance into an electric signal through storage of electrons in holes created on the
surface of a semiconductor. The charge thus collected Is transferred to an electric circuit wherein the
current is amplified and stored on a magnetic base.
Each detector corresponds to a separate pixel. Since all the detectors are identical, no deformation
occurs due to optics of the Instrument, unlike in photographic units. Thus for every 1 1/2 millisecond,
6000 values are obtained corresponding to a ground swath 10 m wide and 60 km long in the case of
SPOT. Each CCD detector measures only 13 pm x 13 pm. They are grouped Into an array of 1728
detectors, which together form a bar of about 2.25 cm and do not occupy much space. All the detectors
scan a row, which Is generated by the advancement of the platform: consequently, this system is
called ‘pushbroom’.
Each detector is characterised by Its own transfer function and has to be calibrated Individually.
As the detectors are identical, the solid angle that controls the instantaneous field of view Is also

Fig. 2.5: Principle of a pushbroom scanner.

 Data Sources 37

identical; the dimensions of the pixels are slightly larger on the edges of the row since the system has
a conical projection (Chapter 13).
An advantage of the pushbroom scanner is the greater duration of exposure than In the scanning-
mirror system for the same platform velocity. For the pushbroom scanner:

D = p/v

where D is the exposure time (in seconds), pthe pixel size (in m) and vthe velocity of the satellite (in
m/s). Thus, for SPOT, D = 10 (m) / 6660 (m/s) = 1504 milliseconds.
In the scanning-mirror system:

D = p/nv

where D is the exposure time (in seconds), p the pixel size (in m), n the number of pixels per scan-line
and vthe velocity of the satellite (in m/s).
As the number of pixels on LANDSAT per line is 3000 with p = 30 m, the exposure time is 1000
times shorter. The pushbroom scanners perform better from the radiometric point of view. Each array
of detectors corresponds to a spectral band. The pushbroom scanner is employed in HRV and MOS-
1 sensors.

2.1.3 Spectral bands

It is important to arrange the detectors in different narrow spectral bands in order to detect specific
phenomena of certain spectral domains.
A detector should receive a certain minimum energy to be active. Hence, if a wide band, such as
the panchromatic from 0.5 to 0.7 pm, is used, the quantity of energy Is greater than that for a narrow
band and a pixel of smaller area can be taken. An inverse proportionality exists between geometric
resolution and spectral resolution. If the geometry of a pixel is changed, the energy in each resolution
element is changed by the same value. Consequently, a geometric modification leads to a modification
of the radiance and vice-versa. It is, therefore, better to apply geometric corrections after processing
than before so that the radiometric signal is not disturbed.
Different sensors have a variable number of spectral bands, generally three to ten (CD 2.1).
Some sensors, employed on the ground or airborne, have more than 100 spectral bands. It is hence
necessary to select bands that are most suitable for a given thematic application.

2.1.4 Airborne systems

At the time of launching of the first Earth resource satellites, it could have been thought that the era of
airborne systems had ended. But it did not. Parallel to the pursuit of the conception and launching of
satellites, a continuous development of airborne remote sensing missions with special sensors has
been witnessed (Table 2.1).
There are three reasons for developing airborne remote sensing. Firstly, It is essential to acquire
data at an intermediate level that enables a better understanding of certain features through the
relationship between the level of acquisition and the level of organisation (Chapter 15). Secondly, in
the case of regional studies covering a limited territory. It may be advantageous to investigate a specific
problem through an airborne mission (more economical). Such is the case, for example, in evaluation
of damages caused to agriculture by a drought or a devastating flood. Finally, the possibility of using
narrow spectral bands, of a few nanometres to a few tens of nanometres wide, in the spectral domains
for which high atmospheric absorption makes use of satellite data ineffective.
38 Processing of Remote Sensing Data

Table 2.1: Examples of airborne spectrometers

Characteristics AVIRIS Casi

Technology CCD array measuring all the CCD array. In spectral mode, selection
channels for 1 pixel; mechanical of 39 columns of the array, information
sweep of lines acquired for 288 bands. In spatial
mode, selection of 15 bands,
Information acquired for 512 pixels.
Date of commencement of
service 1987
Spatial resolution 20 X 20 m at 20 km altitude 2.3 X 2.7 m in spatial mode, 2.3 x 9.5 m
in spectral mode at 1860 m
Swath 12 km FOV 35° 5
Number of spectral bands 224 288
Width of spectral bands 10 nm 1.8 nm
Wavelength range 410 to 2450 nm 430 to 870 nm

Absolute positioning of the aircraft in regions where no ground control points are available and
high costs have long been the major handicaps in using airborne remote sensing. Application of GPS
(Global Positioning System) in association with the American GLOSSNAR satellites has resolved this
problem.

2.2 PLATFORMS
Any moving vehicle that can carry a sensor can be considered a platform for remote sensing. Thus,
several types of platforms can be identified;
— those operating at a height of a few metres from the ground: cranes or other vehicles that
support radiometric or photographic equipment;
— those operating at a height of about ten metres to ten kilometres: aeroplanes, helicopters and
balloons;
— those operating between ten and hundred kilometres: stratospheric balloons;
— those operating between 200 and 40,000 km: satellites, manned or unmanned, subjected to
terrestrial gravity. The latter are the most common for observation of the Earth.
Airborne platforms, such as aeroplanes or stratospheric balloons, are not discussed here, as
these were described in our book of 1975 or in the book by Bonn and Rochon (1992). Only Earth
observation satellites are described in this book. Orbital characteristics of the satellites launched to
study other planets of the solar system are also not considered. Only some orbital data of satellites
are reviewed In the following section. Illustrations and details regarding the geometry of remote sensing
can be found in Lliboutry (1992).

2.2.1 General principles of orbital motion

Two masses E (Earth) and S (satellite) separated by a distance d attract each other with a force F =
g ES/d^, where g \s the gravitational constant. Computed by Newton In 1687, this force controls the
motion of the two bodies according to the three Kepler laws established in 1609.
 Data Sources 39

First law
A satellite moves along an elliptical orbit around the Earth, which occupies one of the foci of this
ellipse (Fig. 2.6). This orbit is described by its semimajor axis a and eccentricity e = c/a, where c is the
distance of a focus of the ellipse to the perigee. Perigee is the position of the satellite when it is closest
to the Earth (818 km for SPOT) and apogee the position when it is farthest from the Earth (833 km for
SPOT). In order to avoid deceleration due to the atmosphere and falling on the Earth, the perigee of a
satellite ought to be greater than 200 km above the ground.

Perigee

Fig. 2.6: Satellite orbit according to Kepler’s first law (after Gilliot, 1994).

■ Second law
The satellite does not travel along the ellipse with a constant linear velocity but in such a manner that
the radius vector joining the satellite to the Earth sweeps a constant area in a given time (Fig. 2.7).
Hence when the satellite is at Its perigee, its velocity is greater than when it is at the apogee.

m Third law
The period P(in years) of a satellite is related to the semimajor axis a expressed in astronomical units
(1 a.u. = Sun-Earth distance) and to the masses of the Earth (E) and satellite (S):
(E + S) P2 =a3

A is the area described by the radius vector between the points a and a ’ in a time t
B is the area covered by the radius vector between the points b and b ’ during the time t
Fig. 2.7: Satellite velocity according to Kepler’s second law (after Gilliot, 1994).
40 Processing of Remote Sensing Data

2,2.2 Orbits
The plane of the elliptical orbit is known as the orbital plane of the satellite (Fig. 2.8). The angle
between the orbital plane and the equatorial plane defines the angle of inclination of the satellite. The
line of intersection of the two planes Is the nodal axis (or line). A node hence corresponds to the
intersection of the equator with the path of the satellite (orthogonal projection of the satellite orbit onto
the Earth).
In Earth-observation remote sensing, circular, geostationary and sun-synchronous orbits are
important.

Fig. 2.8; Orbital parameters of a satellite (after Gilliot, 1994).

■ Any circular orbit

If the eccentricity of the ellipse becomes zero, the orbit becomes circular. The satellite revolves around
the Earth at a constant altitude. The minimum velocity of the satellite at an altitude of 200 km, i.e., for
a = 6500 km, would therefore be 7.77 km/s and the periodicity about 1 h 30 min.
At 200-km altitude, at which military observation satellites operate, the temperatures are of the
order of 750° C plus or minus 300° C. These temperatures are essentially generated by the kinetic
energy of the monoatomic molecule of oxygen, the most abundant element between 120 and 1000
km of altitude. The aerodynamic drag around 200 km limits the lifespan of a satellite to a few months.
In order to remain in orbit, observation satellites need to be operated at about 800 km of altitude.
The orbits of the satellites TIROS, NOAA and ERS-1 are circular.

■ Geostationary orbit
A geostationary orbit is one in which the satellite Is always situated at the zenith of a point above the
Earth’s equator. Hence the orbital plane coincides with the equatorial plane and the orbit Is circular.
The satellite revolves with the same angular velocity as that of the Earth and therefore is fixed in a
terrestrial reference frame. Kepler’s third law gives the radius R of the orbit as 42,164 km, or 35,786
km above the equator (the equatorial radius of the Earth being 6378 km). The satellite velocity vin its
orbit is thus v= 2nR/t, where f = 1 d = 86,164 s and v= 3.07 km/s or 11,063 km/h.
The satellites METEOSAT (France), GOES (Geostationary Operational Environmental Satellites,
USA), GMS (Japan) and INSAT (India) are geostationary.
 Data Sources 41

Sun-synchronous orbit
A satellite in a sun-synchronous orbit always passes over the same point on the Earth at a given local
solar time. The orbital plane of the satellite hence remains constant relative to the orbital plane of the
Earth around the Sun. Consequently, the nodal line makes a constant angle with the axis joining the
centres of the Earth and the Sun.
The solar day is defined as the duration in which a given point on the Earth returns to the same
position relative to the Sun. By definition, the solar day is equal to 24 hours. The sidereal day is defined
as the period of rotation of the Earth relative to a fixed point in space (vernal point). This period Is
equal to 23 h 56 min 4 s. Consequently, the Earth moves around the Sun in 365.2425 sidereal days. In
order to be synchronous, therefore, the orbital plane of the satellite should rotate by 360° in 365.2425
days, i.e., by 0.9856° per day (Fig. 2.9). This corresponds to an angular velocity of retrograde precession.
Hence, a sun-synchronous orbit can be defined as an orbit having an angular velocity of precession
equal to the angular velocity of the Earth around the Sun.

Fig. 2.9: Displacement of the orbital plane of a sun-synchronous satellite (adapted from Gilliot, 1994).

For a circular orbit at an altitude of 832 km, as in the case of SPOT, the inclination should be
98.7°. Consequently, the north and south poles are never covered. Hence for each pole there exists a
circle of precession with its centre at the pole and a radius of 966.6 km for which no images can be
obtained. Otherwise, the entire Earth can be covered by remote sensing images.
The relatively low altitude of sun-synchronous satellites facilitates a good resolution. However, as
these satellites are subjected to deceleration by terrestrial atmosphere, the orbital parameters need to
be corrected periodically. Similarly, the altitude of the satellite has to be permanently controlled.
In the case of sun-synchronous orbits, the azimuth of the Sun is constant throughout the year but
the height of the Sun varies. This gives rise to shadows of varied lengths but in the same direction.
The satellites NIMBUS, LANDSAT and SPOT have a sun-synchronous orbit.
Thus satellites with any circular orbit or sun-synchronous orbit are distinguished from geostationary
satellites (Fig. 2.10).
In the case of unmanned satellites, sensors are mounted on the platform once and for all. The
conditions of Image acquisition are determined by these two entities, viz., platform and sensors, which
hence constitute a single system. To understand the interactions between sensors and platform, let us
study the SPOT system (also see text on the CD giving details of the SPOT system and SPOT 4 in
particular) and the data on other systems.
42 Processing of Remote Sensing Data

Fig. 2.10: Relative positions of sun-synchronous and geostationary satellites.

2.2.3 SPOT system

The SPOT satellite has a sun-synchronous, quasipolar (98.7°) and circular orbit with an eccentricity
less than one per thousand. It acquires images in a descending path during the day, i.e., while going
from the north pole to the south pole, and nominally passes through the equatorial node at 10 h 30 min
on 15 June. The duration of passage above a given region is more or less constant at about 15 min.
The velocity of the satellite relative to the ground is 6.6 km/s or 23,760 km/h.The average duration
of one revolution is 101.4 min. During this revolution, the Earth rotates by 25.35° or 2823 km at the
equator (whose perimeter is 40,075 km). This constitutes the interval between chronologically
successive paths. The orbits and the paths converge at the poles. The spacing between adjacent
paths is 108.6 km at the equator, 76.7 km at 45° latitude and 16.5 km at the circle of precession.
The satellite thus makes 14 and 5/26 revolutions in 24 hours (Fig. 2.11). It covers the same point
every 26 days after completing 369 orbits. The chronologically successive orbits are offset westwards
because of the Earth’s rotation. The paths are hence organised in the same way as those of a ball of
thread from a skein; the Earth moves eastwards while the orbital plane of the satellite does not change.
The satellite passes over Europe in two adjacent paths every 5 days (Fig. 2.12).
Evidently, remote-sensing data in the visible domain, which requires solar irradiation, is acquired
only during the descendant half-orbit. The satellites SPOT 1 ,2 and 3 comprise 4 spectral bands, viz.,
panchromatic (P): 510 to 730 nm, green (B1); 500 to 590 nm, red (B2): 610 to 680 nm, and near
infrared (B3): 790 to 890 nm.

H Stereoscopy and revisit capability

Because of the repeated coverage (see CD) of the same zone in successive orbits, an important
characteristic of SPOT Is the acquisition of Images, panchromatic or multlspectral, at two different
 Data Sources 43

80

40
20

-4 0

-6 0

0 Ground track, day d

1 Ground track,
2 Ground track,
3 Ground track,
4 Ground track,
5 Ground track,
6 Ground track,
7 Ground track,
8 Ground track,
9 Ground track.

Fig. 2.12: Descending orbits over Europe (after ONES illustration).

angles (pre-processing stage 1A or 1 B).This feature facilitates three-dimensional observation of a

zone using a stereoscope (see CD and Fig. 2.13) and thus construction of digital elevation models
(CD 2.2). Stereoscopy is very useful for numerous thematic studies pertaining to morphology soil
science, geology (ground slides), landscapes and evidently photogrammetric analysis.
The overlapping acquisition ensures a greater revisit capability and this enhances the probability
of acquiring images, without clouds (CD 2.3), over a given region.

B SPOT 4 (see CD)

The orbit of SPOT 4 is the same as that of the other three SPOT systems. It is hence possible to
simultaneously use all these satellites not only for increasing the number of images acquired, but also
for generating stereoscopic pairs the same day from two satellites. In fact, the two high-resolution
visible and Infrared (HRVIR) sensors have the same spectral bands as those of other SPOT satellites,
to which a middle infrared (MIR) band of 1.58 to 1.75 nm, with a spatial resolution of 20 m. Is added.
44 Processing of Remote Sensing Data

Fig. 2.13: Stereoscopy by lateral viewing (ONES illustration).

The panchromatic channel is replaced by a single band (b2) with a wavelength range of 610 to 680 nm
and a resolution of 10 m.

□ Instruments on SPOT 4
PASTEL {Passager SPOT de télécommunication lasei) transmits images to the satellite ARTEMIS
through an optical connection or directly or through a replay of mass memory.
PASTED {Passager technologique) enables study of orbital environment: radiation, ageing,
vibrations, electric potential etc.
DORIS {détermination d’orbite et radiopositionnement intégrés par satellite), already existing on
SPOT 2 and 3, enables determination of the satellite position with an accuracy better than 10 cm, as
well as the ground position of the scanner with the same accuracy.
POAM III (Polar Ozone Aerosol Management), an American payload, measures the ozone and
aerosol contents at the poles.

□ Vegetation instrument (VGT)

The vegetation scanning system on SPOT 4 is a novel tool vis-à-vis the previous SPOTs. With Its very
large field of view, medium spatial resolution and high radiometric resolution. It mainly enables a
global monitoring of phenomena related to vegetation, as well as bare soils and water. Its spectral
bands are b2, b3, MIR and bO (430 to 470 nm).

■ SPOT 5
SPOT 5 has been successfully launched during the night of 3 to 4 May 2002. It provides service
continuity to the previous SPOT 1 to 4 since the orbit is similar, i.e. a circular, quasipolar orbit at an
altitude of 830 km and a pass over the equator at 10.30 a.m. (local time at descending node). The
main improvements compared to SPOT 4 consist in: i) a dedicated instrument for along track stereo
acquisition, li) a higher ground resolution: 5 m and 2.5 m (Instead of 10 m) in panchromatic mode,
10 m (instead of 20 m) in multispectral mode. Otherwise the spectral bands are the same as in SPOT 4
with the short wave infrared band maintained at a resolution of 20 m due to limitations imposed by the
CCD sensors used in this band. The VGT (vegetation instrument) is maintained, as well as the field
width (60 km) and the oblique viewing capacity. These choices answer the growing demand In
cartography, agriculture, planning and environment.
 Data Sources 45

2.3 OTHER SYSTEMS

Besides SPOT, other important Earth resource satellites presently in operation are (Table 2.2)
METEOSAT, NOAA, LANDSATThe satellites ERS, RADARSAT, JERS are used for studies in high
frequencies. Specific thermal infrared sensors are given in Table 2.3 and high frequency sensors are
Indicated in Chapter 26. The CZCS (Coastal Zone Colour Scanner) of NIMBUS 7, which was operational
between 1978 and 1986, will be mentioned in the study of water colour (Chapter 4).

2.3.1 METEOSAT
METEOSAT (Fig. 2.14) is a geostationary satellite and has an altitude of 35,800 km above the equator,
at 0° latitude (Table 2.2).
Lateral swing is achieved by rotation of the satellite per se. An oscillating system ensures
changeover from one row to another. The satellite makes 100 rotations per minute. As 2500 rotations
are needed to cover the Earth, it takes 25 min to produce an image. Obviously, the geometry is altered
by the rotundity of the Earth.
METEOSAT 5 has three bands, viz., 400 to 900 nm with a resolution of 2.5 km and 5.7 to 7.1 iim
and 10.5 to 12.5 pm with a resolution of 5 km (CD 26.2). Ground receiving stations are at Lannion
(France) and KIruna (Sweden). The angle of view is 17°. The zone of observation extends from 60°
North latitude (Sweden) to 60° South latitude (southern part of South Africa) and from 60° West
longitude (centre of Brazil) to 60° East longitude (East of Saudi Arabia).

Fig. 2.14: METEOSAT satellite (after Lliboutry, 1992).

The first METEOSAT satellite was launched In 1977 followed by others, viz., METEOSAT 6 in
1993, and the seventh to be launched 20 years after the beginning of the programme.
The entire Earth, except the polar regions, is continuously observed by METEOSAT and four
other geostationary satellites (Fig. 2.15): two American satellites GOES (Geostationary Operational
Environmental Satellite) located at 75° West and 135° West, the Indian satellite INSAT situated at 75°
East and the Japanese satellite GMS at 140° East.

2.3.2 NOAA
The National Oceanic and Atmospheric Administration (NOAA) operates several kinds of satellites
successively, includingTIROS (Television and InfraRed Observation Satellite) launched in 1960.TIROS-
N launched in 1978, the NOAA-6 and the subsequent satellites (Fig. 2.16) are equipped with an
AVHRR sensor (Tables 2.2 and 2.5).
46 Processing of Remote Sensing Data

Table 2.2: Characteristics of the main Earth observation satellite systems in visible and near infrared bands

System characteristics METEOSAT NOAA-AVHRR LANDSAT MSS LANDS AT TM

Orbit:
Type of orbit: Geostationary Circular Subpolar sun- Subpolar sun-
synchronous synchronous
Altitude (km) 35800 850 705 705
Revisit capability 25 min 12h 16d 16d

Sensors:
Scanning Satellite rotation Rotating mirror Oscillating mirror Oscillating mirror
Spatial resolution 2.5 km (S1); 1.1 or 4 km 56 X 79 m 30 m
5 km (S2, S3)

Spectrai bands (pm):

SI 0.4-1.1 0.58-0.68 0.5-0.6 0.45-0.52
S2 5.7-7.1 0.72-1.1 0.6-0.7 0.52-0.60
S3 10.5-12.5 3.55-3.93 0.7-0.8 0.66-0.69
S4 10.3-11.3 0.8-1.1 0.76-0.90
S5 11.5-12.5 1.55-1.75
S6 10.4-12.5
S7 2.0-2.35
Panchromatic
Scene Globe 2400 185 185
dimensions (km)

System characteristics JERS1-OPS SPOT Early Bird IRS ID

Orbit:
Type of orbit: Sun-synchronous Subpolar Polar sun- Subpolar sun-
sun-synchronous synchronous synchronous
Altitude (km) 570 830 470 905
Revisit capability 44 d 26 (or 1-5) d 2-5 d according to 22 (or 2-4) d
latitude

Sensors:
Scanning CCD arrays CCD arrays CCD matrix LISS:CCD WIFS
CCD
Spatial resolution 18 m 10 m (pan); 3 m (pan); 5 m (pan) 190 m
20 m (S) 15m (S) 20 m (S)
Spectrai bands (pm):
SI 0.52-0.60 0.50-0.59 0.52-0.59 0.62-0.68
S2 0.63-0.69 0.61-0.68 0.49-0.60 0.62-0.68 0.77-0.75
S3 0.76-0.86 0.79-0.89 0.615-0.670 0.77-0,86
S4 1.59-1.70 0.79-0.875 1.55-1.70
S5 2.0-2.1
S6 2.1-2.2
S7 2.2-2.3
Panchromatic 0.5t-0.73 0.445 -0.65 0.50-0.75

Scene
dimensions (km) 75 60 131 774
 Data Sources 47

Fig. 2.15: Disposition of meteorological geostationary satellites.

Fig. 2.16: The NOAA-H satellite (after Lliboutry, 1992).

NOAA-7 passes over the same region every 12 h, i.e., at 1400 h and 0200 h. It passes above the
same zone every 19 days and completes 268 orbits to cover the globe. Its field of view is 2700 km by
2700 km.
The system consists of AVHRR (Advanced Very High Resolution Radiometer) with a rotating
mirror and has the spectral bands of 580-680 nm, 720-1100 nm, 3.55-3.93 pm, 10.3-11.3 pm and
11.5-12.5 pm for NOAA-8 and 9. The resolution is 4 km for NOAA-10.
Examples of Image interpretation are given in Chapter 26 (CDs 26.1 and 26.3; also see CD 4.3).

2.3.3 Thermal sensor systems

In view of the importance of applications of thermal Infrared remote sensing, this band is installed in a
number of satellites. Table 2.3 summarises the main systems that presently have (or are planned in
the near future) channels in this spectral domain. The criteria of spatial resolution and temporal resolution
are Important for agricultural applications and, in particular, for monitoring water conditions of crops. It
can be seen from the table that presently no system exists which is suitable for these objectives,
simultaneously combining high spatial and temporal resolutions. Meteorological satellites (such as
METEOSAT or NOAA) have the advantage of high repetition, which enables temporal surveillance of
areas but have low spatial resolution. Consequently, various airborne sensors have been employed to
48 Processing of Remote Sensing Data

Table 2.3: Principal thermal infrared sensors (after Prevot et al., 1996, in INRA publ.).

Satellite Sensor Spectral band (pm) Revisit capability Resolution Field of view Date

METEOSAT MVIRI 10.5- 12.5 30 min 5 km Geostationary

NOAA AVHRR2 3.55-3.93 Twice per day 1.1 km 3000 km
c4:
10.3- 11.3
c5:
11.5- 12.5
ERSI ATSR 1.6 3 to 35 days 1km 50 km 1991
ERS 2 ATSR2 3.7 1995
10.8
12
LANDSAT7 ETM+ 10.4- 12.5 16 days 120 m 185 km1998

EOS-AM ASTER 5 16 days 90 m 60 km 1998

8-12
EOS-AM1 MODIS 0.4-14.4 2 days 1 km 2300 km1998
(14c)
MSG 1 SERIVI 3.7 15 min 3 km Geostationary 2000
8.8
10.8
12
Airborne sensors
German DAIS 7915 6 channels 512 pixels/line 78° view 1993
Scanner 8 to 14 pm
Italian MIVIS 10 channels 755 pixels/line 71° 1994
Scanner
French Inframetrics 1/3 256 pixels/line 7°-20°-80° 1992
Camera 760 possibilities

acquire data of higher spatial resolution but, in this case, repetition of measurements is low. Moreover,
airborne systems often have multiple view angles (Table 2.3); this feature is useful to test the spectral
and spatial resolutions pertinent for a better monitoring of the state of soil and vegetation with the
objective of proposing future satellite sensors.

2.SA LANDSAT
The Earth Resources Technological Satellites (ERTS) programme, using ERTS-1 satellites and renamed
as LANDSAT (Land Satellite), is launched by NASA. The EROS (Earth Resources Observational
Satellites) centres were established in 1966 to distribute images of the missions APOLLO, GEMINI and
SKYLAB. From 1986 until recently the products have been commercialised by the EOSAT society, but
now Landsat 7 products (launched in April 1999) are commercialised by the US Geological Survey
(USGS).
There are two series of LANDSAT satellites: numbers 1 to 3 and numbers 4 to 7.
LANDSATs 1 to 3 (Fig. 2.17) contained a Return Beam Vidicon (RBV) camera (479-575 nm;
580-680 nm; 690-830 nm) and a Multispectral Scanner (MSS) with four spectral bands (Table 2.2).
One pixel of MSS Images corresponds to 79 m width over 56 m along the track. The four spectral
 Data Sources 49

LANDSAT-1

3.3 m

Fig. 2.17: LANDSAT 1 satellite (after LIiboutry, 1992).

bands are MSS4 = 500-600 nm, MSS5 = 600-700 nm, MSS6 = 700-800 nm and MSS7 = 800-1100
nm (CD 2.1). The total field of view is 11.56° and spatial resolution 56 m by 79 m.
MSS was the first system to procure multispectral images for scientific applications. Studies of
these images showed that channel 6 was not very useful since bare soils and vegetation could not be
distinguished. Although this band was not retained for subsequent LANDSAT satellites, this sensor
has always existed for purposes of comparisons with the images acquired since 1972.
LANDSAT 5 is a sun-synchronous satellite in a subpolar orbit. Its altitude varies from 696 km at
the equator to 741 km at the poles, with a standard altitude of 705.3 km (438 miles). Inclination is
98.22. The satellite passes over the equator at 9 h 37 min local time. With a period of 98.9 min per
revolution, it makes 14 and 9/16 orbits per day. The periodicity of passage Is 16 days. It completes 233
paths to cover the globe and acquires 248 scenes per path. Path 1 intersects the equator at 64.6° W.
Overlapping between two scenes is 7.6 % at the equator and this value Increases with latitude, reaching
54% at 60° latitude.
The satellite has no onboard recording system since it transmits the data In real time to receiving
stations. When there is no direct contact, the data is transmitted through relays using communication
satellites TDRS (Tracking and Data Relay Systems).
Every LANDSAT 5 scene measures 170 km in the north-south direction and 185.2 km (100
nautical miles) in the east-west. The territory of France is covered by 210-219 paths and 25-30 rows.
LANDSAT 4 and 5 carry an MSS scanner and a Thematic Mapper, which comprises 7 spectral
bands (Table 2.2)— ^three in the visible with one closer to blue, one in the near infrared, two in the
middle infrared and one in thermal Infrared. The resolution is 30 m except for the thermal Infrared band
where it is 120 m (see the Thematic Mapper Images of Brienne on the CD).
Sixteen photodetectors in parallel are employed to scan 16 rows simultaneously for all the channels,
excluding channel 6 for which, evidently, only four detectors are available. Differences in sensitivity
between the 16 photodetectors can produce parallel strips on the images (termed ‘stripping’) (CD
2.4).
LANDSAT 1 was launched on 22 July 1972 and functioned up to 6 January 1978. LANDSAT 2
was sent into orbit on 5 November 1975 and worked up to 27 July 1983, when some abnormalities
affected its sensors subsequently. LANDSAT 3 was launched on 5 March 1978 and retired on 7
September 1983 following some trouble in the scanner system. LANDSAT 4, launched on 16 July
1982, also had technical problems such as breakdown in the TM source unit from February 1983 and
trouble in two solar panels out of four. Launched on 1 March 1984, LANDSAT 5 functioned with no
problem. LANDSAT 6 crashed into the sea during launching. LANDSAT 7 was sent into orbit in April
1999.
50 Processing of Remote Sensing Data

2.3.5 ERS-1 and 2

The ERS series of satellites (Fig. 2.18 and Tables 2.3 and 2.4) carry active sensors, viz., a radar
altimeter, a backscatter reflector for laser telemetry from the ground and an active microwave instrument
(AMI), designed and developed by the research centre for physics of environment. The latter consists
of a scatterometer for measuring the height of waves, another scatterometer to record the force and
direction of wind on the sea and a synthetic aperture radar (SAR) unit. In addition, a passive sensor
with a scanning system, viz., ATSR, is also used. A microwave radiometer measures the total water-
vapour content of the atmosphere, while another infrared radiometer measures two different thicknesses
of the atmosphere, vertically and at 50°. Sea surface temperature is obtained with an accuracy of 0.3-
0.5° C, with a resolution of 1 km, over a 500-km swath.
ERS-1 was placed in orbit on 25 July 1991 .The IFREMER centre in Brest processes and archives
the data of the altimeter and scatterometers.

11.8 m

Fig. 2.18: ERS-1 satellite (after Lliboutry, 1992).

2.3.6 RADARSAT
RADARSAT is the only system which provides viewing at different angles (Table 2.4) and a choice
between a high geometric resolution (10 m in fine mode) and a coarse resolution (100 m). These
characteristics are programmable; this facility offers many advantages, provided the type of data required
is clearly identified.These various possibilities ensure a better understanding of the phenomena under
study but sometimes make acquisition difficult.

2.3.7 JERS
JERS is the only radar system operating in the L band (Table 2.4). The L band can be very useful,
especially when data are available from JERS and ERS or RADARSAT for a given region, since the
phenomena detected by these different sensors are not the same (see Chapter 26). The backscatter
signal, for example, is much more sensitive to the row effects of plants, which may be an advantage or
an inconvenience depending on the application under Investigation.

2.3.8 Evolution of sensors

Instruments of various resolutions are currently under development.
 Data Sources 51

Table 2.4: Examples of satellites with microwave sensors

EPS JERS RADARSAT

Date of launching 1:07/25/1991 02/11/1992 11/4/1995

2:04/20/1995
Altitude 785 km 570 km 800 km
SAR (Synthetic Image mode
Aperture Radar) sensors
Spectral band C L C
Frequency 5.3 GHz 1275 MHz 5.3 GHz
Polarisation v-v H-H H-H
Special Possibility of a mode to 44 days revisit capability Various view angles, 24 d
characteristics study waves and wind cycle, 1.5 d revisit capability
velocities at 45° North latitude possible
Angle of incidence 23° 35"^ 10° to 60°
in middle of swath
Spatial resolution 25-30 m 18-18m 10 to 100 m depending on
view angle
Swath 100 km 75 km 50 to 500 km
Other sensors ERS-1: OPS (Optical sensor):
onboard Radar altimeter, ATSR 3 channels in visible
scanner (4 channels and NIR, 4 channels
centred on 1.6, 3.7,10.8 in infrared
and 12 pm)
ERS-2: The same plus
GOME (Global Ozone
Monitoring Equipment)

Instruments o f medium resolution, 1,000 to 100 m, provide increased opportunities for image
acquisition due to their higher frequency of passes (4 times a day). Applications of these instruments
include monitoring dynamic phenomena of the environment over vast areas. Instruments envisaged
for launching at the beginning of this century, in addition to the SPOT 4 vegetation instrument already
In operation, include:
— vegetation instrument of SPOT 5, France (5 bands from visible to near infrared and reflective
medium band, resolution 1000 m, swath 2000 km);
— MSU-SK of RESURS-01 mission, Russia (5 bands from visible to thermal infrared, without
MIR, resolutions of 170 or 600 m for a swath of 600 km);
— MODIS of EOS missions, USA, Proto-Flight Model (PFM) launched on 18 December 1999 (36
bands: 20 between 0.4 and 6.0 pm, 16 between 3 and 15 pm, with a resolution of 250, 500 or 1000 m
and a swath of 600 km);
— MERIS of ENVISAT mission, Europe (15 programmable bands between 0.4 and 1.05 pm, with
a resolution of 300 to 1200 m, and a swath of 1450 m).
Instruments o f high spatial resolution, 15 to 8 m, enable acquisition of more accurate information
about ground features, in particular detection and mapping of land-use, differentiation of plant species,
moisture content of soils and plants etc. The HVIR of SPOT 4, already operational, belongs to this
category as well as the following instruments:
— PAN/LISS-3 of IRS-1 D mission, India (5 bands from visible to MIR, resolution 5,8 m in
panchromatic, 23.5 m in multispectral, 70.5 m in MIR, and a swath of 148 km, probable revisit capability
5 days in panchromatic);
52 Processing of Remote Sensing Data

— AVNIR of ADEOS, Japan (5 bands from visible to NIR, resolution 8 m in panchromatic and 16
m in multispectral, and a swath of 80 km, probable revisit capability 3 days).
instruments o f very high spatiai resoiution, 8 to 1 m, may prove to be very important for
applications in urban studies as well as in environmental investigations for detection and monitoring of
narrow linear structures (such as roads, afforestation zones, silviculture strips etc.). Important among
them are:
— HRG of SPOT 5 missions, France (5 bands from visible to MIR, 5 and 2.5 m resolution in
panchromatic, 10 m in multispectral and 20 m in MIR, swath 60 km, estimated revisit capability 4
days);
“ -IRS-1C mission, India (5.8 m resolution in panchromatic, 16 m in the blue band);
— ADEOS 1 mission, Japan (8 m resolution in panchromatic);
— Earth Watch programme using the satellite Early Bird, launched on 24 December 1997;
unfortunately, it has been out of order since 28 December. It was expected to acquire panchromatic
images ( 3 m resolution) and multispectral images (15 m resolution) of 3 x 3 km, with a repetition of 2
to 5 days over the same site. The programme is to be continued with the launching of Quick Bird
(resolution 1 m).
An example of future high-resolution satellites (American and French) as of September 1994 is
given in Table 2.5 (after Mussio and Light, 1995).

Table 2.5: Characteristics of some satellites

System Lockheed/E- Sci Orbital, Itek, IKONOS SPOT 5

characteristics Space Imaging GDE, Eyeglass
Systems Inc.

Orbital
characteristics
Orbit type Polar sun- Polar sun- Polar sun- Polar sun-
synchronous synchronous synchronous synchronous
Altitude km 680 700 681 832
Cycle (days) 14 16 26
Revisit capability 1-3 2 2,9 5
Sensor characteristics
Sensor type CCD arrays CCD arrays CCD matrix CCD arrays
Spatial resolution:
Panchromatic 1 m at nadir 1 m at nadir 1 m at nadir 5 m and 2.5 at nadir
Multispectral 4 m at nadir 4 m at nadir 10 m at nadir

Spectral bands (\xm)

Panchromatic 0.50-0.90 0.50-0.90 0.51-0.73
S1 0.45-0.52 0.45-0.52
S2 0.52-0.60 0.52-0.60 0.50-0.59
S3 0.63-0.69 0.63-0.69 0.61-0.69
S4 0.76-0.90 0.76-0.90 0.79-0.89
S5 1.55-1.75 1.58-1.75
Stereoscopic coverage Front-rear Front-rear on the track on the track
GPS Yes Yes Yes No
Field of view (degrees) 1.3 1.2 0.34
Other characteristics
Scene dimensions (km) 60x60 15x 15 11 x4 0 60 X 60
Ground stations 3 local + abroad TBD 4 21
Date of launch 1998 1998 1999 2002
 Data Sources 53

User application of such data is beset with two difficulties, viz., large computer capacities required
for processing such large volumes of data and cost of data acquisition and processing.
While it is never certain that every proposed system will become operational, these forecasts
indicate the major tendencies towards which remote sensing is progressing.
One of the approaches chosen by the European Space Agency (ESA) concerns the concept of
small satellites that facilitate testing and validation of new instruments. Their smaller weights ought to
enable their launching by lighter carriers or simultaneous launching of multiple satellites. The objective
Is to separate the Industrial sector from the research sector so as to meet the needs of the users
better. Users in particular demand a guarantee of continuity in service for a sufficient duration, which
Is possible only through involvement of public authorities supporting private initiatives.

2.3«9 Satellite photography

Acquisition of photographs of the Earth from satellites requires return of the films to the ground. For
this,
— the flight has to be manned and the space vehicle returned to the ground, as in the case of
GEMINI and APOLLO missions of the USA or Soyuz missions of Russia, or
— ^the films have to be in modules or satellites that can be recovered on the ground, as In the case
of Big Birds, the American satellites flying at an altitude of about 200 km, and the Russian Cosmos. In
some cases, photographs taken with remote cameras can be transmitted to the ground in digital form;
resolutions achieved in such cases would be much better and in the range of decimetres and not in
metres (Lliboutry, 1992).
The Earth has been photographed during manned satellite missions by SKYLAB and by space
shuttles. Photographs were also taken with metric cameras onboard SPACELAB in 1983 and 1985.
They were stereoscopic, colour infrared, with a resolution of about 10 m.

2.4 CONCLUSION
The possibilities of obtaining satellite or aerial photographs are becoming Increasingly numerous and
the sensors more diversified. It is now possible to choose the date of acquisition of images and multiply
them by repetition. Several different fields of view can be used. Geometric and spectral resolutions
have become finer. One can acquire analog data and digitise them, or directly obtain digital information.
Collection of information is becoming Increasingly easier through multiplication of receiving stations.
More spectral bands are being used; sensors with tens of bands are envisaged. These developments
lead to two Important consequences:
— the choice ought to be made according to the thematic application of interest (see part E);
— the quantity of data becomes larger and larger.
A TM image with a 185-km field of view roughly corresponds to 9 SPOT images and 625
conventional aerial photographs of IGN. A multlspectral pixel of TM (900 m^) is equivalent to slightly
more than 2 multlspectral pixels of SPOT (2 x 400 m^) and 9 panchromatic pixels. The volume of aTM
image hence corresponds to 6 channels of 36 million pixels, plus 1 channel of 2.25 million pixels (for
thermal infrared), each pixel being of 8-blt code, which amounts to 1746 megabits, or 218 megabytes.
For an HRV image, the volume of information is 63 megabytes for four spectral bands.
The number of spectral bands can be increased considerably but it is to be noted that data
processing using colour filters can be carried out only for three channels at a time (see Chapter 3).
Hence, one has to choose the most Important among the possible channels or devise new channels
by combining a larger number of spectral bands. The resolution can be improved but, for a given field
of view, the number of pixels increases by a power of two. It may be noted that in the near future
spectral bands will be coded on 10 and more bits. Processing of such volumes of data demands
54 Processing of Remote Sensing Data

computers of very large RAM memory. Effectively, this is the present-day dynamics of computer
designers.
After processing, the data need to be interpreted based on physical and geometric models. This
requires some time for developing mathematical and statistical models and new approaches to tackle
the problem. Researches in the coming years ought to be focused on this aspect, if at all any predictions
can be made on remote sensing, this young science whose usage has just commenced, although it is
already thirty years old.

References
Bonn F, Rochon G. 1992. Précis de télédétection. 1: Principes et méthodes. Presses de l’Université du Québec/
AUPELF, 485 pp.
Gilliot J-M. 1994. Traitement et interprétation d’images satellitaires Spot: Application à l’analyse des voies de
communication. Thèse de Doctorat, Université Paris V, 197 pp.
LillesandT, M, Kiefer R, W. 1994. Remote Sensing and Image Interpretation. John Wiley & Sons Inc., NY, 3rd ed.,
750 pp.
Lliboutry L. 1992. Sciences géométriques et télédétection. Masson, Paris, 289 pp.
Mussio L, Light D, L. 1995. Sensors, platforms, and imagery symposium. P hotogram m etric E ngineering & Rem ote
Sensing, 61:1339-1344.
Prevot L, Laville P, Xing-Fa GU. 1996. Les mesures de télédétection dans l’infrarouge thermique. Actes de l’école
chercheurs INRA en bioclimatologie. Le Croisic, 25-29 mars 1996, Tome 2 du couvert végétal, pp. 69-80.
 B
PHYSICAL
INTERPRETATION
OF DATA
 3
Composition of Colours
Only the basic information required for understanding and use of colours in remote sensing is presented
in this chapter. In fact, colorimetry per se is a very complex science. Books published by the International
Commission for Colorimetry (1931) and that by Wyszecki and Stiles (1982) can be consulted for
details. Among recent publications related to remote sensing, the papers by Escadafal et al. (1988,
1989,1993) which are particularly concerned with applications in soil science are useful.
In remote sensing, a thorough knowledge about colours is necessary for:
—“Visual interpretation of images printed on paper,
— interpretation of colours on screen,
— preparing colour composites of processed data to make results readable,
— choosing the best possible printing of results on paper.
In order to establish comparisons between the objects that our eyes see and the colour of these
same objects on images or photographs, it is also necessary to understand the systems of colour
composition by our eyes, by computer monitors and by pigments.
Colours in light are called ‘additive colours’ and the colour impressions obtained by means of
pigments are known as ‘subtractive’ colours. It is hence evident that all the colours that can be seen on
a screen cannot be reproduced on paper. The latter depends on the pigments used for preparing
various types of ink.

3.1 THE HUMAN EYE AND COLOUR

3.1.1 Vision
Light penetrating the eye induces chemical reactions which are converted into precoded electric signals.
These signals are transmitted through the optical nerve in the brain, which perceives a colour. The
process is complex since any excitation gives a colour perception even when there is no colour sensation
coming from the eye, as in the case of dreams, for example.
The light received by the eye reaches nearly 120 million rod-like cells on the retina, which enable
scotopic vision (vision in dim light), and another 6 to 7 million cone-shaped cells, which provide colour
vision. The remaining light is absorbed by the melanin of chloride, preventing any diffusion. There are
three types of pigments for cone cells. These are: erythrolabe, which has a maximum absorption at
560 nm and perceives red, yellow and white; chlorolabe, which has a maximum at 530 nm and perceives
yellow, green and white; and cyanolabe which has a maximum absorption at 450 nm and perceives
blue, red and white. The eye perceives a few thousand or tens of thousands of colours.
58 Processing of Remote Sensing Data

3.1.2 Sensitivity
The spectral sensitivity of the eye (Fig. 3.1) varies from approximately 400 nm to 700 nm, with a
maximum between 540 and 560 nm, which corresponds to green-yellow. In this band, a variation dx
in wavelength leads to a very small change in spectral sensitivity, which cannot be differentiated by the
eye. Consequently, the eye does not perceive small nuances in yellow shades. On the other hand, the
same variation dX in the red or blue-green band produces a significant change In spectral sensitivity
and hence the eye perceives many variations In blue-green and red wavelength bands.

Fig. 3.1 : Sensitivity of the eye (after Lliboutry, 1992).

3.1.3 Contrasts
The human eye exhibits contrast effects in the course of time. Thus, after observing an intense colour
for a long time (for example, a red setting sun), by turning the face, we ‘see’ a complementary colour
(green sun) during the time necessary for reconstitution of retinal pigments altered by light (red).
The eye also shows spatial contrast effects. The same grey colour appears less clear if surrounded
by white and more distinct if surrounded by black. The effect is similar for complementary colours:
orange looks more like yellow if surrounded by blue. To eliminate such effects, every colour needs to
be surrounded by grey colour.
There Is an edge effect as well. Near the boundary between dark grey and light grey zones, the
dark grey would appear darker and the light grey lighter. This Is the reason why it is believed that a
large number of grey levels can be distinguished. But in fact, the eye perceives little more than a dozen
grey levels, ranging from black to white, when these levels are not juxtaposed.

3.2 RED-GREEN-BLUE SYSTEM

Computer and television screens emit three colours, viz., red, green and blue, which are intermixed
without this being shown on the screen. These are the primary colours. This trichromatic system is
based on the principle that any colour can be produced by an appropriate mixture of these three
primary colours. The International Commission on Illumination (ICI, 1931) adopted the following
wavelengths for the primary colours: red, x (^) = 700 nm, green, x (G) = 546 nm and blue, x (B) = 436
nm. These were chosen from empirical results which showed that this combination produces the
largest range of colours.
The colour of an object can hence be computed as a vector C of co-ordinates r, g and b (Fig. 3.2):

C = rR + gG + bB

The units are chosen in such way that white colour Is obtained when b = g = r =1.
The curves of colour sensations of an average observer, or the effects of mixing, are generated
based on a series of colorimetric measurements carried out with a field of view between 1° and 4®.
 Physical Interpretation of Data 59

Fig. 3.2: Computation of co-ordinates r, g and b o f a colour vector.

They define the characteristics of the reference observer (ICI, 1931) and represent the percentages of
red, green and blue to be mixed for obtaining the sensation corresponding to each of the monochromatic
radiations of the visible spectrum. Negative values of r can be clearly observed, with a minimum
around 510 nm (Fig. 3.3). The existence of colour sensation curves also constitutes the basis for
colorimetry.
These colours are hence called additive primary. Addition of primary colours defines the secondary
colours (Fig. 3.4 A), as mentioned below;

Fig. 3.3: Combination of monochromatic colours for the standard observer (ICI, 1931).

Fig. 3.4: Additive (A) and subtractive (B) colour compositions of R, G and B.
60 Processing of Remote Sensing Data

— blue + green = cyan

— green + red = yellow
— red + blue = magenta.
When the three colours are added, white is obtained, i.e., blue + green + red = white. Black
represents absence of colour. Therefore, we can define an achromatic axis passing through black and
white.

3.3 CUBIC REPRESENTATION

3.3.1 Additive system
The colour space can be represented by a cube whose three orthogonal axes constitute the proportions
of red, green and blue (Fig. 3.5). At the origin, values of the three axes are equal to zero and the
resultant is black. An identical maximum value for the three axes gives white. The corners of the cube
represent white, magenta, red, yellow, green, black, blue and cyan. The diagonal connecting the two
corners, black and white, constitutes the achromatic axis, which comprises all the possible grey levels.
Other colours are produced by changing the proportions among the three principal colours.

Fig. 3.5: Trichromatic cube.

A computer or television monitor operates on the principle of additive colours. Colours can be
reconstructed in this manner (CD 3) if 256 values for each of the three primary colours are available.
This number is much higher than what our eye can distinguish.

3.3.2 Subtractive system

Using the same cubic system, a similar colour space is obtained when the axes are defined by cyan,
yellow and magenta (Figs. 3.4 B and 3.5) instead of red, green and blue. The origin of these three axes
represents white and the maximum value for the three axes black. These three colours are hence
called subtractive: the transmitted light is effectively lower in intensity than the light received. Any
colour can be produced using white light, from which yellow, magenta and cyan colours are subtracted
by means of filters as follows:
 Physical Interpretation of Data 61

— the magenta filter transmits red and blue but stops green;
— ^the yellow filter transmits red and green but absorbs blue;
— the cyan filter transmits green and blue but stops red.
Consequently, when two filters are used, such as yellow and magenta for example, only one
colour, i.e., red, can pass through. As a corollary the magenta and cyan filters allow blue, and the cyan
and yellow filters transmit green. Depending on the filters, the light intensity after passing through a
filter is greater or smaller. Thus, any colour can be composed. When all three filters are used, all three
colours are subtracted and no light exists, resulting in black.
The colour pigments are determined from subtraction of colours. The pigments absorb certain
colours and, when they are mixed, colours are accumulated and shades absorbed. This Is the principle
involved In printing Inks produced from yellow, magenta and cyan colours. In practice, in the offset
process, the density of inks does not vary continuously. The different colours are obtained by the
method of half-tints: colour points of different inks are distributed on grids of points, which are superposed
at different angles; this results in overlaps. The points hence become large or small and the intensity
of absorbed light Is less or more, resulting In restoration of diverse colours.
In additive as well as subtractive systems, complementary colours can be defined as those whose
sum gives grey. The latter corresponds to the two opposite points of the achromatic axis. The
complementary colours are therefore red and cyan, yellow and blue, and green and magenta.

3.4 TRIANGLE OF THE INTERNATIONAL COMMISSION ON

ILLUMINATION
With the objective of describing colours in a Cartesian system, the ICI makes use of the trichromatic
co-ordinates X, y and Z derived from R, G and B, which can be computed as follows:

X = 2.7659 R+ 1.7519 G + 1.13020 B

y = f ? + 4.5909 G + 0.06012 e

2 = 0 .0 5 6 5 G + 5.59440 e

The colour of an object can thus be computed from Its spectral reflectance curve (CX) and the
composition of fight that illuminates it (hfk), using the respective mixing functions, viz., rX, gX and b x .

780 nm
R=k J C l Hk - r X- dX
380 nm

780 nm
G=k J Ck- Hl - gX- dX
380nm

780 nm
B=k I Cl - H X - b X - d X
380 nm

Using the above, the trichromatic co-ordinates x, y, zcan be derived from the relation:
62 Processing of Remote Sensing Data

X X y z 1
X = -------------- or — = ^ = — = ------------
X+Y+Z X Y Z x+y+z

The co-ordinates are conventionally represented in a right-angled triangle (Fig. 3.6). The point
named ‘White’ is the achromatic point and the dotted line gives the triangle RGB. Thus, every colour
can be characterised by brightness (Y) and its chromatic co-ordinates xand y.

Fig. 3.6: The ICI chromatic triangle based on the co-ordinates X, Y and Z.

3.5 MUNSELL SYSTEM

While the Red-Green-Blue system is represented in cubic form, the Munsell system is based on
cylindrical co-ordinates (Fig. 3.7).

WHITE
 Physical Interpretation of Data 63

The axis of the cylinder corresponds to the diagonal of the RGB cube: it is an achromatic axis
which ranges from black to white, passing through all the grey points, and corresponds to a Munsell
value. The square of the Munsell value gives brightness. For a given value of the achromatic axis, a
perpendicular plane can be defined on which the angle of rotation corresponds to the hue. On this
plane, from the achromatic axis a circle is defined whose radius characterises the value of chroma.
Every colour Is thus defined by three co-ordinates: value, hue and chroma (Intensity (lightness), colour
and saturation are also used).
A catalogue of Munsell colour codes exists. A digest chart that gives soil colour codes is used in
soil science worldwide. However, all colours are not represented since the ranges of value and chroma
do not exceed 8, whereas the theoretical ranges are 10 and 24 respectively (Fig. 3.7).
Colour pellets made from a mixture of pigments that produce a colour sensation are employed.
The three pigments used correspond to the three relative maxima on the reflectance curve, viz., 430
nm, 530 nm and 590 nm (Fig. 3.8).

R {% )

Fig. 3.8: Reflectance curves of a soil sample and the Munsell pellet giving the same colour print.

Conversion of the Munsell code to the RGB code is not simple and necessitates conversion
tables or computer programs, since the geometry of the two colour spaces differs (Wyszecki and
Stiles, 1982). As an example, one can see in Fig. 3.9 the projection of values of colour of hue 7.5 YR
and 5 YR on the Red-Green plane, on the one hand, and on the Red-Blue plane, on the other. It is very
clear that the colours are better separated on the Red-Blue plane than on the other plane. The influence
of hue is small; chroma mainly affects the Red component; value affects the two components, and the
distance between two values is greater In high values than in low values.

3.6 METAMERISM
The term metamerism denotes the phenomenon whereby objects of different spectral properties may
produce the same colour sensation. Thus, an object with high reflectance In green and red bands may
appear the same yellow as an object that reflects in a monochromatic band of wavelength corresponding
to yellow. Hence, the same colour sensation may be obtained with very different reflectance curves
(Fig. 3.10). As no bijective relationship exists between reflectance and colour, it is generally not possible
to predict the reflectance curve of an object from its colour.
Two colours are generally metameric only under a given illumination. Thus, the colour of a soil
and that of the Munsell pellet 10 YR 6/5 (Fig. 3.8), which give the same colour sensation, exhibit
distinctly different reflectance curves that intersect one another thrice. It has been established that the
64 Processing of Remote Sensing Data

Red

0.10 -

/im-
■ A«'« •■
ÌÀ6M

.^ 5 (6 .
«.«/a /

0.05

Green

0.05 0.10
(a)

Red

■•■8/2/'
▼
0.10 -
* /"
■f\6/8 .

;
.*
'¿e/®
••
'^vy
/
.’ .■ V e/ /
■;a5/8.' *
^ & 5/6 *
>6/2/
0.05 - • ^ 5 /4 .•

a2 /4 > ‘
Blue

0.05 0.10
(b)

Fig. 3.9: Projection of Munsell colours in ICI space: a) Red-Green, b) Red-Blue. Black triangles correspond to
the hue 5 YR and white triangles to the hue 7.5 YR (after Escadafal, 1989).

reflectance curves of metamerie objects ought to intersect one another at least three to five times
(Takahama and Nayantani, 1975; Ohtha and Wyszecki, 1997). In the case of soils, the reflectance
curves do not intersect and, consequently, a relationship can be established between the colour of soil
surfaces and their reflectance curves. Therefore, the colour of soils can be determined from satellite
images.
 Physical Interpretation of Data 65

Reflectance (%)

- - 4 .

Fig. 3.10; Six metamerie curves that give the same colour imprint.

3.7 COLOUR IN PHOTOGRAPHY

3.7.1 Principle of emulsions
Use of emulsions can be readily understood by considering a black and white emulsion. An emulsion
consists of a thin layer of photosensitive silver halide grains coated on a base. When the emulsion is
exposed to light, a chemical reaction takes place. During the process of development the silver salt is
reduced to silver, which is black in a mono-atomic state. After some time, development is stopped by
immersing the film in an acid solution. The emulsion is then placed in a fixer bath, which removes the
unexposed salt grains and makes it chemically stable. The negative is then washed and dried. As a
result, a grey level proportional to the energy received (Ej) is obtained. The latter is determined by the
Integral of the reflectance curve between the two limits of the spectral band (Fig. 3.11 ).

3.7.2 Panchromatic and infrared

Transmission 7 is defined as the ratio of light quantity transmitted to the emulsion, Ep and the total
incident light, Ef

^ Arrow indicates the value of the signal in the corresponding band

Resolution of the signal of vegetation into different bands of three emulsions

Fig. 3.11: Energy coming from vegetation and received by different emulsions.
66 Processing of Remote Sensing Data

Opacity is the inverse ratio: Op = 1 /7= E j/

Density, D, is equal to the logarithm of opacity:

D = log (Op) = log 1 /7 = - lo g 7

Thus, we get:D = 0.3 for 7 = 100% and D = 1 for 7 = 50%.

The sensitivity of a film varies according to wavelength and density (Fig. 3.12).

Wavelength (nm)
400 450 500 550 600 650 700 750
Sensitivity log D - density

400 450 500 550 600 650 700 750 800 850 900 950 Wavelength (nm)

Fig. 3.12: Spectral sensitivities of panchronnatic (A) and black and white infrared (B) (after Kodak).

Aerial photography conventionally uses panchromatic emulsions that are sensitive to visible light,
400 to 700 nm (Fig. 3.12 A). The emulsion is limited to 400 nm since this Is the lower boundary of the
visible spectrum but it is possible to make ultraviolet photographs as well. However, the photographic
lens absorb wavelengths below 400 nm.
Infrared emulsion, 700 to 900 nm (Fig. 3.12 B), is also used in aerial photography by filtering the
light below 700 nm. Beyond 900 nm the emulsion Is photochemically unstable. The glass of the
photographic lens absorbs wavelengths exceeding 900 to 1000 nm. Quartz lens can be used to extend
the sensitive range to 1100 nm.

3.7.3 Colour
Colour photography is based on the subtractive system of colours, yellow, magenta and cyan. The
yellow layer absorbs the blue component of white light, the magenta layer absorbs the green component
and the cyan layer absorbs the red component.
Emulsions are made up of three layers, coated on a base, each layer being sensitive to a spectral
band (Fig. 3.13A).The first layer, yellow, is sensitive to blue, while the second, magenta. Is sensitive to
blue and green. A filter is also used to absorb blue. This filter also serves as the third layer, cyan, which
is sensitive to red and blue. In each layer, the quantity of recorded energy is inversely proportional to
the light intensity of the wavelength band for which the layer is sensitive.
 Physical Interpretation of Data 67

Sensitivity log

Fig. 3.13: Spectral sensitivities of colour (A) and colour-infrared (B) emulsions (after Kodak).

After development and printing, the initial colours of blue, green and red are obtained (Table 3.1).
However, when the distance between the photographic camera and the Earth increases, it is necessary
to use a yellow filter in order to avoid atmospheric disturbances, which occur mainly in the blue band.

3.7.4 Colour infrared

Colour-infrared emulsion (earlier called false colour) was developed forty years ago for detecting
camouflages. It is useful in distinguishing live vegetation from cut tree branches. It covers the spectral
domain of 500 nm to 900 nm, i.e., the green, red and near-infrared bands. Blue is filtered since this

Table 3.1: Principle of operation of colour and colour-infrared films

Spectral bands Blue Green Red Infrared

Colour film
Normal sensitivity Blue Green Red
Colour of layers Yellow Magenta Cyan
Resultant colour Blue Green Red
CIRfilm
Normal sensitivity Blue Green Red Infrared
After filtering Green Red Infrared
Colour of layers Yellow Magenta Cyan
Resultant colours Blue Green Red
68 Processing of Remote Sensing Data

band, very much affected by atmospheric disturbances, is of no importance in aerial or satellite

photography. The remaining part of the spectrum is covered by three bands (Fig. 3.13 B). As these
three bands are sensitive to blue, a filter is used to suppress all the corresponding wavelengths.
The principle of this emulsion lies in shifting all wavelengths: when reflected, the green light
appears blue, the red appears green and the near-infrared appears red (Table 3.1). Since the eye
perceives a large number of shades In red, the possibility of detection of various features is increased
as most elements of the natural environment (except water) have higher reflectance in the infrared
than In the visible. This is particularly distinct for vegetation and hence this emulsion is important for
agronomy, forestry and vegetation study In general.
Application of colour infrared developed with satellite remote sensing since images of Earth-
observation satellites are almost always presented in this colour composition. Thus, the quick-look
entries in the reference catalogue of SPOT images: SIRIUS (see the chapter SPOT on CD) are in
colour infrared (Commonly known as IRC).
Colour as well as infrared aerial photographs are now available for France‘S.

3.8 TREATMENT OF COLOURS ON COLOUR SCREEN

3.8.1 Colours on screen
It has been seen that colours can be composed from the three primary colours. For each pixel of the
colour screen (monitor), additive mixing of the three (luminophore) primary colours, viz., blue, red and
green, is carried out. The chromatic gamut that can be managed by a monitor is limited by the
luminophore and hence all colours of the chromatic triangle (Fig. 3.6) are not reproducible on a monitor.
The quality of visualisation depends on the number of rows and columns of the monitor, the number of
colours available and the number of colours that can be displayed simultaneously (CD 3.1).
On the colour screens used for image processing of remote sensing data, 256 levels can be used
for each of the three primary colours. Flence, 256^, i.e., 16,777,216 different colours can be composed
on the screen but not all of them can be seen since our vision Is limited. Thus, the colours produced
from the first 40 to 50 levels of digital values coded on 8 bits (256 levels) practically cannot be
distinguished from black. In most cases, for low values the digital counts need to be enhanced at least
by 10 values, such that two colours can be differentiated by the eye. If the number of levels for each
primary colour is limited to 16, the number of colours that can be generated would be 4096, which per
se Is very high to be readily distinguished by the eye.

3.8.2 Colour display

Displaying colours on a screen consists In giving to each of the three basic components, viz., red,
green and blue, a more or less strong intensity. For example, In order to generate red colour, the level
of red is fixed at 256; however, to obtain pink. It Is necessary to brighten the colour and hence increase
the energy. This Is achieved by enhancing the green and blue components by the same quantity.
It is interesting to note that if several persons are asked to construct a colour, orange for example,
through composition of three colours, the results can be somewhat different. This Is because the term
‘orange’ may correspond to several concepts, such as a mixture of red and yellow or the colour of the
fruit. Depending on the idea of the person who composes the colour and his (her) visual capacities,
the three colours, red, green and blue, are not the same. Thus, orange colour can be produced with
varying digital values:

■^They can be ordered from IGN-2, Pasteur Road-94 160, Saint-Mande, France. For website, see the list
enclosed in the book.
 Physical Interpretation of Data 69

For red: 217 to 255

For green: 81 to 160
For blue: 0 to 30.
The differences are still greater when the colour to be generated is denoted by terms such as
dark brown or beige and, a fortiori, when qualitative terms such as café au lait, chocolate, apple green
etc. are employed. This only proves the paucity of our vocabulary for denoting the thousands of colours
we are able to distinguish.
It is nevertheless to be noted that the number of colours that can be produced on the screen is
most often greater than what we can identify with our eye.
Black colour results from absence of colour emission, but on a colour screen what is perceived as
black is often a green or a blue or a very dark brown. Similarly, white colour is a mixture of equal
proportions of the three basic colours; however, it is common to see on a colour screen a ‘white’ when
a very high intensity of pixels tends towards very bright green or red.

3.9 USE OF COLOURS IN IMAGE PROCESSING

Almost all Image-processing systems are based on colour coding. However, two approaches can be
distinguished.

3.9.1 Colour code of an image (8 bits)

The colour seen on the screen corresponds to a colour table known as Look Up Table (LUT), which
assigns a colour to a value of a pixel, most often coded on 8 bits.
Any colour can be arbitrarily assigned to any value of a pixel, as in the case of classification. It is
thus possible to assign to each class one of the 256 possible colours. Thus, a very large number of
colour representations can be used in a single classification. These representations are important
since they will be interpreted visually by the reader and, depending on the chosen colours, the felt
impression may be different. Classes assigned to neighbouring shades of colours may be mixed up
and be interpreted as a single group. If such classes are to be Interpreted as separate, they should be
allotted clearly distinguishable colours. For example, complementary colours ought to be used for
zones that have common boundaries because they stand out better.The choice of colours also depends
on the colour perception of the person viewing them! Lastly, it Is to be remembered that colours have
a symbolic significance, which varies with cultural background and context. For example, white is the
festival colour In Europe versus red in the East, and white the colour of mourning In the East versus
black In Europe— except for mourning the queen of France, which was observed in white. Leafy trees
are green, but only for four or five months in temperate climate. Water is blue, but only when the sky is
blue and not when in a glass. Soil is maroon, but this excludes white, red, black or brown, etc. soils.

3.9.2 Interpretation of a 3-by-8 bits code

A colour view on the screen corresponds to a look-up table coded in 3-by-8 bits. Such is the case
when the RGB systems are used on a computer monitor. In different systems, it is possible to erase
one or the other colour, and it is seen that for each pixel the intensity varies with each of the colours,
blue, green and red. There are 256 possible intensities for each colour.
Colour composition is very commonly employed to create a colour or colour-infrared Image from
three channels of a satellite image. However, care should be taken in interpreting an image thus
70 Processing of Remote Sensing Data

coloured when it is generated from different wavelength bands, such as TM3, TM4, TM7, etc. The
same is true when a diachronic image is prepared assigning, for example, blue to visible red (675 nm)
for a winter image, green for that of spring and red for a summer image:
— Spring crops appear dark cyan. In fact. In winter the area is a bare soil and hence appears blue.
In spring, the crop is in growth and there Is less green than blue. In summer, the crop is full of chlorophyll
and there would be little red in the image (see Figs. 4.12 and 4.13).
— Winter crops appear medium red. The plot in winter is chlorophyllous and hence blue of low
intensity in the image. In spring, the crop is full of chlorophyll and there is less green than blue. In
summer, the crop is harvested, leaving stubble in the field, and hence much red in the Image.
— Mountain pastures are covered by snow In winter and their hue is intense blue. In spring, bare
organic soil is visible and shows as very pale green. In summer, the young vegetation is highly
chlorophylllan and so very little red appears in the image.
Obviously, colours composed for such images change if other wavelength bands are used. Thus,
for the same objects and the same colour combinations for the same seasons, but coding the near
infrared channel instead of the visible red (675 nm), spring crops would be yellow, winter crops appear
very light yellow, while mountain pastures would look magenta.
Interpretation of colours can become complex If different channels are used for different seasons
and If positive channels are combined with the negative. In such a case, a look-up table should be
compiled, indicating for each wavelength the values of digital numbers and the corresponding colour
for each of the three components used, to obtain the resultant colour composite.

3.9.3 Choice of colours

For clearly distinguishing various classes during processing. It is recommended that colours be chosen,
at least initially, in such a way that neighbouring pixels belonging to different classes on an image are
given quite distinct colours of high contrast. This provides better detection of the boundaries between
two groups and hence correct decisions regarding the digital value of the dividing line between them.
At the end of processing, colours should be chosen in such a way that the user can read the
document with utmost ease. This depends not only on his preference of colours, but also the manner
in which the document will be viewed. It may be preferable to use highly contrasting and bright colours
if the document is to be only seen at a glance. On the other hand, if the material Is to be studied for
some time, more harmonious colours may be used.
The choice of colours in the legend should also correspond to the symbolism of colours. The
succession of green, orange and red can evidently be used to indicate a gradation from good, inadequate
to bad. Two extreme situations of the legend can be distinguished using contrasting ‘hot’ and ‘cold’
colours.
Colours are to be chosen taking Into consideration the surface area of each unit in the final
document, as well as the vicinity between colours. For example, red bordering green gives an impression
that the red zone ‘overflows’ the green: the green area appears smaller than It actually is. Only the
surrounding colours need to be changed in order to perceive the effect of various colours. Hence, the
choice of colours has an immediate influence on what Is perceived at first glance.
A document of successful colour composition ought to facilitate its reading at various levels. At
first glance, one should be able to differentiate major units, estimate their relative areas, their dispersion
in the zone under study and their organisation vis-à-vis other units. It is necessary that during an
advanced stage of spatial analysis, units situated in secondary levels of a semantic hierarchy be
differentiated. Several solutions exist for this. One Is to Indicate first-order units in colours representing
different values: yellow is always very bright whereas maroon Is always dark. Second-order units are
represented in the same colours as those of the first-order but with more or less pure shades.
 Physical Interpretation of Data 71

3.9.4 Colour printing of documents

A colour image prepared on screen may get distorted when transferred to a colour printer. It needs to
be verified whether or not the computer code of the colouring system of the image-processing software
is coherent with that of the printer. If it is not, the computer folder of the processed image should be
routed through a graphic system that is per se compatible with the printer.
It should also be remembered that the inks used in a printer are made up of pigments, and all the
colours possible cannot be produced with the pigments available. Such a situation Is observed In
pages oif the Munsell code, wherein colours exist for which co-ordinates can be defined but which
cannot be printed (Fig. 3.7). It is therefore possible to view colours on the screen which cannot be
reproduced on a colour printer.
Lastly, it should be noted that when documents are exposed to daylight, different inks fade at
different speeds: red disappears most rapidly and blue remains the longest. Hence, on documents
that may be exposed to sunlight, features considered most important should not be represented in red
since they will be the first to fade.
For more details about colours, refer to the various books on graphic semiology and, in particular,
the one by Bertin (1967).

3.10 CONCLUSION
The minimum knowledge presented here on colour perception by the eye is necessary to interpret
satellite and aerial images and photographs in colour or colour infrared. It should be remembered that
visual interpretation is generally employed alone or combined with computer processing of images. In
fact, the eye has the capacity of interpretation of forms of objects and their distribution, which is
superior to what can now be achieved through modelling and structural processing of images.

References
AGFA. Introduction à la numérisation: Le prépresse couleur assisté par ordinateur, vol. 4, 41 pp.
Bertin J. 1967. Précis de sémiologie graphique, les diagrammes, les réseaux, les cartes. Mouton et Gauthier-
Villars, Paris, 431 pp.
EscadafaI R. 1989. Caractérisation de la surface des sols arides par observations de terrain et par télédétection.
Application: exemple de la région deTataouine (Tunisie). Études et Thèses. Orstom, Paris, 317 pp.
EscadafaI R. 1993. Remote sensing of soil color: principles and applications. Remote Sensing Reviews, 7: 261-
279.
EscadafaI R, Girard M-C, Courault D. 1988. La couleur des sols: appréciation, mesure et relations avec les propriétiés
spectrales. Agronomie, 8:147-154.
EscadafaI R, Girard M-C, Courault D. 1989. Munsell soil color and soil reflectance in the visible spectral bands of
Landsat MSS and TM data. Remote Sensing of Environment, 27:37-48.
Girad C-M, Girard M-C. 1975. Applications de la télédétection à l’étude de la biosphère. Masson, Paris, 186, pp.
ICI. 1932. ICI Proceedings 1931. Cambridge Univ. Press, Cambridge.
Lliboutry L. 1992. Sciences géométriques et télédétection. Masson, Paris, 289 pp.
Monget J-M. 1986. Cours de télédétection, 1. CTAMN, Octobre, Sophia-Antipolis.
Munsell Color Company. 1975. Munsell Soil Color Charts. Munsell Color, Kollmorgen Corp., Baltimore, MD.
Ohtha N, Wyszecki G. 1997. Location of the nodes of metameric color stimuli. Color Res. Appl., 2:183-186.
Takahama K, Nayantani Y. 1975. New methods for generating metameric stimuli of object color. J. Opt. Soc. Amer.,
62:1516-1520.
Wyszecki G, Stiles WS. 1982. Color Science: Concept and Methods, Quantitative Data and Formulae. John Wiley
& Sons, NY, 950 pp.
 4
Spectral Characteristics
Interaction of radiation with matter constitutes the basis for interpretation of remote sensing images.
An object situated in a given geographic position at a given moment, viewed under a given field of view
and receiving a given radiation, exhibits a spectral behaviour that is specific to it. Hence, some authors
use the term spectral signature. This term is inappropriate since a signature implies constancy whereas,
in reality, the spectral behaviour of an object varies with time, place, mode of data acquisition and
incident radiation.
The spectral behaviour of objects is an important means of analysis and interpretation of remote
sensing images since it Is based on general laws of physics. Therefore, as will be seen later, this
approach facilitates formulation of an interpretation model that can be generalised for a large number
of cases.
Although the objects under study may be many, they can be reduced to a few general cases such
as vegetation (organic matter), soils (mineral matter), water, snow and ice. The first three constitute
the major targets of remote sensing of the Earth.

4.1 VEGETATION
The term vegetation is commonly employed in a very general sense to refer to the existence of
photosynthetic activity. Optical properties of vegetation covers depend on the plants they contain,
their spatial arrangement and the underlying soils. Spectral behaviour of isolated leaves needs to be
distinguished from that of an individual member of a given plant species and from that of a population
belonging to the same species and the same variety. For the purposes of accurate investigations, it Is
therefore necessary to differentiate between different species and, for a given species, between various
phonological stages and physiological states. Moreover, spectral behaviour of a single-species
population needs to be distinguished from that of a multi-species population, for example alfalfa
field from a permanent grassland. In fact, in such cases, a change is observed from a more or less
regular repetition of spacing between individual plants, belonging to a single vegetation species, to an
irregular distribution of individuals pertaining to various species that differ morphologically and
physiologically.
The major types of spectral characteristics of vegetation are described in this section. It should be
remembered, however, that like all living beings, there are as many forms as number of individuals.
Hence, the spectral characteristics obtained from laboratory measurements under controlled conditions
(in particular. In diffuse light) are discussed first, followed by results of in situ field measurements In
direct illumination. It should be noted, however, that in the case of dense chlorophyllian vegetation (for
which spectral contribution of soil is negligible), the spectral characteristics are similar irrespective of
the precision level of data acquisition and the species considered, although the reflectance values
may vary significantly.
 Physical Interpretation of Data 73

The spectral characteristics presented here were obtained by different laboratories of the United
States Department of Agriculture (USDA) who carried out the first Investigations on cultivated species
(Allen et al., 1969; Knipling, 1970).

4.1.1 Laboratory measurements

The spectra of reflectance, absorptance and transmittance determined In the laboratory (Fig. 4.1 ) are
valid for ail chlorophyllian leaves irrespective of the species. Three types of spectral behaviour,
corresponding to the major spectral domains, can be identified:
— in the visible spectrum, vegetation exhibits a particular behaviour due to the existence of
chlorophyll pigments;
— in the near infrared, spectral behaviour Is related to the structure of tissues;
— in the reflective middle infrared, the water content of the tissues has the main Influence (CD
4.1).
Reflectance and transmittance vary In approximately the same manner. The parameter Investigated
In remote sensing applications for vegetation is not actually reflectance, however. On the contrary, it is
the efficiency of interception of radiation and hence absorptance which is normally used for yield
assessments, as this controls the production of biomass. An indirect estimation is possible using
linear relationships between the radiation reflected by vegetation (vegetation Indices, see below) and
yield. Other methods based on agro-meteorological models are mentioned in Chapter 22.

Structure of
• PigmentSi vegetation Equivalent thickness of water

I Visible I Near infrared Reflective middle infrared i

Fig. 4.1 : Absorptance, reflectance and transmittance for chlorophyllian vegetation (after Guyot, 1997).

■ Visible domain (380-700 nm)

In the visible spectral domain, vegetation exhibits a low reflectance (maximum 15%), with a maximum
at 550 nm. The latter is due to pigments, in particular chlorophyll a and b, which have two absorption
bands In the blue (450 nm) and red (660 nm) domains. Other pigments such as carotene, xanthophylls.
74 Processing of Remote Sensing Data

etc., also have absorption peaks in the visible band but the absorption peaks of chlorophyll usually
mask them. Effects of other pigments on spectral characteristics are significant only for very young or,
on the other hand, senescent vegetation, or one subjected to stress or deficiency or for plants with
albino or variegated leaves. Laboratory measurements with DK2 Beckmann spectrophotometer on
leaf samples of various species of permanent grasslands {Plantago media L., Trifolium pratense L.,
Brachypodium pinnatum L. (Beau.), Festuca pratensis Huds., Prunella vulgaris L., Primula veris L.,
Salvia pratensis L., Sanguisorba m/nor Scop.) for which the contents of different pigments [chlorophyll
(a+b), chlorophyll (a) and carotene] were also determined, showed an Inverse relationship between
reflectance at 550 nm and chlorophyll a content (Fig. 4.2(a)). The greater the chlorophyll a content, the
lower the reflectance at 550 nm.

Content of chlorophyll a Chlorophyll (a + b)

carotene

R550 10 20 30 R675

(b)
Fig. 4.2: a. Inverse relationship between chlorophyll a content and reflectance at 550 nm. b. Inverse relationship
between reflectance at 675 nm and ratio of chlorophyll (a+b) / carotene.

Studies by Gaussman et al. on maize leaves (Zea mays L.) of different ages, in which the chlorophyll
content decreases with age, showed that reflectance increases with age (Table 4.1). Similar observations
were reported for cotton leaves {Gossypium hirsutum L).
At 675 nm, an inverse relationship was observed between the ratio of chlorophyll/carotene and
reflectance (Fig. 4.2(b)).
Chlorophyllian vegetation has a lower reflectance in green-yellow and orange spectral bands
than does senescent vegetation in which the chlorophyll content is less.
Reflectance of vegetation is affected by chlorophyll content when the latter is greater than 3
pg cm“ 2 and lower than 10 pg cm"^. For values outside this limit, a lower but constant reflectance is
observed due to the optical characteristics of vegetation tissues. In the case of mature leaves with
similar chlorophyll content, differences in reflectance observed in laboratory measurements between
various species are mainly due to a waxy cuticle (sclerophytes or succulent species subjected to xeric
conditions) or dense pilosity on the leaves, which increases reflectance. These differences may not be

Table 4.1: Inverse relationship between age, chlorophyll content and 550 nm-reflectance of maize leaves

Chlorophyll content Reflectance at 550 nm

Young leaves 15.8 mg/L 13.7%

Old leaves 13.9 mg/L 14.4%

 Physical Interpretation of Data 75

detected during field measurements of reflectance, partly because the radiometers often have very
wide spectral bands and partly because the entire vegetation canopy, and not Individual leaves, is
measured.
The visible band spectral behaviour of senescent or dry vegetation totally differs from that of
chlorophyllian vegetation. In the entire visible spectrum, for a given species the reflectance of green
leaves is always lower than that of dry leaves (Fig. 4.3).

Wavelength (nm)

Fig. 4.3: Reflectance curves of green (of a year) and dry leaves (preceding year) of the chalk false-brome
(collected from the same individuals).

Diseases, parasites, mineral deficiencies, etc. affect the chlorophyll content and can be detected
using the visible band. However, the effects of such must be very severe to enable identification
without ambiguity. Moreover, as pigment variations may be produced by extremely diverse sources,
detection of particular spectral characteristics provides no information as to their causes.
Pigment contents may also vary in the presence of coloured flowers, as in the case of permanent
grasslands. Such variations can be distinguished from laboratory spectrophotometric measurements
of inflorescence (or parts of inflorescence) of various species of grass (Fig. 4.4).
White flowers— ray florets of daisy, inflorescence of white clover— exhibit strong reflectance for all
wavelengths of the visible spectrum.
Yellow flowers— Ragwort, bulbous buttercup and horseshoe vetch— ^have very low reflectance in
the blue band and maximum reflectance in the yellow.
Violet flowers— Red clover, Self-heal, knapweed— present very low reflectance in the green band
and, depending on the colour shade, higher or lower reflectance in blue and red domains. Thus, the
inflorescences of red clover have lower reflectance in the blue than do Self-heal and knapweed.
These differences are perceptible In field measurements on permanent grasslands containing
abundant flowered species (Fig. 4.5). To facilitate a ready comparison of reflectance values In blue
and red, these curves have been offset by 10% reflectance at the wavelength 550 nm.
The presence of many flowers, yellow (primrose), white (cuckoo flower) or orange (marsh marigold),
is responsible for strong reflectance, particularly at 650 and 675 nm. This Increase is greater for marsh
marigold than for other species, since these large-size flowers contribute more to the spectral behaviour
by partially masking the spectral contribution of the foliage. The marsh marigold grasslands can be
spectrally differentiated from one another in places where bare soil is visible through the vegetation
because of its low reflectance values at 400 to 500 nm.The effect of bare soils on spectral behaviour
of low-density vegetation cover Is discussed further in this chapter. On satellite images, only cultivated
vegetation canopies (such as rapeseed, sunflower, linseed etc.) or fallow lands (Phacelia tanacetifolia)
contain an abundance of flowers sufficient to modify spectral behaviour perceptibly.
76 Processing of Remote Sensing Data

Beilis perennis (English Daisy) X (nm)

White
......... Trifolium repens (White Clover) flowers

------- Senecio jacobea (Ragv\/ort)

^ Yellow
------- Ranunculus bulbosus (Bulbous Buttercup) ’ flowers

-------Hippocrepis comosa (Horseshoe Vetch)

------- Trifolium pratense (Red clover)

Violet
------- Prunella vulgaris (Self-heal) flowers

------- Centaurea jacea (Knapweed)

Fig. 4.4: Laboratory-measured curves of reflectance for different colours of inflorescence.

■ Near infrared (700 nm to 1.3 |Lim)

In the near infrared, pigments do not influence spectral behaviour since the quantity of radiation
absorbed by leaves is very small. The Internal structure of foliage, In particular various tissues and
arrangements of cells and intercellular space, are principally responsible for differences in reflectance.
A simplified transverse section of a leaf of a flowering plant consists of the following tissues, from the
upper surface to the lower (Fig. 4.6):
— waxy cuticle of variable thickness,
— layer of cells forming an epidermis,
— more or less developed hypodermis,
— palisade parenchyma consisting of elongated and well-arranged cells, mainly containing
chloroplasts,
 Physical Interpretation of Data 77

REFLECTANCE
(%)

X (nm)
78 Processing of Remote Sensing Data

— mesophyll whose irregular-shaped cells form a loose network with a number of intercellular
gaps filled with air or water vapour,
— layer of cells forming an epidermis in which stomata that ensure leaf transpiration are located,
— thin cuticle.
The leaves of plants can be arbitrarily classified into two groups. The first group contains a well-
developed palisade parenchyma and a less-developed mesophyll (type A). The second group has a
less-developed palisade parenchyma and a well-developed mesophyll (type B). A lower reflectance in
the near-infrared band is observed for type A than for type B. Spectrophotometric measurements
coupled with microscopic observations of thin sections of leaves have shown the significance of the
internal structure of tissues in spectral characteristics of leaves. This is due to the discontinuities in
refraction indices between cells, water and air of intercellular spaces and lacunae of the mesophyll.
In fact, the path of radiation differs in different tissues depending on their refractive indices. A light
ray incident at an angle 0;On the interface between two media of refractive indices and may be
reflected at an angle 0^or refracted at an angle 0^^ In the case of specular reflection 0^= 0^. According
to Descartes’ law, the angle of incidence 0^and the angle of refraction 0^^are related by the equation:
X sin 0/= /?2 X sin 0^^(Fig. 4.7). The refractive indices for air and water are 1 and 1.33 respectively,
and for the elements constituting the vegetation cells 1.55 approximately. When a light ray propagates
from a high refractive medium to a low refractive one (n^ > 772), Descartes’ law apply for angles for
which sin Q¡< 1 or sin Q¡= /?2/n^- Beyond this angle of incidence (0^> arcsin (772/ 77^)), the ray is totally
reflected.

A more or less strong reflection may occur on the cuticle. In palisade parenchyma, radiation Is
transmitted with a small path deviation. In fact, the cells in this parenchyma are roughly in the shape of
parallelepiped and regularly aligned, without large lacunae between them and hence the refractive
indices of the media traversed by light are similar. On the other hand, cells in mesophyll are more or
less irregular spherical in shape and have many lacunae (sometimes of large size) between them
containing air or water vapour. The presence of media of varied refractive Indices gives rise to the
situation of total internal reflection (for example during propagation from a cell to a lacuna) and hence
a stronger possibility of reflection towards the upper surface of the leaf. This effect of intercellular
lacunae and voids was detected from the decrease in reflectance at 800 nm during filling of these
voids by water.
Physiological changes (pigment content, internal structure of tissues, water content) accompanying
the processes of maturation and senescence produce significant changes in the spectral behaviour in
visible and infrared bands. Senescent vegetation (yellow and dry) is generally characterised by higher
reflectance in these two spectral domains than green and turgescent vegetation. In the near infrared,
 Physical Interpretation of Data 79

this phenomenon is due to the combined effect of changes in internal structure and water content.
Senescence induces collapse of the cells of mesophyll, leading generally to horizontal layering of cell
walls, accompanied by loss of water content. The decrease in reflectance in the near Infrared, observed
in the laboratory, for sick or nutrient-deficient plants (Fig. 4.8) corresponds to deterioration of cells and
collapse or smaller tissue thickness.

Fig. 4.8: Variation in reflectance of a leaf according to its physiological state.

Lastly, a small absorption occurs at 0.98 and 1.20 pm due to liquid water present in leaves. This
effect Is ordinarily masked by the very strong absorption of atmospheric water vapour. It can be detected
in laboratory or field measurements but not In satellite data which are acquired in very broad bands.
On the other hand, it has been noticed In the records of AVIRIS (Airborne Visible/Infrared Imaging
Spectrometer).

■ Reflective middle infrared (1.3 to 2.5 ]im)

The spectral characteristics of leaves are mainly affected by the water content of cells, as water has
absorption bands at 1.45, 1.95 and 2.5 pm. A green, turgescent plant has low reflectance in the
reflective middle infrared. If the water content decreases, either due to senescence and drying or
following diseases and parasite attacks, its reflectance in this spectral band increases (Fig. 4.8). The
bands 1.4 to 1.8 pm and 2.1 to 2.35 pm are more sensitive to variations in water content than the band
1.9 to 2.05 pm. Recent studies have shown that components such as lignin and cellulose also influence
the reflectance of vegetation in this spectral range. Vegetation rich in lignin has high absorption at 1.72
pm, which increases with decrease in water content.

4.1.2 Field measurements

Spectral behaviour of vegetation inferred from laboratory measurements of leaves was described in
the preceding section. What would it be when determined in the field for a vegetation cover? Several
cases of vegetation are possible: total soil coverage (100%), low (< 20-30% ) and intermediate. In the
first case, spectral behaviour is similar to that of leaves measured in the laboratory, with variations
Induced by the structure of vegetation cover (growth and orientation of leaves, number of leaf layers,
etc.). In the second case, spectral behaviour is close to that of bare soil while in the third case, all
intermediate variations between the preceding two are possible.
In the case of more or less dense vegetation cover. In addition to the phenomena described in the
preceding section, the processes associated with the structure of the canopy, mainly related to the
80 Processing of Remote Sensing Data

disposition of leaves, need to be considered. Depending on the angle of inclination of the leaves to the
horizontal, three types of vegetation can be distinguished. These are planophyll (angle close to 0°),
erectophyll (angle close to 90°) and intermediate types, viz., plagiophyll (inclined leaves most frequent)
and extremophyll (erect leaves most common). Dicotyledons are usually planophyll or plagiophyll
whereas monocotyledons are erectophyll or extremophyll. Nevertheless, the same individual of a
given species may exhibit significant angular variations in the course of Its development. Such is the
case in particular for grass in which the angle of leaf inclination decreases during maturation. Considering
these differences in the disposition of leaves, the same leaf area index‘d leads to a smaller coverage
for erectophyll species than for planophyll. The differences in the leaf attitude more or less affect the
magnitude of shadow produced in the canopy. In the case of cultivated plants, the effect of leaf orientation
is enhanced by the row effect. The latter is also sensitive in the near infrared for an erectophyll vegetation
(leaf area index 10) for a field of view inclined at about 30° relative to the vertical, while it is not
sensitive in the case of planophyll vegetation (even for low leaf area indices).
The reflectance of a vegetation cover varies with its phenological stage and physiological state
(Fig. 4.9).

Fig. 4.9: Field-measured reflectance for vegetation cover in various physiological states.

In field measurements over senescent vegetation, reflectance in the near-infrared band is found
to decrease. This decrease corresponds to the combined effect of changes in the internal structure of
leaves and in the contribution of underlying soil to reflectance of vegetation, due to changes in leaf
orientation as they become less dense. Consequently, the role of shadow produced becomes more
important, which also contributes to reduction in reflectance.
In the reflective middle infrared band, for vegetation covers with similar water content, a greater
absorption of radiation Is observed in vegetation with the highest biomass. This band is sensitive to
quantity of total water per unit area, i.e., the ‘equivalent water thickness’ defined by Allen et al. (1969).
In the case of total chlorophyllian vegetation cover (soil not seen and no spectral contribution
from it), saturation of reflectance is observed starting from a leaf area index of 2 in the visible band and
8 in the near infrared. The saturation effect Is reached at smaller leaf area Indices for planophyll plants
(such as beet) than for erectophyll plants (such as grass).
Moreover, multicoloured inflorescences (see supra) modify spectral behaviour in the visible band.
Non-chlorophyllian standing dry matter mixed with chlorophyllian aerial phytomass likewise
modifies the spectral behaviour of the entire coverage to a varying degree, depending on the relative

^Leaf area Index is the ratio of surface area developed by all the leaves to surface area occupied on the ground
(Heller R. 1977. Abrégé de physiologie végétale. Masson, Paris).
 Physical Interpretation of Data 81

proportions of the two mixing phytomasses. Such types of vegetation pose problems in estimating the
fraction of chlorophyllian biomass.
In the case of low density of chlorophyllian vegetation, the spectral contribution of soil to the total
spectral reflectance of bare soil and vegetation varies according to the spectral band used and the
nature and state of soil. In fact, in the visible domain, bare soils most often have a higher reflectance
than vegetation, while the reverse situation occurs in the near infrared.
When the coverage by soil equals that of vegetation, the more reflective the soil (very dry, low
organic matter and iron content, smooth surface, etc.), the greater Its spectral contribution in the
visible band, and hence more readily detected. However, the spectral behaviour of the vegetation
cover in the near Infrared is not modified. Contrarlly, a non-reflective soil contributes little to the spectral
reflectance of vegetation in the visible, but significantly reduces it in the near infrared. This indicates
that the proportion of vegetation cover at which the influence of soil on spectral behaviour becomes
perceptible is highly variable. Our own investigations concerning Neoluvisols show that in the near
Infrared, the presence of green vegetation cannot be detected when the vegetation coverage is less
than 20%, while a spectral behaviour characteristic of chlorophyllian cover is obtained when the coverage
is more than 40%. Between these two values of coverage, the reflectance curves acquired in the field
resemble neither those of the soil nor those of the chlorophyllian vegetation (Fig. 4.10). In-situ
measurements at a height of 2 m above fields of Panicum maximum in different phenological stages
and for different degrees of coverage Illustrate this effect. The spectral behaviour of chlorophyllian
Panicum with low coverage does not resemble that of vegetation.

I I Chlorophyllian, coverage close to2 0 %

I I Yellowing, coverage close to 70 %
I..■-"! Chlorophyllian, coverage greater than 80 %

Fig. 4.10: In-situ measured reflectance graphs for Panicum maximum.

Similarly, a maize field, with a coverage of less than 20% (Fig. 4.11), exhibits a reflectance curve
close to that of soil and does not indicate the presence of chlorophyllian vegetation.
The effects of spectral contribution of soil and different phenological stages for winter corn measured
on different dates during the cultivation period are illustrated in Fig. 4.12. From October to June, a
gradual variation of the reflectance curves from bare soil to chlorophyllian vegetation and back to
senescence can be seen.
In the reflective middle infrared domain, for homogeneous vegetation cover such as annual crops,
various authors obtained statistically significant correlation between reflectance values and water
content of plants. Contrarily, in the case of heterogeneous covers of permanent grasslands (Orth,
1996), no such statistically significant correlation (/^ = 2.04 x 10“ ^) was observed, except when different
types of grasslands were separated according to height; in the latter case, significant correlation (r^ =
0.90) was observed only for homogeneous and low canopies.
82 Processing of Remote Sensing Data

Fig. 4.11 : Spectral behaviour of a maize field (coverage < 20%).

Fig. 4.12: In-situ measurements of reflectance on different dates for a winter corn field.

The foregoing discussion Illustrates the Influence of the structure of vegetation cover and indicates
that results obtained for cultivated vegetation cannot be extended to natural or seminatura! covers.

4.2 SOILS
Spectral characteristics of soils are discussed in detail in Chapter 23. In this chapter, the basics of
spectral behaviour in the visible and reflective near- and middle-infrared bands are described, which
constitute the minimum knowledge required for understanding vegetation indices.
Irrespective of whether the measurements are carried out in the laboratory or in the field, the
reflectance of soils uniformly increases in the wavelength range from 400 to 1450 nm. It differs markedly
from the reflectance of chlorophyllian vegetation: in the visible band, reflectance values of soil are
usually higher than those of vegetation, hovering around 675 nm, whereas they are commonly lower
than those of vegetation In near infrared.
Various soil horizons can be distinguished by means of a field radiometer (if care is taken to clear
them in such a way that in-situ measurements on a horizontal plane can be obtained). For a Brunisol
(Soil Reference Manual (SRM), 1995), the reflectance values can be interpreted when they are
compared with the colours of horizons, organic matter content, total calcium content and moisture
content (Fig. 4.13).
Horizons LI and L2, which are very similar, cannot be differentiated (Table. 4.2). Horizons S1 and
S2, which differ from the preceding horizons in colour and organic matter content, are distinguishable.
Horizons C l and C2 are identified from their total calcium content. Horizon C3, on the other hand,
differs from horizon C2 by Its lower moisture content.
 Physical Interpretation of Data 83

Table 4.2: Results of analysis of a Brunisol (see reflectance variations in Fig. 4.13)

Horizon name Depth (cm) Munsell Organic Total Moisture (%)

(SRM) colour matter (%) calcium (%)

LI 0-5 10YR3/3 5.1 3 28

L2 5-30 10YR3/3 5.0 2 26
SI 30-60 10YR4/4 1 .2 0.5 2 0

S2 60-105 10YR4/4 1 .2 0.5 25

Cl 105-130 10YR5/6 0 .8 32 26
C2 130-170 10YR7/4 0 .8 63 23
C3 170-220 10YR8/3 Not analysed 6 6 19

i k Reflectanee (%)
60
C3

50
^ — —

40
/
/
/ o1
30 /
/ X
/ X S2
— S1
20 X _____ L2
X
L1

10

0 lllr
500 600 700 800 900 1000 1100 (nm)

Fig. 4.13: Reflectance of various horizons of a Brunisol (measurements carried out with an EXOTEC radiometer
using MSS spectral bands. For legend, see Table 4.2).

It can be seen that all the curves are approximately similar and reflectance increases with
wavelength. Their major differences are expressed by the value of their integral. Spectral behaviour of
soils can be Interpreted by studying the shape of the curves (Fig. 4.13).
Lastly, while comparing reflectance curves of soils, the time of measurement should be taken into
consideration, since the values change with time and moisture content, which varies with height of the
Sun (Fig. 4.14).
In the reflective middle infrared band (1.6 to 2.2 pm), reflectance depends on water content. As
soils always contain water, the absorption bands of water at 950,1150,1450,1950 and 2450 pm can
be observed on the curves obtained using a radiometer with narrow spectral bands.
Reflectance of soils is dependent on a number of internal and external factors, which can be
classified as follows (see Chapter 23):
— Surface roughness, which depends on external factors, such as microtopography and crop
(agricultural) activity, and on Internal factors such as salt efflorescence, ferruginous encrustations.
84 Processing of Remote Sensing Data

-— -o Clear soil (10YR 7/1)

\
V

7 8 9 10 12 14 16
Solar time

Fig. 4.14: Temporal variation of moisture content of a clear soil (10YR 7/1), 2 May 1989
(after Yongchalermchai, 1993).

cracks, gilgai, structure and porosity, slaking crusts^, etc. Whatever be its origin, surface roughness
modifies reflectance according to the amount of shadow viewed by the sensor. Reflectance is Inversely
proportional to the degree of roughness.
— Physicochemical components, such as organic matter, calcium, iron, water content, grain size,
salts, etc., which Increase or decrease soil reflectance.
Organic matter, iron and water content reduce soil reflectance in the visible and infrared bands.
The higher the concentration of any (or several) of these elements, the smaller the reflectance. Contrarily,
calcium and some salts enhance the reflectance of soils. Finally, the grain-size distribution either
increases or decreases soil reflectance under different conditions. In fact, fine particles retain more
water compared to coarse grains, leading to higher moisture content, which consequently reduces
spectral reflectance.
A detailed description of the Influence of physicochemical properties of soils on reflectance Is
given in Chapter 23. It may be remembered that in soil science, as in any biological discipline, factors
involved are interdependent and variation of one leads to modification of another (or others).
Interpretation of the spectral characteristics of soils hence necessitates a good understanding of soils
and their functioning.

4.3 VEGETATION INDICES

Since the time satellite recording of spectral radiance of ground objects in visible and near-infrared
bands became possible, many authors have developed various Vegetation indices’, based on a certain
combination (sum, difference, ratio, linear addition) of intensities of these channels. These Indices are
used, on the one hand, to Identify and monitor temporal variation of vegetation cover and, on the
other, to estimate certain parameters such as aerial phytomass. Moreover, these combinations have
the advantage of reducing the effect of factors external to vegetation, such as solar Irradiance,
atmospheric influence, spectral contribution of soils, etc.

4.3.1 Identification of vegetation cover

Identification of vegetation is based on differences between spectral characteristics of vegetation and
soil (see supra). These differences are particularly distinct In the red band (around 675 nm) of the
visible and near-infrared (800 to 1000 nm) channels, which are hence most often employed.
 Physical Interpretation of Data 85

4.3.2 E stim ation of vegetation param eters

Production of chlorophyllian aerial biomass by vegetation depends on the efficiency of interception Sj,
which Is equal to the ratio of the photosynthetically active radiation absorbed (PARJ to the
photosynthetically active Incident radiation (PARy) (Fig. 4.15).

Fig. 4.15: Analysis of energy-absorbing efficiency of vegetation (after Varlet-Grancher, 1982). PAR^- reflected PAR;
PAR^: transmitted PAR; PAR^^ PAR of soil radiance; Climatic efficiency: = PAR/Rg « 0.5; Interception efficiency:
8 / = PARg/PARy (dependent on vegetation, LAI, coefficient of extinction etc.); Biological efficiency: 8 ¿, = (c x

A(MS))/ PAR^; A(MS) = increase in dry matter related to photosynthesis; c= energy equivalent = 12x10® J-kg""*.

Interception efficiency evidently depends not only on the leaf area index, but also on the structure
of the vegetation canopy, as illustrated in Fig. 4.16. Alfalfa and beans can be considered planophyllic
(see supra), while maize and sugarcane are erectophylllc. It has been observed that for the first two
types of plants, interception efficiency (cj) becomes asymptotic for leaf area indices between 2.5 and
3, whereas asymptote is reached for leaf area Indices of 4.5 and 5 for the other two crops.
The structure of the vegetation cover should be taken Into consideration while estimating Cyfrom
the leaf area index. Numerous studies (Baret and Guyot, 1991 ) showed that the spectral reflectance of
vegetation cover is related to the total aerial phytomass at the time of measurement and that an
asymptotic relationship exists between leaf area Index of a vegetation and various combinations of
reflectance in the red and near-infrared bands, i.e., the vegetation index. This relationship assumes

Fig. 4.16: Variation of interception efficiency with leaf area index (after Varlet-G rancher, 1982).
86 Processing of Remote Sensing Data

that the geometry of illumination and field of view (see Fig. 1.5) as well as inclination of leaves remain
constant. However, this is not so in reality and the semiempirical relations between vegetation index
and aerial phytomass are valid only locally and for a given instant.
Some vegetation indices are applicable only for vegetation covers that are dense (no soil visible)
but not too dense (otherwise, the saturation effect mentioned at the beginning of the phapter becomes
apparent) and that are chlorophylllan, with no mixing of standing dry matter with green. Other formulae
ought to be used when coverage of chlorophylllan vegetation is low (seedling or harvesting stages of
agriculture and annual steppes in arid zones etc.) or when standing dry matter accumulates, such as
in natural herbaceous stands (Rondeaux et al., 1996).

4.3.3 Use of ‘soil clusters’

The concept of soil clusters Is based on knowledge of the spectral behaviour of soils (see supra and
Chapter 23). Spectral characteristics are related to the degree of surface roughness (presence of
slaking crusts, hardpans, etc.), organic matter content, composition of chemical elements (CaCOg,
Fe, etc.) and water content. If only one of these parameters varies, a linear relationship between the
radiant intensities of visible and near infrared bands can be defined for a soil. This relationship is the
‘soil line’ (Huete, 1988) used as a reference line for studying sparse vegetation covers.
Strictly speaking, there Is no single ‘soil line’ but a cluster (see Fig. 23.16) depending on various
factors that Influence the reflectance. This makes the correction for soil reflectance in vegetation studies
difficult.

4.3.4 Various vegetation indices

Many authors have used vegetation Indices computed from reflectance measured in the field or from
digital values furnished by satellite data for identification of chlorophylllan vegetation and estimation of
leaf area index, canopy development, stress effects, AFAR (Absorbed Photosynthetically Active
Radiation), évapotranspiration and efficiency. These studies have resulted in formulation of several
equations (Table 4.3), with varied coefficients of balance depending on the data and subject under
analysis (CD 7.8).
The vegetation Indices can be grouped into two categories: those characterised by a slope (RVI,
NDVI, SAVI etc.) and those characterised by a distance (PVI and Tassel Cap Index).
The various Indices, in fact, correspond to an empirical approach in estimating the vegetation
parameters. They are more or less sensitive to the spectral contribution from soil (and to its variations
depending on soil characteristics) as well as from standing dry matter mixed with green. Errors In
estimation of vegetation parameters will be high if field spectral data on the components of the pixels
are not available.
The normalised vegetation index (NDVI) should be used cautiously in view of its sensitivity to
atmospheric effects and angular variations (as in large angle scanners such as NOAA-AVHRR). In
fact, the flux reflected by a vegetation cover is higher in the visible red band than in near infrared if the
system is in a back (hot spot) position relative to the sun, rather than facing the sun (against the light).
Consequently, the NDVI values corresponding to the first configuration are systematically smaller
than those corresponding to the second. The most commonly used correction to overcome this effect
consists of taking the highest value of NDVI over a long temporal sequence (about a month, for
example). Neglecting this effect may lead to erroneous interpretations of NDVI in studying variations
in vegetation cover. Hence, satellite remote sensing, notwithstanding its synoptic view and ability to
observe a given region at regular time Intervals, may not be able to furnish all the information desired.
In fact, in subjects of global importance, such as temporal monitoring of vegetation cover variations
over long periods and over vast areas (at regional or continental level) or drought phenomena, etc..
 Physical Interpretation of Data 87

Table 4.3: Most commonly used '/egetation indices

(NIR; Near infrared; R: Red; G: Green; B: Blue)

Index Formula Characteristics Authors

Difference R-NIR Very sensitive to Monget,

atmospheric variations 1980

Ratio RVUNIR/R Saturation at high indices; Knipling,

or of other channels; sensitive to spectral 1970;
Pigment index: G/R contribution of soils and Violller et
atmospheric effects al., 1985

Normalised vegetation NDVI = (NIR-R)/(NIR+R) Sensitive to atmospheric Rouse et al.,

index effects; smaller range of 1974; Tucker,
variation than preceding; 1979
sensitive to variations in view
angle, according to position
TVI= ^(NDVI+0.5) vis-à-vis Sun/hot spot*

Transformed vegetation Attempt to eliminate negative Deering et al.,

index values; stabilisation of 1975
variance

Perpendicular vegetation PVI = a.,NIR - 82^ + constant Decrease in spectral Richardson

index contribution of soils; sensitive and Wiegand,
to various soil parameters 1977

Tassel Cap General formula Orthogonal transformation of Kauth and

a.jG + + a^NIR + a^jNIR 4 channels for minimising Thomas, 1976
sensitivity to spectral
Greenness index (derived GR4 = - i? iG - b 2 R + contribution of soils; complete Jackson, 1983
from the preceding) bgNIR + b4 NIR for MSS channels elimination not possible

Soil-adjusted vegetation SAVI = (1+L) (NIR-R)/ Several indices are derived Huete, 1988
index (NIR+R+L), where L = 0.5 to from this to minimise soil
reduce soil effect effect (TSAVI.MSAVI.etc.)

Atmospheric reduced ARVI = (NIR-RB) / (NIR+RB); Effect of aerosols in Kaufman and

vegetation index RB = R - Y (B-R) Y varies with atmosphere on NDVI Tanre, 1992
aerosol type reduced but sensitive to
spectral effects of soil

‘ Hot spot corresponds to a view angle identical to incident solar radiation (backing sun). No shadows in this
configuration and all areas viewed are illuminated.

large errors may accumulate due to atmospheric perturbations as well as vegetation characteristics,
making the results Ineffective.
It is evident that remote sensing data ought to be integrated into models, often very complex,
together with other parameters of vegetation and environment (climate, soil, hydrological regime, etc.)
for estimating yields (see Chapter 22), water stress, etc. The empirical approach of vegetation indices
should be considered with great care.
88 Processing of Remote Sensing Data

4.4 WATER
4.4.1 Response in visible and near- and middle-infrared bands
The spectral behaviour of water is dependent not only on the water molecules but also on dissolved or
suspended constituents (such as particles, algae, organic matter, etc.), as well as the state of surface
roughness. Incident radiation Is partly reflected specularly, more so when the water is calm and smooth,
and partly refracted and transmitted Into the volume. Specular reflection can be observed in aerial
photographs as well as in satellite images. The phenomenon can be readily detected on aerial
photographs since the same surface appears very white in one photograph and dark in the next. In
satellite images, unless diachronic data are available for the same region, the specular surfaces may
be confused with highly reflective mineral surfaces (such as quarries, alluvial-sand mines, etc.). Pure
water reflects very little in the red and infrared bands. It appears black in black-and-white Infrared as
well as colour infrared photographs. On satellite images the digital values are very low for the
corresponding bands (CD 4.2) and especially in reflective middle infrared.
The spectral response of the sea Is directly related to interaction of several processes operating
on its surface and inside the water mass. The Important ones are:
— agitation of the surface under the effect of wind and surges,
— presence of a floating pollutant (hydrocarbons, wastes etc.), or of pollution from a ship,
— mixing of waters of different densities, buoyancies and temperatures (fresh and salt waters,
rise of cold waters),
— currents,
— suspended sediment load,
— presence of phytoplankton and chlorophyll pigments in waters,
— presence of dissolved substances, etc.
The response of a layer of water In the optical domain, a phenomenon concerned with the land/
sea boundary and the actual water mass, is discussed in relation to two types of application.

4.4.2 Approximate method of bathymetry in clear water

Under the conditions of observation of coastal environments with clear waters, the light signal in the
visible band penetrates a layer of water where it is absorbed. In shallow seas, it may reach the sea­
floor and traverse the water layer in return, thus providing information on depth to the bottom and its
nature. The signal measured over a shallow sea depends on the reflection properties of the floor and
the Influence of the water column standing on it.
The spectral behaviour of the sea-floor can be described by a model of exponential decrease of
the signal with depth at which the floor is situated. Several formulations of this model are reported
(O’Neill and Miller, 1989; and others). All these models are similar and can be represented by the
equation:

R (0, z) = R (0, oo) + [ A - R (0, 00)] exp {kjz)

where R (0, z) is the measured reflectance above a sea-floor of depth z, R (0, oo) is the reflectance
measured above a sea-floor of Infinite depth, A the albedo of the floor, zth e depth of the sea-floor and
/Cythe coefficient of attenuation.
Thus, the reflectance of a water column bounded at the bottom by a reflecting surface at a depth
z, can be described as a sum of the reflectance R (0, oo), of an optically identical water column with no
bottom and the contrast of the bottom relative to R (0, oo) modulated by the attenuation in a two-way
 Physical Interpretation of Data 89

path between the surface and the floor. The parameter kj characterises the rate of decrease of the
incident solar energy with depth. It represents the radiant intensity of the water in a given spectral
band and is related to the depth to sea-bottom. Spectra of diffuse attenuation coefficients for water
with different concentrations of particulate and dissolved material are shown in Fig. 4.17.

Fig. 4.17: Spectra of diffuse attenuation coefficients for descendant illumination for (a) pure sea water, (b) oligotrophic
oceanic water, (c) clear coastal water, (d) eutrophic oceanic water (4-5 mg chlorophyll a per m^) and (e) coastal
water with high concentrations of optically active particulate and dissolved matter (Baltic Sea) (after Maritorena,
1993).

The attenuation coefficient varies with wavelength. For clear waters in particular, the diffuse
attenuation coefficient is small at short wavelengths and exhibits higher values beyond 570 nm.Thus,
for clear waters penetration of radiation is high for channels in the wavelength range of 400-500 nm,
whereas in waters with high concentrations of particulate and dissolved matter, largest penetration
depths are obtained in the range 500-600 nm. Considering the values of attenuation coefficient, only
a few bands are useful for determining the depth and nature of shallow sea-floor. These are TM1
(450-520 nm), TM2 (520-600 nm) and to a lesser extent TM3 (630-690 nm) of LANDSAT, and the
channels b1 (500-590 nm) and less so channel b2 (610-690 nm) and P (510-730 nm) of SPOT For
extracting bathymetric information from remote sensing data of the visible domain, it is necessary to
determine the best possible attenuation for the zone under study and the concerned spectral bands.
This complex process of bathymetric estimation can be simplified by employing a data-processing
method. For a given sea-floor, the transformation Xy = In (/.y - ), where Ly is the radiance for the
wavelength / and the radiance for an Infinite floor, is approximately equivalent to linearising the
signals corresponding to different depths (Lyzenga, 1978). In one graphic presentation in which In (Ly
- L^j) is plotted on the abscissa and In (Ly- L^y) on the ordinate, the points associated with a given
type of sea-floor form a cluster of points scattered along a straight line with a slope equal to k/kj. The
spread of the cluster of points is linearly dependent on the bathymetric level and the deviation of this
cluster Is a measure of the variation in the type of sea-floor. A rotation of the system of axes by an
angle equal to arctan (/c//Cy) gives a new reference framework (Vz, Yf1, V?2,..., V/n-1 ) such that Yzis
theoretically independent of the nature of the sea-floor while YU Is independent of depth. Thus, we get:

Z=aYz+b
90 Processing of Remote Sensing Data

In practice, several methods exist to determine the real absolute Z Use of calibration points
facilitates determination of the coefficients a and b which are assumed constant for all the pixels of an
Image. An application of the method proposed above (Loubersac et al., 1989) by Inversion of the
model of exponential decrease of the signal with depth, linearisation, change of axes, calculation of a
pair of coefficients (a, b) for each of the major types of floor detected by a threshold of Yfl and
calibration by introducing the in-situ measurements Into the model, shows a relative error of 10% for
the bathymetric model for the 0-10 m layer of water. Beyond 20 m the results of the model have no
significance due to loss of the signal and presence of noise of the HRV sensor used. Andrefouet et al.
(1998) described the various approaches and solutions, a critical analysis of the results obtained and
the errors.

4.4.3 Measurement of water colour

When the radiance due to the sea-floor does not interfere with the signal received by the sensor, either
because waters are saturated with particles or because the layer of water is sufficiently thick, information
on the water colour can be extracted from the data of the visible band. In fact, the reflected flux
measured above the surface Is directly related to the concentrations of the constituents present In the
water, pigments associated with plankton, degraded organic matter, suspended material and dissolved
matter. However, as the contribution of the atmosphere to the measured signal is very high, up to 90%,
In the narrow spectral bands used (see later), development of algorithms for atmospheric corrections
is necessary.
Significant advances have taken place in the remote sensing of the colour of water thanks to the
data of the CZCS (Coastal Zone Colour Scanner) sensor of NIMBUS-7 satellite, which became
operational between 1978 and 1986. This satellite has six spectral bands, four in the visible, one in
near infrared and one In the thermal range (Cassanet, 1990). The four bands of the visible domain
positioned respectively at 433-453 nm, 510-530 nm, 540-560 nm and 660-680 nm facilitate
measurement of water colour, especially in the presence of chlorophyll. These bands also enable
atmospheric corrections.
In general, the algorithms developed by means of CZCS experiments for estimating the optical
parameters of sea-water comprise three stages:
— ^Transformation of the digital counts furnished by the satellite sensor into luminance values and
calibration in view of the gradual drift in the sensitivity of the sensor (Sturm, 1983; Singh et al., 1985).
— Atmospheric correction of the total luminance obtained for Isolating the effect of the water layer
under investigation.
— Calculation of the concentrations of pigments and suspended matter, and computation of the
coefficient of absorption by means of specific algorithms based on the analysis of the relationships
between channels, calibrated by in-situ measurements of pigment concentration (Bricaud et al., 1981 ;
Sturm, 1981), or by Inversion of a radiative transfer model (Fischer and Doerffer, 1987).
Results of such processing are used for applications in marine biology and ecology, such as
monitoring of water quality, sedimentary transport and dynamic processes.
As the spectral bands of CZCS sensor are not optimally placed vis-à-vis causes of variation in the
optical quality of sea-waters, we cannot obtain sufficiently reliable qualitative and quantitative parameters
about the colour of water that are useful for problems of littoral management such as water pollution,
for example. Hence, a strong expectation stands before the scientific community to generate data
from new sensors, such as SeaWIFS launched In August 1997 and Envisat/Merls launched on 28
March 2002 for which spectral resolution Is optimised relative to CZCS.
 Physical Interpretation of Data 91

4.5 SNOW AND ICE

Snow, which is made up of fine crystals of Ice varying from 50 pm to 1 mm in size, constitutes a highly
diffusive medium due to its granular structure. Reflectance of snow in the visible band mainly depends
on the content of pollutants, while in near infrared. It varies according to the geometry of the crystals,
i.e., their shape and dimensions. Reflectance of snow In this range of wavelengths is inversely
proportional to the size of the crystals (Fig. 4.18).

Fig. 4.18: Variation of the bidirectional reflectance of snow (computed from a model) with wavelength for various
grain sizes. Angle of solar incidence 40° relative to the nadir (after Fily et al., 1997, p. 455, with permission from
Elsevier Science Publishers).

This Figure explains the stronger reflectance of fresh snow compared to that of older snow (with
coarser crystals) or the reflectance of firn which per se is more reflective than ice. Ice has optical
properties dose to those of water, except in the range 1.55 to 1.75 pm, where it absorbs more radiation
than water.

References
Allen WA, Gaussman AJ, Richardson AJ, Thomas JW. 1969. Interaction of isotopic light with a compact plant leaf.
J. Opt. Soc. Amer., 59:1376-1379.
Andrefouet S, Loubersac L, Mahtorena S, Morel Y. 1998. Mesure de la bathymétrie des zones côtières par
télédétection passive dans le domaine visible. In: Manuel de télédétection océanique. Pêche et Océans.
Gordon, Breach (eds.). Institut Maurice Lamontagne, Canada.
Bardinet C, Monget J-M. 1980. LANDCHAD.Télédétection et géographie appliquée en zone sahélienne du Tchad.
Collection de l’École Normale Supérieure de Jeunes Filles, Paris, no. 12,133 pp.
Baret F, Guyot G. 1991. Potentials and limits of vegetation indices for LAI and APAR assessment. Remote Sensing
of Environment, 35:161-173.
Ben Moussa H. 1987. Contribution de la télédétection satellitaire à la cartographie des végétaux marins: archipel
de Molène (Bretagne, France) Thèse de doctorat, Univ. Aix-Marseille II, 122 pp.
Bricaud A, Morel A, Prieur L. 1981. Absorption by dissolved organic matter of the sea (yellow substance) in the UV
and visible domains. Limnology and Oceanography, 26:43-53.
92 Processing of Remote Sensing Data

Cassanet J. 1990. Satellites et capteurs. In: Télédétection satellitaire. Paradigme, 141 pp.
Deering DW, Rouse JW, Haas RH, Schell JA. 1975. Measuring forage production of grazing units from Landsat
MSS data. Proc. 10th Int. Symp. Remote Sensing Environment, vol. II, pp. 1169-1178.
Fily M, Bourdelles B, Dedieu JP, Sergent C. 1997. Comparison of in situ and Landsat Thematic Mapper derived
snow grain characteristics in the Alps. Remote Sensing of Environment, 59:452-460.
Fischer J, Doerffer R. 1987. An inverse technique for remote detection of suspended matter, phytoplankton and
yellow substances from CZCS measurements. Advances in Space Research, 7 (2 ): 21-26.
Guyot 1997. Climatologie de l’environnement. De la plante aux écosystèmes. Masson, Paris, 505 pp.
Heller R. 1977. Abrégé de physiologie végétale. Masson, Paris.
Huete AR. 1988. A soil-adjusted vegetation index (SAVI). Remote Sensing of Environment, 25:295-309.
Jackson RD. 1983. Spectral indices in n-space. Remote Sensing of Environment, 13:409-421.
Kaufman YJ, Tanre D. 1992. Atmospherically resistant vegetation index (ARVI) for EOS-MODIS. IEEE Trans.
Geoscience Remote Sensing, 30:261-270.
Kauth RJ, Thomas G. 1976. The Tassel Cap, a graphic description of the spectral-temporal development of agricultural
crops as seen by Landsat. Proc. Symp. Machine Processing of Remotely Sensed Data. IEEE Catalogue, no.
76, ch. 1103-1 MPRSD, LARS. Purdue Univ., West Lafayette, IN (USA).
KniplIng EB. 1970. Physical and physiological bases for the reflectance of visible and near infrared radiation from
vegetation. Remote Sensing of Environment, 1:155-159.
Loubersac L, Burban PY, Lemaire O, Chenon F, Varet H. 1989. Nature des fonds et bathymétrie du lagon de l’atoll
d’Aitukai (Iles Cook) d’après des domnées SPOT 1. Photo Interprétation 89-5 et 6 , fasc. 4.
Lyzenga DR. 1978. Passive remote sensing techniques for mapping water depth and bottom features. Applied
Optics, 17 (3): 379-383.
Maritorena S. 1993. Étude spectroradiométrique de la colonne d’eau et des fonds en milieu lagonaire récifal.
Implications sur l’imagerie télédétectée à haute résolution dans le visible. Thèse Univ. Française du Pacifique,
Océanologie, Tahiti, 195 pp.
Monget J-M. 1980. cf. Bardinet & Monget 1980 Monget is the author of the index.
MynenI RB, Asrar G. 1994. Atmospheric effects and spectral vegetation indices. Remote Sensing of Environemnt,
47: 390-402.
O’Neill NT, Miller JR. 1989. On calibration of passive optical bathymetry through depth soundings. Analysis and
treatment of errors resulting from the spatial variation of environmental parameters. Int. J. Remote Sensing,
10(9): 1481-1501.
Orth D. 1996. Typologies et caractérisation des prairies permanentes des marais du Cotentin, en vue de leur
cartographie, par télédétection satellitaire, pour une aide à leur gestion. Thèse INA PG, 149 pp. et annexes.
Richardson AJ, WIegand CL. 1977. Distinguishing vegetation from soil background information. Photogrammetric
Engineering & Remote Sensing, 43:1541-1552.
Rondeaux G, Steven M, Baret F. 1996. Optimization of soil-adjusted vegetation indices. Remote Sensing of
Environment, 55:95-107.
Rouse JW, Haas RH, Schell JA, Deering DW, Harlan JC. 1974. Monitoring the Vernal Advancement of Natural
Vegetation. NASA/GSFC Final Report. Greenbelt, MD, 371 pp.
Singh SM, Cracknell AP, Spitzer D. 1985. Evaluation of Sensitivity decay of CZCS detectors by comparison with in
situ near-surface radiance measurements. Int. J. Remote Sensing, 6:749-758.
Sturm B. 1981. Ocean color remote sensing and quantitative retrieval of surface chlorophyll in coastal waters using
CZCS data. Marine Science Series, 13:267-279.
Sturm B. 1983. Selected topics of CZCS data evaluation. In: Marine Science and Technology. Reidel Publ. Comp.,
Dordrecht, pp. 137-167.
Tucker CJ. 1979. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sensing of
Environment, 8:127-150.
Varlet-Grancher C. 1982. Analyse du rendement de la conversion de l’énergie solaire par un couvert végétal.
Thèse de Docteur Ingénieur ès Sciences Naturelles, Univ. de Paris-Orsay, 144 pp.
Viollier M, BelsherT, Loubersac 1.1985. Signatures spectrales des objects du littoral. Proc. 3rd Int. Coll, on Spectral
Signatures of Objects in Rem. Sens. Les Arcs, France. ESA SP 247, pp. 253-257.
WIegand Cl, Richardson AJ. 1990. Use of spectral vegetation indices to infer leaf area, évapotranspiration and
yield. 1: Rationale. Agron. J., 82:623-629.
 c
PROCESSING AND
INTERPRETATION
 5
Visual Interpretation of
Photographs and Images

5.1 VISUAL INTERPRETATION

5.1.1 The eye
Visual perception differs according to the colour of an image. The eye perceives up to 16 grey levels In
black and white and about 10,000 colours. These figures correspond to the number of levels that can
be recognised absolutely, I.e., when they are not intermixed but are isolated from one another. The
concept of contrast is not taken into consideration In this count. When contrast is involved, the eye
exhibits better performance. In fact, a large number of shades are perceived when intermixed grey
levels are compared, but it is not certain that the eye can reliably measure the same grey levels when
the two levels compared are not close to one another.
The light signal perceived by the eye, by the order of its spatial resolution, is about 0.1 mm. If the
pixel of an Image or a photograph observed Is greater than 0.1 mm in size, it is visible, otherwise it
cannot be perceived. For a SPOT (1 to 4) Image with a resolution of 20 m and in the scale of 1:200,000,
the resolution perceived by the eye of 0.1 mm would be 20 m.Thus, we do not get the Impression of
seeing the pixel. At larger scales, the pixel of the image Is perceived. Thus, at 1:20,000 scale the pixel
size on the image is 1 mm and hence discernible. Therefore, Images in the range of scales of 1:100,000
to 1:50,000 can be interpreted visually without being disturbed by the pixel, which plays the same role
as that of the grain In a photograph. Obviously, visual interpretation on monitors does not present
these Inconveniences since image resolution can be enhanced or diminished as desired by using the
zoom.

5.1.2 The brain

Interpretation of the signal perceived by the eye is done by the brain, where ail forms of processing
take place. Five types of processing by the brain can be identified and grouped as:
— pre-processing of the signal (1),
— textural processing (2 to 4),
— structural processing (5).
1. The first form comprises perception of colours, contrasts and shapes, in particular linear fields
and angles between them.
2. The second is identification, carried out by recollection of objects or shapes seen previously.
3. The third is classification of objects Into types or reference groups.
96 Processing of Remote Sensing Data

4» The fourth form of processing corresponds to spatial analysis of the distribution of objects in two-
dimensional space. For example, a type of land cover, viz., hill slopes with vineyards and orchards
in the Yonne district, is observed in a single part of the Image (in the centre) and absent elsewhere
(Fig. 5.1). On the other hand, another type, hardwood forest bordering a talweg, is seen throughout
the image (Fig. 5.2). Thus spatial analysis can be carried out after textural processing of images
(CD 5.1).
5. The fifth form is analysis of organisation of objects in three-dimensional space (if stereoscopy is
feasible) or at least in two-dimensional space (Images on monitor). This involves defining textural
groups of objects whose spatial organisation is specific and visually recognisable from more or
less defined patterns, often empirically or through chorological laws, involving studies of
relationships between characteristics of semantic units (intrinsic factors) observed and their
distribution in three-dimensional landscape (extrinsic factors).

Fig. 5.1: Distribution of hill slopes with vineyards and orchards in the Yonne district, interpreted from SPOT
images (after Bertrand, 1994 and Mollet, 1994).

Hardwood forest on talweg border

(Agro-landscape 28:67,680 ha)

Fig. 5.2: Distribution of hardwood forest bordering a talweg in the Yonne district, interpreted from SPOT images
(after Bertrand, 1994 and Mollet, 1994).
 Processing and Interpretation 97

5.1.3 Interpretation procedure

An aerial photograph or an image represents aggregate data received at a given time in a given field
of view and with a resolution defined by the characteristics of the source-receiver pair. It is raw data for
which an interpretation has to be made. Aerial photographs or images thus differ from a map for which
a legend corresponding to a given theme exists. A map is the result of classification of the objects
depicted on it; it represents spatial distribution and organisation of these objects as well as their
conceptualisation. An aerial photograph or an image is the result of a series of measurements carried
out for each element of resolution in a given field of view; it does not define objects. Hence it is
necessary to recognise in the aerial photograph or image the objects present, to classify them and to
study their distribution and organisation before comparing the information from a map with that of
remote sensing data.
In a photograph or an Image, the eye detects the most contrasting features. These contrasts may
be represented by lines which, in most cases, are open. Hence they do not delineate entities. It is
therefore not simple to define features generated and determined by what is enclosed inside an envelope.
For example, just by applying isodensity analysis on an aerial photograph or an image it can be seen
that objects earlier defined are not correctly delineated.
Two major types of visual interpretation are commonly used. The analytical method is described
below and the method of synthesis or landscape is given In Chapter 18.

5.2 PHOTO-INTERPRETATION
Interpretation of aerial photographs has been done for several decades now for many thematic subjects
concerned with the investigation of natural sciences, such as geology, geography, agronomy, soil
science, botany and urban studies (Agache, 1970; Chevallier, 1971; CRU, 1969; Girard and Girard,
1970, Girard et al., 1996; Guy, 1969; Lillesand and Kiefer, 1994; Mulders, 1987; Smith, 1968, and
others). Evidently, interpretation varies with the object viewed, which determines the semantic and
graphic precision to be achieved, the theme concerned and the methods of operation. However, in
almost all cases we are interested in analysing aerial photographs for diverse thematic problems,
retaining only those that would be important at the end of study. In fact, like a satellite Image, an aerial
photograph is a general and non-specific document on which all types of thematic data are overlapping
and recorded at a given date and time.
The basic approach to interpretation of an aerial photograph is to proceed from the simplest to
the most complicated. Thus, the brain interprets the various features appearing on the image in the
order of legibility and stores them In memory. It is then easier to choose features of interest from a
thematic point of view.
Let us consider, for example, a 1:17,000 colour Infrared photograph of the Vitteaux region (Côte
d’Or) in Burgundy. While the method is general, the tables presented here comprise only outlines
pertinent to the example.

5.2.1 Lines and points

H Communication routes
Features most readily discernible in a photograph or an image are lines and straightline aspects such
as roads, boundaries of agricultural or forest zones, etc., which in most cases are produced by the
effect of human activity on the environment. It is often important to draw or extract these features by
filtering onto a layer of information and to establish a legend (Fig. 5.3).
98 Processing of Remote Sensing Data

Fig. 5.3: Interpretation of communication lines in the Vitteaux region. Approximate scale T:40,000.

These features can be analysed by studying:

— general directions of straightline segments using rose diagrams,
— density of lines per unit area,
— length of lines as a function of their shape (rectilinear, curvilinear, etc.),
— intersections of lines,
— open or closed drainage patterns.
A curved communication line in the north-western part of Fig. 5.3 can be very clearly distinguished.
It represents an old railway track.
Roads are usually indicated as long lines covering the entire image. They are numerous and
exhibit a radial disposition all around Vitteaux. Two roads, merging into one, are directed south-east,
encircling a river. It can be inferred from this that the river overflows from time to time between the two
roads and that caution is taken to travel roads in this zone without risk of floods. Another series of
roads run eastwards and join at the village Massingy. Some areas without roads are also evident,
which correspond to inclined zones (see geomorphology in a subsequent section). A road with a
hairpin turn Is also seen in this slopy region.

B Parcels
Parcels are analysed on the basis of criteria of size, shape, abundance and mode of closure.
 Processing and Interpretation 99

□ Size
The sizes of a parcel are expressed In metres. This Is possible when the scale in the case of a
photograph and resolution in the case of satellite image are known. The following notations are used:
L : the longest side,
/: the shortest side,
p : sides of equal dimensions,
r: curved sides.

□ Shapes
Six definite shapes are mentioned In Table 5.1 but other forms can also be described depending on
the region and mode of cultivation.

Table 5.1: Shapes of plots

Shape Formula Description

Strip-form Lxl(L>3l) Highly elongated rectangles

Rectangular Lxl{3l>L) Four perpendicular sides,
trapezium, parallelepiped
Draught-board L x Lor l x 1 Four equal sides, or squares
Polygonal np More than four sides of equal
dimensions

Curved Lx r One or two curved sides

Rounded rx r Complete circles, or more than

two curved sides

□ Modes of closure
The following types of closure can be identified:
— open plots such as open fields,
— plots enclosed by hedges, with bushes or trees
-— plots enclosed by dry stone walls (in particular in clearing zones),
— plots surrounded by trenches or embankments.

□ Abundance
Abundance refers to giving an estimation of the sum of areas of each type of plot, relative to a wider
group such as landscape, landscape unit or landscape element. A code of 0 to 1 or 1 to 5 (see Table
5.2 below) can be used.

□ Pattern of parcels
When studies are carried out over a large spatial field“' , the general pattern of the plot compartments
enables a better understanding of the relationships between various components of the landscape
and the environment, or between the environment and its exploitation by man. It is preferable to
Integrate the plots of land and forest with those of agricultural plots. Six types of plot compartments
can be defined:

■•Spatial field or field of view: Largest than the area analysed and comprising all the sites under study.
100 Processing of Remote Sensing Data

Circular
The shapes of curved or approximately circular features are due to constraints of the physical medium,
such as hills or depressions (common in karst zones), from which they get their configuration, or due
to technological constraints, such as circular plots prepared according to a sprinkler irrigation system.
This pattern, little favoured in Europe, is more common in North America or in semidesert regions
(Saudi Arabia, Australia, Southern Algeria, for example) where modern cultivation operations have
modelled the land use characteristics according to available technology.

Parallel
Striplike plots are often non-uniform along their longer side particularly in alluvial plains or on slopes.
Their general direction may be rectilinear or curved.

Radial
Plots that are most often rectangular or striplike converge by their longer side towards a centre (a
village, for example). This pattern combines the circular and parallel patterns. It is characteristic of
circular zones which are recognised from the boundaries of plots organised as circles or crescents
and extend over hundreds of metres and kilometres. Such patterns are quite common In regions of
ancient culture (Europe, Syria etc.). Several studies have interpreted such patterns (Soyer, 1970).
They can also be detected in satellite images (Girard, 1995).

Perpendicular
Some plot compartments clearly show two preferential alignments for the boundaries of plots, such as
newly cultivated zones or areas recently reclaimed from former wet or aquatic zones.

Disseminated
Various types of scattered plot compartments are observed in forests, lands and grass plots. They
indicate low human activity or a zone difficult to cultivate or old abandoned agricultural attempts.

Random
In some cases, neither a spatial order nor a structural pattern is observed in the distribution of plots.
Such is the situation in many zones. It may also arise when the field of study is not large enough for a
pattern extending over a wider zone to be detected.
Infrastructures and superstructures may be observed within such patterns. For example, in the
Loire valley near Longue, squared features are noticed, often indicated by communication tracks,
inside which parallel patterns comprising strips or rectangles or random structures can be delineated
(Girard and Girard, 1989).
It is also possible to detect roads cutting across old patterns of the area or, inversely, zones
adapted to Roman roads (Agache, 1970; Chevallier, 1997). In some cases, the pattern of the feature
is dictated by an environmental component. Thus, in narrow flat-bottomed valleys, parallel features
are frequently observed due to the insertion of roads and plots between slopes.
Three types of features can be Identified on the image shown in Fig. 5.4, which are briefly
summarised below:
1. Zones with no plot boundaries. Firstly, towns and villages are recognised by zones bounded by
polygons with numerous entries (A). Secondly, zones of large area with tightly undulating and
circular boundaries are evident. These are grasslands, fallowlands on slopes or in some cases,
forest blocks (B).
2. Groups of very small plots, most often striplike, densely packed one against the other (C). East of
Vitteaux, these features are larger and more rectangular (D).
 Processing and Interpretation 101

Fig. 5.4: Interpretation of plot compartments in the Vitteaux region.

3. Other plots are medium to large in size and more or less polygonal in shape (E). Grasslands of
rounded shape are also distinguishable near the river (F).

■ Habitat
Another layer of information can be established from remote sensing data by identifying every point-
type element of the image such as dwellings and constructions. For this theme, analysis of point
density as well as their spatial distribution often provides useful information. Thus, various farm and
residential constructions, sections of a town and diverse urban zones of an agglomeration can be
distinguished (see the satellite Images of towns on the CD). Areas of human impact on the environment
such as development zones, drainage or irrigation zones, etc. are also discernible.
The pattern of dwellings in the global analysis can be coded in two ways:
— Coding the habitat units with values of 0 to 3:
0. No habitat,
1. Scattered habitat,
2. Grouped habitat,
3. Mixed (scattered and grouped) habitat.
— Coding each mode of habitat by 0 or 1. The value 0 in each of these three groups, viz., scattered,
grouped and mixed, indicates absence of habitat.
It is also very useful to superpose this layer on the preceding one for analysing relationships
between habitation or urban zones and roads, as well as between open lands and habitation zones. A
new interpretation can thus be made by excluding topographic and land cover data. Not only are
102 Processing of Remote Sensing Data

conventional relationships invariably confirmed, but a new method of detecting new specific relationships
between the media under investigation is developed. One of the basic premises of visual interpretation
is to be prepared to be astonished and to discover a priori unknown relationships or to accept “new”
discovery of them.
It can be seen from Fig. 5.5 that habitation is strongly grouped In several suburbs and in Vitteaux
proper. It should be noted that the railway track does not pass through Vitteaux but some habitations
have developed between the city and its suburbs. Between Boussey and Saffres habitation is developed
along the road. Some isolated hamlets exist on the plateau west of Vesvres and in the east between
Saffres and Massingy.

N
Massingy
VITTEAUX

Saffres

Vesvres ' .i * .i? '

Boussey

•••••• Buildings Zone with drainage works

Fig. 5.5: Interpretation of construction zones (and drainage works) of the Vitteaux region.

Processing of all such information (scanning, for example) is facilitated by digitisation of

interpretation and its incorporation In a geographic information system.

■ Hydrology
All Information pertaining to water can be extracted from remote sensing images. In fact, most common
forms associated with water are either punctuate (ponds, lakes etc.) or linear (rivers). Moreover, if the
photograph is in the near infrared band, the contrast between the water body and all neighbouring
objects is very sharp since reflectance of free water in this wavelength band is zero (see Chapter 4).
The hydrology of the Vitteaux region (Fig. 5.6) is mainly represented by the river Brenne, a very
serpentine watercourse with numerous meanders. Several secondary watercourses with dendritic
stream systems exist, some of which have only temporary flows.
 Processing and Interpretation 103

In this case also, each preceding layer can be compared with the next. The existing hypothesis
may be confirmed or new chorological laws may be defined. Integration of two interpreted layers of
information for the habitat and hydrology constitutes a good reference for subsequent interpretations.
In particular, the number, shape, spatial distribution of watercourses (permanent or temporary), provide
valuable Information about geology and geomorphology (Girard and Girard, 1970).

■ Geomorphology
Geomorphology may be represented by lines or partly closed contours. Geomorphological maps indicate
various types of information and the conventional signs denote different morphological forms such as
talwegs, alluvial cones etc., as in the given example. Slopes are represented by most expressive and
most readily interpretable signs such as hatches drawn on the greatest slope and spaced at one-
fourth their length. Colour presentation of maps enhances their legibility (Fig. 5.7).

5.2.2 Closed areas

Another phase of interpretation of a photograph or an Image consists of delineating closed areas,
map units^, with which a ground feature Is associated. Hence, it becomes necessary to define the

^Map unit: a graphic unit with a closed contour having a local semantic content and assumed homogeneous at a
definite probability level.
104 Processing of Remote Sensing Data

Fig. 5.7: Interpretation of morphology of Vitteaux environs.

objects we wish to recognise, either a priori or while drawing the map units. This phase differs from the
preceding ones since closed areas are drawn and especially since thematic maps are prepared through
iteration between graphic and semantic data. Drawing boundaries leads to defining an object, i.e., we
go from the container to the content. On the other hand, the object permits defining such boundaries,
i.e., we go from the content to the container. The capacities of the interpreter gradually increase and
play an important role. The Interpreter ought to come out of the confines of his specific ability and be
capable of defining all semantic and spatial models that he uses.

■ Land cover
In the study of the natural medium, the first Information layer often Is the land cover map. In order to
process this theme correctly, it is necessary to start with relationships that tend to be distinct: for
example, that a land cover Is delineated by the boundaries of a particular plot. As these boundaries
would have been drawn previously, all such plots are Identified as having the same land cover, even if
a track or a road separates them (Fig. 5.8).
Depending on the importance of the theme and the experience of the interpreter, analysis of land
cover data can be carried out in detail. For example, several types of forest cover can be delineated
taking into consideration the tree species, their density and their height. In some cases, only one class
such as ‘forest vegetation’ can be used but contrarily bare soils can be classified into many groups by
their grey levels, hues or brightness. If it is desired to obtain a maximum number of classes, it is often
convenient to initially differentiate tree and scrub vegetation, then herbaceous vegetation followed by
analysis of soils, for example. Thus, various information layers are generated, which can be readily
 Processing and Interpretation 105

m i Forest and woods Agriculture

a Hedges and grasslands Vineyards and orchards

Fig. 5.8: Interpretation of land cover in Vitteaux environs.

connpared with the functions of GIS. In fact, as the graphic base of all these documents is the same— •
either a photograph or an image— superposing them one over the other poses no problems.

H Soil science and geology

Every subtheme can be compared with the preceding layers with the objective of better defining the
chorological laws that enable construction of a reconnaissance map.
Such a procedure was followed in the example of Vitteaux with regard to soil mapping (Fig. 5.9).
A geological map could also be prepared from the same data (Fig. 5.10).

5.2.3 Identification of thematic objects

Certain constituent elements of the natural medium, such as deep soil horizons, groundwater table,
geological layers, etc., are not directly detectable on images or aerial photos (Fig. 5.10). In such
cases, aerial photographs or images can be used differently from the manner described above. In this
approach, we look for features that are detectable on aerial photos or Images and which are directly
related to the theme being investigated. Thematic objects are thus delineated using boundaries of
image features visible in the remote-sensing data. This enables recognition of chorological laws that
106 Processing of Remote Sensing Data

Fig. 5.9: Preparation of a reconnaissance soil type map from interpretation of aerial photographs of the Vitteaux
region.
1. CALCISOLS, thick and stony; 2 . LITHOSOLS, limestone, on slopes; 3. Humic CALCISOLS, very thin; 4. Humic
RENDOSOLS, on marls; 5. RENDOSOLS on limestone, on benches; 6 . Thick CALCISOLS on marls; 7.
NEOLUVISOLS of loams on limestone; 8 . BRUNISOLS with coarse material of alluvial cones; 9. COLLUVIOSOLS
of secondary valleys; 10. Hydromorphic, loamy and carbonaceous FLUVIOSOLS.

prevailed in making the thematic map. This new information layer can subsequently be integrated with
others because the characteristics of geographic projection are the same since they are derived from
the same data.
Aerial photos and, to the extent possible, satellite photos are obviously analysed In stereoscopic
vision.

5.2.4 Vinicultural “terroirs”

This term has been variously defined by different specialists (Vaudour, 1997). It can be briefly defined
as a geographic space of limited size affected by interactions between the natural environment and
human activity. Remote sensing facilitates detection of these spatial features from pertinent indicators
related to agronomic potentialities. If the latter are integrated with the structures associated with human
activity, such as community appurtenance and skill, mode of rural life and typical and specific
productions,” terroirs” can be defined by means of remote sensing.
Interpretation of aerial photos is often utilised for studying terroirs and demands a very fine
geographic resolution and synoptic view.
 Processing and Interpretation 107

Fig. 5.10: Preparation of a geologic map of the Vitteaux region based on interpretation of aerial photographs; map
1:50,000.
Hard limestones of the Lower Bajocian; lg_5 : Marls of the Upper Liassic; 1 4 ^,: Limestone with giant Gryphaea of
the Middle Liassic; Domerian marls; I3 : Sinemurian limestone; L: Loams; Fz: Recent alluvium.

The terroir mapping of C6tes-du-Rh6ne area in the Nyons-Valreas region can be cited as an
example. This map of 21 soil-landscapes (Chapter 18) grouped Into 5 morphological units enabled
determination of 8 vinicultural terroirs (see CD 5.2). These terroirs were validated by data on grapes
and were subsequently used for classification from the SPOT image based on the maximum likelihood
method (see Chapter 9).

5 .2.5 Conclusion
Analysing every phase and every theme is the best method for extracting all the information contained
in an aerial photo or image. Comparisons are then made between various layers with a view to
reconstructing a spatial model relative to a given approach to the natural environment.
Each zone can hence be described by a group of analytical variables defined earlier (Table 5.2).
In the given example, a generalised description key is obtained for all map units. For each map unit,
the following modes are indicated.
— either absence (0) or presence (1);
108 Processing of Remote Sensing Data

Table 5.2: Format for description of analysis

Format for description of aerial photos

Communication Mode Land cover Mode

Road Agriculture
Track Vineyards and orchards
Railway track Grasslands
Shape of plots Woods, forests and isolated trees
Strip-form Conifers
Rectangular Geomorphology
Draught-board Loam
Polygonal Alluvium
Curved Flat-bottomed talweg
Rounded V"Shaped talweg
Plot compartments Alluvial cone
Circular Cornice
Parallel Slope
Radial Bench
Perpendicular Hydrology
Disseminated Spring
Arbitrary Temporary flow
Habitat River
Scattered Free water
Grouped
Mixed
Drained zones

— relative abundance, which to be unaffected by variations generally possible from different

interpreters and considering the duration required for interpretation, can be given by the following
values:
0— absence,
1— 0 to 5%,
2— 5 to 35%,
3— 35 to 65%,
4— 65 to 95%,
5— 95 to 100%.
This code has been tested for the last twenty years. It has undergone a slight change compared
to the one proposed in 1975 (Girard and Girard, 1975). It proved to be quite sufficient in defining
various landscape units by subsequent statistical analysis.
The GIS greatly facilitates comparisons. It Is also possible to establish interpretation models and,
using them, simultaneously analyse the results in spatial as well as semantic domains.
 Processing and Interpretation 109

5.3 VISUAL INTERPRETATION OF SATELLITE IMAGES

5.3.1 Interpretation of photographic prints
In order to extract maximum information from these images, it is possible to enlarge them in such a
way that the resolution of the image corresponds to the resolution of the eye. Thus, for a panchromatic
SPOT image, the 10-m resolution is represented by 0.1 mm, which then corresponds to an image of
scale 1:100,000. To facilitate drawing, a scale two to four times larger can be used, i.e., 1:25,000 for a
resolution of 10 m, or 1:50,000 for a resolution of 20 m.
One of the main advantages of satellite images is a synoptic view that enables comparison of
colours or grey levels from one end to the other of the scene, i.e., over thousands of square kilometres.
On the other hand, common aerial photographs provide such comparisons for only a few tens of
square kilometres. It is therefore necessary to preserve the entire spatial field. For a 10-m resolution
Image, such as that of SPOT in the scale 1:25,000, the entire field corresponds to a photograph of
2.4 X 2.4 m. This is not convenient to work with. Contrarily, Interpretation becomes very easy with
60 cm X 60 cm photographs, which correspond to a scale of 1:50,000 for a quarter of the SPOT
multispectral scene.

■ Points and lines

On a satellite image it is difficult to mark all the points or lines that often do not cover a pixel:
— either because their size is insufficient and they do not cover an entire pixel; in such a case, a
composite response, a mixel is obtained (Chapter 15);
— ^or because they lack sufficient contrast with adjoining objects.
Such a situation is observed in the case of constructions and a large number of roads. The
boundaries of plots appear only when the land cover feature has a strong contrast. Water bodies can
be delineated if they are sufficiently large (more than four times the size of the pixel). Streams can be
detected even if their width is of the order of the size of a pixel since most often a strong contrast exists
with neighbouring zones; however, the relationship between pixels of the same stream cannot always
be identified. Contrarily, they are readily confused with shadows and hence caution must be exercised.

H Designing a legend
Interpretation can be carried out object by object in the manner described above and validity of
interpretation rapidly verified. To prepare the legend for the entire field under investigation, interpretation
in different zones Is recommended so as not to forget the modalities in diverse variables. The legend
has to be defined as interpretation progresses for all the thematic aspects of the zone— ^the earlier the
better.
In order to ensure uniformity of processing for the entire visual field and good reproducibility
between interpreters, when there are several, or between procedures followed by the same interpreter
on different days, a description format Is prepared (Table 18.1), which should be completely filled In for
each map feature.

5.3.2 Interpretation on computer monitor

Interpretation can be carried out directly on monitors when an Image-processing software package is
available in the computer system. In this case, boundaries of various map units are directly drawn on
the monitor. Two aspects need to be harmonised, viz., high precision of the pixel for tracing the
110 Processing of Remote Sensing Data

boundaries and maintenance of a sufficiently large field of view to facilitate decision making. For this,
it is necessary to have a largest possible monitor and instantaneous access to forward (enlargement)
and backward (reduction) zoom. One has to verify whether this function is practicable and immediate
In the software and whether it can be used directly with the mouse employed to trace the boundaries.
The major advantage of computer Interpretation on a monitor for tracing boundaries lies in the
ability to process:
— several images of the same scene,
— multiple channels of the same image,
— several colour combinations of various channels,
— images produced from diverse treatments and classifications, and
— images representing topography generated from digital terrain models (CD 3.1).
In fact, since tracing the boundaries Is primarily dependent on contrasts, it would be easier when
using these various combinations to choose the pixels through which a boundary is to be drawn. But
then it is essential to understand the information included In the Image displayed. This is not so simple,
as will be shown later in the study of digital image processing.
Contrarily, regions corresponding to images acquired on different dates can be delineated and
the temporal variation of objects thus analysed. This is based on the assumption that geometric
corrections are applied (Chapter 13) so that the images can be superposed one over the other.
Radiometric distortions would still remain and have to be corrected to the extent possible.
For visual interpretation on the screen, there should be a facility for instantaneous display of the
output image and for repeatedly accessing it. This imposes constraints on the image processing software
used. It is important to have a series of interpretations, all of which have the same projection and can
be compared with one another. Any boundary of a map unit can then be readily modified when different
themes are compared and when it is perceived that differences in delineation of some pixels, between
different thematic features, raise the semantic significance above random variation.
All the thematic data thus interpreted are digitised in raster form. It is possible to directly process
them in geographic information systems. They can also be represented as vector data (several image
processing programs Incorporate this function).
When a combination of an image processing software and GIS software is employed. It is necessary
to georeference the images to be used. Then the boundaries can be traced directly In vector mode on
the GIS, while retaining the required images at the bottom of the screen. The importance of this
approach is that by the end of visual Interpretation, the geometric base of the GIS is established and
the topology of various map units traced from the images is automatically obtained.

5.3.3 Stereoscopic vision with satellite images

Stereoscopic vision can be acquired when two images covering the same ground objects at two
different angles are available (Chapter 14). Spot images provide stereoscopic vision (see Chapter 2
and the CD SPOT system). In fact, by its capability to acquire images at various angles, it is possible
to observe a region stereoscopically using images acquired at one or several days’ interval.
The larger the difference between view angles, the better the perception of topography. It is hence
recommended that images with a large positive angle for one and a large negative angle for the other
of the stereoscopic pair be obtained. It should be remembered, however, that in this case distortions
will be maximal and consequently interpretation on GIS becomes more difficult. It may therefore be
recommended that, whenever possible, one image be acquired with maximum angle and the other
with a smaller angle in another direction or with zero angle. In this case, the boundaries are traced on
the second Image since Image distortions would be minimal and topographic information more
accurately extracted.
 Processing and Interpretation 111

The stereoscopic vision of a satellite image differs significantly from that of an aerial photograph
since the relief is less accentuated and since the image is acquired through a more synoptic view.
Consequently, it is more difficult, for example, to differentiate various terraces in large valleys or alluvial
cones. Contrarily, level differences that extend over tens of kilometres can be readily traced and
determination of large zones becomes easier since large features are distinctly identifiable and readily
classified. Similarly, drainage networks are easy to interpret given their hierarchy and structure. Talwegs
can be readily identified by analysing winter Images in which they are observed below deciduous
forests and on bare soils.
In some cases, images required for stereoscopic vision have been acquired at an Interval of
several months. It should be remembered that overtime the solar angle changes and inevitably some
shadows will silhouette the apparent relief. Consequently slopes of strongly incised valleys are wholly
shadowed and little information about them ascertainable. Such a situation can lead to many errors in
interpretation In as much as shadows and water can have very close radiometric characteristics.

5.4 CONCLUSION
Visual Interpretation of satellite Images has been a standard procedure for many years. One might
have thought in the 1980s that it would displace expert interpretation of aerial photographs. As a
matter of fact no such displacement has occurred. On the contrary, interpretation of aerial photographs
has taken on greater importance for several reasons.
Given the lower resolution of satellite images, aerial photographs offer many advantages in
answering questions that require high resolution, such as quantum of vehicles, persons and trees,
accurate areal computation, land-use management, etc. Further, with the advent and ready availability
of GPS, demands for accuracy are on the rise. Present-day Earth observation satellites do not yet
acquire data accurate enough to meet these demands but that capability may not be far off.
On the other hand, since photographs can be digitised and processed digitally In the same manner
as images, a revival of Interest is seen. Further, the high resolution of aerial photos can be combined
with the power of computer processing. A model of a transformed landscape can be developed by
modifying one or several factors of the physical or human setting and display the new model in the
same scenario. This enables visualisation of changes that might be effected through a directorial
variation in territorial management.
Lastly, with the advent of GIS and digitised mapping data such as digital terrain models, which
give altitudes and all the derived variables, viz., slopes, exposure, crests and talwegs, etc., it has
become possible to Integrate visual interpretations with the data obtained from computer processing.
It is possible to process the information at multiple levels of resolution, although problems concerned
with change of scale (Chapter 15) have yet to be satisfactorily solved.
Visual interpretation is obviously simplified since it can be directly carried out on a computer
monitor:
— with integration of topography by digital processing if needed,
— ^with integration of aerial photos and satellite images,
— with the possibility of synthesising a multiband image providing resolution of aerial photographs.
Visual Interpretation today is the fastest method for Integrating structural Information. This will
perhaps change in the near future. However, to answer the present-day demand of professionals in a
time frame compatible with their requirements, visual interpretation is often one of the components to
be integrated into such an approach.
Introduction of the stringent analysis necessary for computerisation of data and their digital
processing has led to profiling visual Interpretation methods. Thus, formats for description of Images
and photographs have been designed and computerised, which have to be filled In systematically.
Consequently, visual interpretation procedures are freed from the word ‘approximately’ so often used
by experts. Laws of interpretation are now emerging from this need for absolute accuracy. It will hence
112 Processing of Remote Sensing Data

become possible to formulate a set of chorological laws that will be usable at least in a given geographic
region, and which will be re-usable for the same region at a subsequent date. It may even be estimated
that when this set of laws is no longer applicable in one region, it may be used in another. Delineation
of regions may be based on this fact. Similarly, it may be considered that there has been a temporal
change In a region when a set of chorological laws is no longer applicable.

References
Agache R. 1970. Détection aérienne de vestiges protohistoriques gallo-romains et médiévaux. Bull. Soc. de
Préhistoire Nord, 7: Musée d’Amiens.
Bertrand P. 1994. Élaboration d’une base de données localisées sur les agropaysages à partir d’images satellitaires.
Application à l’étude des organisations spatiales et à la segmentation du département de l’Yonne. Mémoire
de Mastère ‘Système d’informations localisées pour l’aménagement des territoires’. Institut national
agronomique, Paris-Grignon, 46 pp.
Chevallier R. 1971. La photographie aérienne. Armand Colin, 227 pp.
Chevallier R. 1997. Les voies romaines. Picard, Paris, 343 pp.
CRU. 1969. Photographie aérienne et urbanisme. Centre de recherche d’urbanisme, Paris.
Girard CM. 1995. Persistance de terroirs circulaires dans le Gâtinais occidental et relations avec les îlots boisés.
Photointerprétation. 4/95.
Girard M-C, Girard C-M. 1970. Cours de photo-interprétation. Polycopié, INA PG, Grignon, 208 pp.
Girard C-M, Girard M-C. 1975, Applications de la télédétection à l’étude de la biosphère. Masson, Paris, 186 pp.
Girard M-C, Girard C-M. 1989. Télédétection appliquée. Zones tempérées et intertropicales. Masson, Paris, 260
pp.
Girard M-C, Girard C-M, Bertrand P, Orth D, Gilliot J-M. 1986. Analyse de la structure des paysages ruraux par
télédétection. C.R. Acad. Agri. Fr., 82 (4): 11-25.
Guy M. 1969. La photo-interprétation. Encyclopedia Universalis, Paris.
Lillesand TM, Kiefer RW. 1994. Remote Sensing and Image Interpretation. John Wiley & Sons, Inc., NY, 3rd ed.,
750 pp.
Mollet S. 1994, Élaboration d’une base de données des agropaysages du département de l’Yonne— application à
l’étude des dynamiques financières agricoles. Mémoire Diplôme d’Agronomie approfondie. Institut national
agronomique, Paris-Grignon, 55 pp.
Mulders MA. 1987. Remote sensing in soil science. Elsevier, Amsterdam, 379 pp.
Smith JT Jr. 1968. Manual of color Aerial Photography. Amer. Soc. Photogrammetry.
Soyer J. 1970. La conservation de la forme circulaire dans le parcellaire français. Mémoire de photo-interprétation
de L’EPHE, vol. VI. SEVPEN, Paris.
Vaudour E. 1997, Analyse spatiale et caractérisation des terroirs du bassin viticole de Noyons-Val réas (AOC Côtes-
du-Rhône). Mémoire de DEA, Institut national agronomique, Paris-Grignon, 34 pp.
 6
Image Processing—General
Features

6.1 INTRODUCTION
The following chapters are devoted to description of various methods of image processing. In this
book, special attention is given to processing satellite images or digitised aerial photographs. However,
most of these methods are applicable to any digitised data such as photographs obtained from as
different domains as archaeology (site detection as well as pottery analysis), soil science
(micromorphology), medicine, artificial vision in robotics, industrial zone surveillance, etc.
After a presentation of generalities (Chapter 6), the following topics are covered in successive
chapters: preliminary Image processing of single or multichannel data (Chapter 7), unsupervised
classification (Chapter 8) and supervised classification (Chapter 9). This part is concluded by
methodology of image processing In remote sensing (Chapter 10); bibliographic references for these
five chapters are listed in the last one.
The next chapters include structural processing of images (Chapter 11) by the software OASIS,
digital filtering of images (Chapter 12) and, lastly, geometric transformation of images, which is necessary
for superposition of various sources of geographic information (Chapter 13).
Names of the methods may vary according to different authors since they are mostly derived from
older mathematical methods, revived for application to image processing. The description of these
methods is not the same in general works on image processing, but we have used the most common
terms in the various advanced software programs in remote sensing.
Image processing methods can be broadly grouped under six principal types: 1) measurement-
space-guided spatial clustering; 2) single-linkage region growing; 3) variant of the preceding which
takes into consideration not only the value of the pixel, but also its neighbourhood and is known as
hybrid-linkage region growing; 4) spatial clustering; 5) centroid-linkage region growing; and 6) split
and merge methods which use contour reduction.
Attempting to organise these different methods into rigid groups is of little use for applied remote
sensing. Greater Importance should be given instead to the possibilities and limitations of each type of
processing. However, it is always a combination of various methods that leads from the information
available at the beginning of the study to the best possible qualitative response to the problem posed
by the user.
As it is often difficult to find examples of application of the various methods, we have applied all of
them to a single image. Since colour figures cannot adequately express all the details of processing,
all the results have been given on a CD. The simple programs of the TeraVue software (La Boyère
publishers) on the CD facilitate visualisation of the numerous Images presented and their partial
modification. To derive maximum benefit from the text access to a PC with a CD-ROM drive is necessary.
It is hoped that a large number of readers interested in satellite Image processing have this facility.
References to the CD are indicated in parentheses in the text with a number following the letters CD.
114 Processing of Remote Sensing Data

An image of Earth observation satellites can be used in several ways. It can serve as a basic plan
that suffices to mark the objects constituting the references needed for the theme under study. Most
often the reference points are villages, roads, rivers, forests, seashores, agricultural lands, etc.These
markers are not visible on the image in the same manner. Their visibility depends on the quality of the
image, date of acquisition of the scene, etc. Even for such basic usage, it Is necessary to carry out a
preliminary visual interpretation or image processing to obtain a readily understandable document.
Use of satellite images through visual interpretation (see Chapter 5) always requires a minimum
amount of processing such as pre-processing of the signal (see Chapter 2) and processing abutting in
colour composites. Sometimes this stage of processing is more elaborate and visual interpretation is
carried out on colour composites priorly subjected to digital classification.
Use of satellite images almost invariably necessitates statistical processing of data by means of
information technology.

6.2 IMAGE PROCESSING METHODS

Before undertaking processing of images, a judicious choice of the possible methods to be employed
should be made. Thus two different types of methods can be identified:
— Point analysis is based on the laws of spectral behaviour of objects, for which one should have
at least an interpretation model of objects (see Fig. 7.4). A better solution would be to conduct field
measurements of reflectance and take into consideration the most possible transformations of the
signal by the atmosphere as well as due to satellite conditions of acquisition. This group of analyses is
known as the radiometric method.
— Spatial analysis is based on characterisation of objects by their geographic position and by
their relative spatial position. Identification of objects based on field studies and in association with
already acquired information (mainly from existing maps) constitutes the basis of processing. This
type of analysis is called the geographic method.
Generally both methods are used depending on the phase of processing, objectives and data
available in addition to the satellite images (see Fig. 10.1).
In both cases the results obtained ought to be verified by comparing them with data other than
that of remote sensing. It is hence necessary to identify the method of validation to be employed in
order to choose the type of processing that would be coherent.
Several approaches exist for image analysis, depending on whether one Is a mathematician,
computer scientist, statistician or thematic expert. These different approaches are complementary
and become enriched during analysis of satellite Images. However, since each has its own specific
language, it is necessary to precisely define the latter when working with different scientific disciplines,
all the more so because the same terms may have different significance in different disciplines. For
this reason the terms and methods essential in image processing are described below.

6.2.1 Texture
Texture of an image Is defined as a combination of textural elements. A textural element is a group of
resolution elements (whose area is determined by the characteristics of the sensor) which have the
same value of radiance (or of a function of radiance) and which are connected (spatial dimension).
A textural element in the case of a raw satellite Image is a group of connected pixels having the
same digital value. For a classified image, it may also represent a group of pixels belonging to a map
unit and hence ranked in the same class. Texture is the characterisation of the entirety of these units.
One of the expressions of texture can be represented as a diagram on a Cartesian system of co­
ordinates, with the perimeter of the map unit as abscissa and its area as ordinate (Fig. 11.4). Analysis
 Processing and Interpretation 115

of such a graph facilitates characterisation of shapes of various zones. Characterisation of shapes of

textural elements hence forms a part of texture.

6.2.2 Structure
Structure of an image is a combination of structural elements. A structural element is defined by
repeated relationships existing between the textural elements.
These relationships may be, for example, relative positions between the textural elements of the
same shape or of different shapes, distances between textural elements, contrasts, etc. Obviously,
characterisation of textural elements by means of chorological laws, or spatial laws such as convergence
of shape, forms an Integral part of structure.
From a didactic point of view, image processing in remote sensing can be classified depending
on whether it Is based on structural or textural data. However, most of the programs presently available
in image processing software pertain to textural analysis. Hence we start with the description of textural
processing methods in the next chapters, followed by structural analysis in Chapter 11, devoted to
VOISIN and OASIS.

6.3 CLASSIFICATION
6.3.1 Multiple languages
Image processing Involves several aspects. It is based on mathematical and statistical methods, often
quite old. However, their development could not have been possible without computers since they
demand very long and tedious computations. With the advent of remote sensing data, these methods
progressed as large amounts of data requiring use of these tools became available. Hence a processing
language has developed based on mathematical, statistical and computer languages.
Thematic experts obviously use image processing every day since It constitutes a statistical
extension of the visual interpretation they used to carry out on non-digital images and still carry out on
digital images. The presently existing image processing techniques are not capable of replacing
interpretations by the brain. The language of image processing is Impregnated with that of interpreters
of various thematic fields. This language may differ according to whether the theme of study is closer
to a scientific or literary approach, or to the theme of a geologist, a soil scientist, a botanist, an
agricultural scientist or a geographer, and so forth. One of the most patent examples is the definition
of ‘structure’ and ‘texture’, terms which according to different thematic experts or mathematicians
acquire opposite meanings (see Glossary).

6.3.2 Segmentation and classification

If the approach of mathematicians working on data processing is followed, segmentation has to be
differentiated from classification.
Segmentation is the action of dividing an image into classes without knowing what these classes
thematically represent. This most often corresponds to unsupervised classification in remote sensing
data processing.
Classification is the process of extrapolation to the entire image using the previously chosen test
zones or nuclei for which the thematic significance and relationship with an object (or a group of
116 Processing of Remote Sensing Data

objects) have been established. This generally corresponds to supervised classification in remote
sensing data processing.

6.3.3 Classification and grading

The terms ‘classification’ and ‘grading’ need to be distinctly differentiated.
Classification consists of determining the laws (most often statistical, arithmetical or logical In
remote sensing) which lead to the choice of classes or categories to be retained and to the mode of
grading. Classification may or may not be hierarchic, supervised or unsupervised.
Grading consists of distributing the pixels of an Image Into groups or categories pre-established
by classification. Similarities exist within a category or class and the characteristics are common to all
the pixels included in It.
However, in the common parlance of image processing in remote sensing (see the discussion of
various programs), the term classification is generally used without differentiating between the two.
To avoid any confusion about the so-called automatic classification, it should be remembered that
an external intervention in processing always exists. In a number of cases, a rather too fast use of a
processing method may give the impression that the thematic specialist is not involved. This is never
the case. If the thematic specialist is not involved and wishes a result to be displayed on the monitor,
it means that he/she accepted by default the options inserted by the author of the software in the
algorithm. Some examples of the various types of external intervention in processing (discussed later)
are:
— division of a dendrogram and the method used for it in a hierarchical ascendant classification,
— number of classes retained in a hierarchical ascendant classification or In the classification by
mobile centres,
— choice of number of iterations in the classification by mobile centres,
— choice of the probability threshold in the maximum-likelihood classification,
— choice of nuclei in classification such as maximum-likelihood or structural classification, viz.,
OASIS,
— choice of the window size in structural classification (OASIS or VOISIN methods).
‘Nucleus’is a group of pixels chosen a priori to characterise an object for which a classification is
to be designed. A thematic group is the population of pixels which as a result of classification and
hence a posteriori, is combined into a single group. At the end of classification, a thematic group
carries the same name as that of the nucleus used for classification.

6.3.4 Ascendant and descendant methods

Descendant methods are based on selection of geographic or radiometric criteria for dividing an
image into various parts. This division is then iterated for each of the preceding parts with new criteria
that differ from the previous. The procedure starts with the entire image and Its division into parts that
gradually become increasingly smaller and may ultimately reduce to the size of a pixel. The main
question in this regard would be: what is the basis for the criteria for division in every iteration and are
these criteria coherent with one another? It is necessary to verify the stability of the entire division
operation when the criteria chosen change slightly. One has also to look into the question of
heterogeneous pixels, viz., mixels, that encompass several objects (Fig. 15.3).
These methods quite often correspond to visual interpretation of a synthetic nature, which is
based on structural analysis. They are also used when the objectives of classification are well defined
and when only geographic distribution of objects for which digital values are clearly established Is
analysed. These methods are little developed for satellite Image processing.
 Processing and Interpretation 117

Ascendant methods consist of aggregating a group of pixels that exhibit pre-defined characteristics.
The latter may refer to their radiometric properties or relative geographic positions. These are generally
the same criteria as used in successive iterations that finally lead to grouping all the pixels into a single
category. Ascendant methods hence sfarf with a pixel and arrive at the entire /mage through successive
additions. The question to be considered here would be: on what basis are the criteria of grouping
(distance, probability, etc.) defined for a given iteration (metric) and between different iterations
(ultrametric). The validity of the resultant groupings needs to be verified through assessment of the
quality of classification by means of the distance to a given reference point or probability of appurtenance
to a given category (see below).
These methods are most commonly employed in textural processing of satellite images.

6«3.5 Concept of mathematical distance

The statistical methods of classification used in image processing are mainly based on the concepts
of probability and mathematical distance.
The concept of probability is readily understood. The greater the probability of a pixel joining a
group, the better its grading If it is included In this group. This approach is used In maximum-likelihood
classification.
The concept of mathematical distance Is less common but equally easy to interpret. The greater
the distance between a pixel and a group, the less likely its classification in this group.
Mathematical distance is computed in a multivariable space, i.e., multiband and/or multidate.
Several procedures exist for computing mathematical distance between two objects. An example Is
useful In understanding computation of mathematical distance.
On the image of Brienne (CD 6.1) three points were identified (Table 6.1), one characteristic of
dark bare soils (S), the second a mixture (a mixel, M) of water and soil and the third characteristic of
chlorophyllien vegetation (V).

Table 6.1 : Mathematical distances of three pixels

Object Row Column b1 b2 b3 Distance Euclidean Manhattan

between objects distance distance
Soil (S) 796 847 75 56 66 S andV 43 61

Water and 779 838 91 84 86 S and M 40 64

soil (M)
Vegetation 795 839 66 46 106 M andV 50 83
(V)

The mathematical distance D between pixels S and V is designated as Dgy. For each channel (b^),
a difference between the digital values of each pixel S and V is computed: [S ^ /-
The distance is equal to the sum of these differences and Is given by:

Dsv = - V^,] + [S^2 - V ^ ] + [S,3 - V ^]

Different types of distances are defined. The %^-distance Is related to frequencies and Mahalanobis
distance is related to probabilities. Two most commonly used distances are the Euclidean and
Manhattan.
Since In principle the distance V to S is equal to the distance S to V, each difference
should be equal to [V ^ /- S^j,].Thus, the following two distances are defined:
118 Processing of Remote Sensing Data

— Euclidean distance is obtained as the square root of the sum of squares of all the differences:

— Manhattan distance (or L1 or block-city distance) takes the absolute value of each difference:

^S V - ^^b^ ^ ^ b 2 '~ ^ b 2 ^ -^IS.3-V,3'

The computed distances for the three points mentioned above are given in Table 6.1 .The Manhattan
distance is greater between soil and water-soil mixel (64) than between soil and vegetation (61). An
inverse situation Is observed in the case of the Euclidean distance: the latter is smaller between soil
and water-soil mixel (40) than between soil and vegetation (43). This is due to the fact that the differences
channel by channel between the soil and the mixel are of the same order of magnitude (16, 28 and
20); consequently squares of these differences are not very high: 256 to 784. Contrarily, between soil
and vegetation, the channel-to-channel differences are high (11,10 and 40); hence squares of these
differences are much higher— from 100 to 1600. Thus, the corresponding distance is greater.
Euclidean distance is much more sensitive to a large difference for a channel compared to the
Manhattan distance. Consequently Euclidean distance Is greater between two pixels that have a single
large difference for a channel than in the case of two pixels that have smaller differences over several
channels.
The choice of distance to be used, termed mefr/c distance, is very wide. However, this choice Is
much reduced in the satellite image processing programs available and so unfortunately the user
rarely has much choice. Software designers use a particular metric distance. It Is better to know which
one. Since at present satellite data are mostly coded In 8 bits. It is considered that quantitative information
is processed and Euclidean distance used as default. Some program propose several distances but
then it Is necessary to search in specialised books for the incidence use of a given distance for
classification.

6«3«6 U tilisation of distances for classification

Distances provide a simple way of expressing the postulate on which the classification is based. A
classification Is a grouping of pixels of a region in such a way that they are at the closest possible
distance from one another within a group and at the farthest possible distance from one another in
different groups.
Distances constitute the basis of classification. Intragroup distance Is defined as the distance
characterising not two pixels, but a group of pixels. Intergroup distance characterises the separation
between two groups. The rules of utilisation of these two distances for classification and their computation
correspond to ultrametric crWena.

M Intragroup distance
Intragroup distance can be determined by computing the weighted average of the distances of all the
pixels belonging to a group channel by channel. If there are G pixels in a group, there exist G (G-1 )/2
distances to be computed. It is also possible to choose the Intragroup distance as the longest distance
between pixels of the same group.
The longer the intragroup distance, the more scattered is the group, the smaller the intragroup
distance the more compact is the group.
A mean value for the group for each channel can also be defined, which determines the spectral
characteristic of a ‘mean pixel’. However, it should be remembered that this Is an approximation, since
all values in image processing ought to be integers in order to be displayed on the monitor, which is
 Processing and Interpretation 119

rarely the case at the end of computation of a mean or median. Then the standard deviation or the
median deviation is computed. However, we may also compute the sum of distances of the pixels of
the group relative to the central value (median, mode or mean).

B Intergroup distance
Intergroup distance expresses the differentiation between groups retained In a phase or at the end of
classification.
It can be determined by computing the mean of all the distances existing between the pixels of
the two groups under consideration, taking 2 x 2 pixels at a time. Between two groups containing G
and /^pixels respectively, there would be (G x K) distances to be computed. Another central tendency
such as the median can also be used. The minimum distance between G and K\s also important since
It poses the greatest risk that the two groups are not different but a singular.
A large intergroup distance indicates that the two groups are totally different or well separated.
Contrarily, a small intergroup distance indicates that the two groups are not separate and that they
may be ultimately combined at a higher hierarchical level.
Another solution consists of defining a statistical population for each of the two groups G and K
and investigating by conventional statistical tests whether the two populations (assumed Gaussian)
are similar or not and with what probability. For this, generalised variance analysis can be done. One
can also compute the distance between two fictitious individuals (mean, mode or median) representing
the two groups.

B Ultrametric
The ultrametric parameter corresponds to the choice made for comparing a pixel with a group and a
group with another group. This Is essential for classification. In fact, when two pixels are compared
with each other, it is considered that they have a zero intrapixel distance. But this is no longer true If
groups are involved. Thus replacing a group of G pixels with a fictitious pixel (mean, median, etc.)
introduces distortions upon comparison with another pixel if its Intragroup distance is not taken into
consideration during decision making. The same would hold true in a comparison of two entire groups.
Decision rules for determining the group to which a pixel Is to be attached are relatively simple.
First, It Is necessary to determine the distance between the pixel and the group.
The distance between a pixel and a group Is determined in most cases by computing the distance
from a pixel to the fictitious pixel representing the group. In this case we do not consider the fact that
a single pixel (for which the Intradistance by definition is zero) is compared with a group of pixels
(whose Intragroup distance is rarely zero). The same procedure is followed for all groups.
The decision rule for attaching a pixel to a group is as follows. The group for which the pixel-to-
group distance is the smallest is identified and thereby the group to which the pixel is to be attached Is
determined. This is the method of minimum sorting distance: DIMITRI (Girard, 1983).
Other more accurate methods exist for determining the distance of a pixel to a group G;they are
based on intragroup distances (Van Den Driessche, 1965). For this, the distance of the pixel to each of
the pixels in a group is computed. The mean of these distances Is used as the criterion for defining the
position of the pixel relative to the group. The same procedure is followed for all the groups and finally
the pixel is allotted to the group for which the distance Is the shortest. These methods are rarely used
in remote sensing.
Several possibilities exist for defining the decision rules for combining two groups.
Intergroup distances can be considered as the distances between pixels and the groups showing
minimum distances can be combined. If the Intragroup distances, far from negligible at the end of
classification, are not taken into consideration, the errors would obviously be large.
Groups can be combined on the basis of mean intergroup distances. This is the common practice.
120 Processing of Remote Sensing Data

It is possible to take the smallest of the distances existing between two pixels of two different
groups. The pixels aggregate one with another. Consequently, If several distinct groups are formed, It
Is certain that they are well separated since the ultrametric parameter does not favour separation.
It is possible to take the longest of the distances existing between two pixels of two different
groups. This expresses the condition to establish connections only when the pixels of the two groups
are very close to one another; thus, formation of numerous groups is favoured. Since the ultrametric
criterion favours separation In this case. It is necessary to verify whether they are truly different at the
thematic level.

■ Quality of results
Upon completion of classification, the quality of results must be verified. When classification is based
on mathematical distances. It follows that each pixel is placed in a group. As the distance of each pixel
to the group can be determined, a display showing these distances on the monitor provides Information
on the quality of the results (CD 11.4). The smaller the distance of the pixel to the group, the better it
is classified.
Depending on the assumption made at the beginning of classification, if the intragroup distance is
smaller, the group is more likely to be made up of pixels having close spectral characteristics (Fig. 6.1,
groups A and D). A long intragroup distance (Fig. 6.1, group C) indicates that the group may be
subdivided.
The greater the Intergroup distances, the better separated they are (Fig. 6.1, A and B, A and C, B
and C). A smaller intergroup distance shows that the two groups can be considered as a single group
(Fig. 6.1, B and D).

Intragroup distance

^ -------------^
Intergroup distance

Fig. 6.1: Relative positions of groups as a function of their intragroup and intergroup distances.

6.4 CONCLUSION
In commercial softwares, classifications are often imposed with little choice. This is not a serious
limitation If their merits and demerits are known. Some program, much more expensive, enable choice
of a specific method of classification. Clear understanding of the fundamentals of various classification
techniques is necessary for using them effectively, however.
 7
Preliminary Processing
Before undertaking any sophisticated processing of images, it is necessary to fully understand the
data contained in each band. For this, the statistical composition of the Image has to be analysed to
generate contrast stretched and colour Images.
In the case of multispectral data, after processing each band as mentioned previously, various
arithmetic treatments are carried out and the data of different bands are combined. Principal component
analysis and masking are often useful before commencing classification.

7.1 PROCESSING OF SINGLE-BAND DATA

The following treatments may be carried out on digital data of a single band, which can be:
— a panchromatic or a multispectral band of a satellite image,
— the result of a processing such as an axis of PCA,
— the result of a preceding classification, or
— a combination of multispectral data, such as a vegetation index, brightness, etc.

7.1.1 Histogram analysis

At present, satellite image data are coded in 8 bits, i.e., in 256 levels. Various data processing programs
can be used to obtain histograms of the digital numbers of a band. Histogram analysis is important in
several respects and should be carried out before any processing.
Firstly, when an image is displayed on the monitor, it is possible that nothing can be inferred, and
in such cases the histogram aids in identifying the information contained In the Image. If no information
exists, all the values are 0 (black) or 255 (white) or there may be no value at all. On the other hand, it
is possible that all the pixels on an image are dark and the eye cannot distinguish them. For example,
such a situation occurs in the case of the result of a classification into a small number of groups, say
10 to 30, which are displayed with intensity values between 10 and 30. Moreover, hues (tints) separated
by a level difference of 1/256 cannot be visually differentiated, especially in the range of low values
(Chapter 3). It may hence be thought that the image comprises no different classes, since the eye
cannot discern them, when de facto it does.
An histogram is primarily characterised by its shape. Rarely Is the distribution of the digital numbers
of an image Gaussian. This can be verified by standard statistical tests but these may not be directly
available in a single program. It is hence necessary to verify that the available image processing
software allows storage of all the histogram-related information in a file whose format Is compatible
with the formats needed for conventional statistical programs.
In the case of satellite images, Gaussian type histograms may be observed occasionally. However,
most often the distributions (Fig. 7.1) are bimodal or trimodal (b1 band) or of two separate populations
122 Processing of Remote Sensing Data

Statistical data of histograms:

Band Minimum value Maximum value Mean Standard deviation

b3 (Infrared) 9 199 91.68 24.95

b1 (Green) 42 254 83.87 24.53

Fig. 7.1: Histograms of bands b3 (infrared) and b1 (green) of the image of Brienne region (also see Table 7.1).

(b3 band). In most cases, histograms of unprocessed (raw) data are highly skewed to the right, with
very low frequencies of large digital numbers.
The exact significance of the histograms has to be studied. Some programs give the values of
mean, median, standard deviation, etc. along with the histograms (Fig. 7.1). These values should be
Interpreted with every precaution, especially when the histograms are not Gaussian. In any case,
these values, which correspond to estimates, are not very important since exhaustive information is
available in a satellite image.
Radiometric analysis provides a definite method for interpretation of the digital numbers of an
histogram.

■ Radiometric method
Models can be used to convert the digital numbers into reflectance values taking into consideration
the atmospheric transformations of the signal (Deschamps et al., 1984) (see Chapter 1). Models such
as LOWTRAN (LOW Resolution TRANsmittivity program) are included In some softwares such as
TeraVue. Using these models, the digital numbers can be evaluated on the basis of information about
the spectral characteristics of the objects, viz., soil, vegetation, water, snow, roads, etc. For this, among
others, co-ordinates of the satellite scene, date and time of acquisition need to be known.
It Is also possible to calibrate the digital numbers using the reflectance values measured in the
field. Presumably field measurements are obtained on the same day as acquisition of the satellite
Image or at least on a date very close to the latter so that the characteristics of the objects under
investigation do not change.

■ Geographic method
In some cases, locations of an object clearly known on the ground can be readily identified on the
Image. By ascribing a given colour to all occurrences of this object, the spectrum of digital numbers
 Processing and Interpretation 123

characterising a geographically defined object can be established (CD 7.1). For example, waterbodies
are identified in the southern part of the infrared colour image (band b3) of Brienne region. The digital
numbers for water vary from 10 to 19 and include the values 22, 25 and 33. Thus an object whose
geographic position is known can be characterised in terms of radiometric values but not by reflectance.
It is then possible to thematically interpret the digital numbers of the histogram of each band. This
process is essential for optimal utilisation of the various software tools of the image processing technique.
Visualisation of an image on the monitor passes from the raw image through two phases, viz., a
function of transformation of raw digital numbers and Look Up Table (LUT).The pixel values are not
changed; rather their visualisation on the monitor is modified depending on the function applied.

7.1.2 Transformations of digital numbers

■ Dynamic range enhancement: thresholds
One of the primary methods of transformation of an image is enhancement of the dynamic range. The
raw data of an image hardly ever covers the entire range of 256 values available for display. It is hence
recommended that this interval be utilised to the maximum extent possible.
One method of image enhancement involves stretching the dynamic range of the digital numbers
by giving 0 to the minimum value and 255 to the maximum value, i.e., by using the real maximum
interval. Thus we obtain an image that exploits the entire capacity of the monitor. This procedure is
very simple and conserves all the information of the image; however, dynamic range of the Image on
the monitor is not always improved (CD 7.2).
The second method is based on determining the limits of statistical population of the digital numbers
by leaving on the left (minimum values) and on the right (maximum values) 1 per 10,000 values: 1-per-
10,000 Integral (Table 7.1 and CD 7.2). A reduced interval of digital numbers is thus obtained, which
enhances the dynamic range of the image on the monitor. Consequently pixels with digital numbers
less than 1 per 10,000 (45 for b1) or more than 9999 per 10,000 (214 for b1) are perceived as black or
white respectively. If this transformation Is used for subsequent processing, these pixels will no longer
be differentiated. It is therefore necessary to ensure that important objects to be classified are not in
these Intervals of digital numbers. As in the preceding case, the minimum and maximum values are
changed to 0 and 255. This procedure Is readily applicable when a function that gives the integral of
pixels below a chosen threshold is available. One can also choose the 1 per 1000 integral instead of
1 per 10,000; however, most often the objects corresponding to the edges of the histogram, indicated
as black or white, become visible on the monitor and may ultimately disturb the processing. Lastly, it
may be difficult to differentiate between water and shadows, both of which appear as black since their
digital numbers are small.
The third method consists of fixing the thresholds from the first or the last value of a digital
number for which the pixel frequency is greater than or equal to 1 per 10,000 (1 per 10,000 class)

Table 7.1: Various threshold values (in digital numbers) for dynamic range enhancement of an image.
Image of Brienne

Band Real maximum 1 per 10,000 1 per 1000 1 per 10,000 Mean Standard
interval integral integral class deviation

b1 (green) 42 to 254 45 to 214 49 to 184 45 to 182 83.87 24.53

b2 (red) 22 to 254 23 to 203 25 to 176 23 to 175 68.88 31.10

b3 (infrared) 9 to 199 11 to 179 11 to 168 10 to 172 91.68 24.95

124 Processing of Remote Sensing Data

(Table 7.1, CD 7.3). This is the simplest solution and was used for processing the Image of Brienne.
Precautions to be taken are the same as In the previous case.
Once the limits are chosen, it is possible to apply several transformation functions to the digital
numbers.

H Transformation functions
The objective of a transformation function is to obtain a majority of pixels in the grey levels that can be
readily detected by the eye, for example between the digital numbers 50 and 200. It is hence possible
to choose a transformation function such that most pixels of the band under study lie between these
values.
The simplest transformation function Is the linear function. After fixing the two limits of threshold
(Table 7.1), the raw signal Is transformed linearly to present the image on the monitor (Fig. 7.2a and
CD 7.3).

Fig. 7.2: Linear transformation of band b3 (image of Brienne). a. 1-per-10,000 class; b. suppression of values
corresponding to water and cultivated plots.

This method has the advantage of maintaining the relative positions of objects in the range of
digital numbers. The dynamic range of the Image is enhanced but not necessarily optimally. It can be
used to make only a portion of the image visible. This portion corresponds to the Interval of digital
numbers chosen for defining the linear range of transformation (Fig. 7.2b).
The square roof of the transformation function gives a satisfactory contrast if the digital numbers
of the band are relatively small. The minimum values, if not too close to zero, are thinned out and the
values close to 255 are saturated to very bright values. Consequently the image often has no contrast
(CD 7.4) and Its dynamic range is not optimal.
Logarithmic transformation gives an effect similar to the preceding one, and further reduces the
dynamic range. The image Is generally bright (CD 7.4).
The exponential function, when applied to an Image without defining the limits of dynamic range,
gives a very dark image for the more common medium and low values. On the other hand, high and
very high values appear very bright (CD 7.4). This function can be employed when the receiver gain Is
high and saturation occurs at high values.
The last three transformation functions are mainly useful for immediate application on the original
bands. The bright values and the dark zones can be directly marked in the Image.
The equal-population transformation, or continuous anamorphosis, optimises the dynamic range
of the band (CD 7.4). The Image is divided in such a way that each class of grey levels on the monitor
comprises an equal number of pixels. This function is preferred when the image is processed by the
geographic method since it provides a better perception of objects in the image. Contrarlly,
 Processing and Interpretation 125

transformations related to radiometric values vary for each wavelength band and consequently
application of radiometric models is rarely possible.
These methods of transformation are of the continuous type. Consistency requires their application
to raw data obtained from a continuous signal (even if it is digitised in the satellite during data acquisition).
On the other hand, when indices that result from classification are analysed, the data are interpreted
as discrete classes and not as a continuous parameter. In this case, other transformation functions
are employed.
The transformation function enables us to allot any grey level to any class, point by point or class
by class on the monitor. In this way, the value of a digital number can be modified in its representation
on the monitor. Hence, the result of a classification can be transformed into a band comprising as
many grey levels as the classes, assigning the desired grey level to each class.
The po/ygona/function is useful in constructing for any interval defined by the user a straight line
of the desired slope, viz., small or large, positive or negative. It is thus possible to represent the digital
numbers of low value in white and those of high value In black; a negative slope is thereby obtained.
As the interpreter cannot discern the variations in the dark and bright zones with the same degree of
acuity, some objects can be more readily identified on a negative slope than on a positive one. This is
often the case with long linear features such as rivers, roads, railway tracks, etc.
The polygonal function also facilitates putting a greater or smaller range Into a class, allotting an
entire interval to the same grey level, etc. All variations within a limited spectral range can be evaluated.
For example, all 256 grey levels can be used solely for pixels corresponding to water bodies (CD 7.4).
Thus, this tool Is useful for carrying out:
— Binary segmentation of image, which consists of placing all the pixels below a given grey-level
threshold in black and all the pixels above It in white (this constitutes a kind of masking).
— Multithresholding, which consists of creating various classes assigned to different grey levels
(Fig. 7.3). The thresholds are determined from modelling of the spectral characteristics of various

Objects Water Water side Crops, grasslands, White soils

and dark soils forests, soils and green plots

Digital numbers Oto 20 20 to 60 60 to 136 136 to 199

Image intensity 0 to 150 0 150 255 to 0

Transformation Positive Binary Binary Negative

functions threshold threshold threshold threshold

Fig. 7.3: Multithresholding of band b3 of the image of Brienne.

126 Processing of Remote Sensing Data

features or from field measurements. The representative classes In such a case correspond to known
objects. This procedure belongs to the radiometric method.
A special transformation function corresponds to radiometric masking (see below). If the spectral
characteristics of the image feature are used, it is possible to prepare a band In terms of final
interpretation of a desired theme. For example, if natural vegetation Is studied, the following aspects
can be taken into consideration:
— Free water bodies, which correspond to the lowest values of digital numbers in the infrared
band; these are assigned 0 value;
— Cultivated zones with intense chlorophyll activity with very high values of digital numbers in the
infrared band; these are assigned the value of 255.
All the other values of digital numbers belonging to themes assigned between 0 and 255 are then
spread out (Fig. 7.2b). Water bodies appear as black and green vegetation as white; all other features
are distributed between black and white over the 255 values (CD 7.5).

■ Look Up Table (LUT)

Once a transformation function is defined, a colour is assigned to each of the 256 digital numbers and
a LUT Is prepared. Several tables are currently in use (CD 7.6). One table codes the digital numbers
from black for 0 to white for 256. There are also tables that proceed from black (0 value) to white (value
128) and return to black (value 256). A monochromatic table can also be used which allots the same
colour to the entire image, from dark for 0 to bright for 256. Monochromatic tables can likewise be
prepared with less than 256 classes; this facilitates visual interpretation since the eye cannot detect so
many classes. Tables of 16 colour classes are also common.
Further, one can build one’s own LUT by assigning any colour to any class. For display during
processing phases, it is important to choose colours of high contrast and not dull ones. In fact, the role
of colours is to facilitate recognition of a given image feature, which requires its distinctive differentiation
from all other features.
Choice of colours Is particularly useful In making good hard copies of the results of processing.
The palette chosen for printing an Image is obviously a matter of taste, but it should be such that the
results to be evaluated are enhanced and not diminished by the colours selected. Moreover, legibility
of an image largely depends on the quality (make) of monitor used.
The possibility of assigning a given colour to a particular value of a digital number provides a kind
of sorting that represents a preliminary stage of data processing.
Thus, a colour can be given to a pixel on the grey level image and all the occurrences of the same
value, coloured the same way, can be obtained (CD 7.1).This aids In determining the number, geographic
distribution and location of all the pixels having the same value of digital number. The connection
between radiometric and geographic methods Is thus established.

■ Contrast analysis
Contrast is the ratio of the typological distance and the geographic distance between two sites. It thus
represents the ratio of their semantic distance to geographic distance.
In the case of satellite images, contrast is defined as the ratio of the mathematical distance
between the values of digital numbers of two pixels and their geographic distance.
Several types of mathematical distances can be used. The simplest. In the case of one and the
same band, is to take the difference between the absolute values of two digital numbers (Manhattan
distance). The geographic distance can be expressed in metres or in units of resolution for a given
satellite.
On a black and white Image, the expression of this contrast appears to the eye as the difference
in intensity between two connected pixels. However, if colours are assigned to various pixels, the
contrast is then expressed by the difference that the eye establishes between the colours of two
 Processing and Interpretation 127

connected pixels. But, this difference manifests differently for each observer. In colour, the effect of
brightness is greater than that of hue; thus, yellow appears brighter than red or violet. It is necessary
to give attention to the colours chosen for finalising a given processing since visual interpretation,
indispensable for the cartographic method, depends as much on the choice of colours as on the result
of processing. In all cases, it should facilitate interpretation of the image by the user.
The contrast on a computer monitor depends on the image per se, the transformation function
used for contrast stretching, the LUT, the colours assigned to various units, as well as on the contrast
regulation of the monitor. The contrast is Important since it aids in evaluating an image feature or zone.
The book Graphical Semiology (Bertin, 1974) may be consulted for preparation of a final document.
Lastly, it should be remembered that it is not possible to obtain exactly the same colours on a colour
print as on the monitor, since the print uses pigments that do not support all colour combinations
observed on a monitor (see Chapter 3).

■ Classification
Most programs of satellite-image processing provide classification using three bands and it is generally
not possible to process a single band. This Is inconvenient because one may desire to process a
panchromatic band, for example derived from a satellite image or a digitised aerial photograph. This
difficulty may be overcome by taking the same band three times and applying three-band classification.
Obviously, the image remains black and white. The neighbourhood method of image processing,
OASIS (see Chapter 11), can be used for single-band processing.
A possible method of processing panchromatic data is to initially carry out contrast enhancement
of the image. Segmentation can be done on the histogram, but this is often insufficient for making a
significant interpretation. Subsequently, for the image exhibiting different zones, if necessary, masking
can be applied (see below) and classification carried out on each unmasked part of the image using
the program OASIS. All the masked parts can then be grouped into a single image, the result of
classification displayed in colour.

7.2 PROCESSING OF MULTISPECTRAL DATA

Most often, three bands are used for Image processing since they can be displayed simultaneously on
a monitor operating on a three-fold 8-bit code. This provides an immediate interpretation of colours.
The three channels may be three multispectral bands of the same scene, three channels of the
same spectral band acquired on different dates, or one of the principal components (over an image of
multispectral bands) chosen for the same scene of three different dates or even three classifications
of the same scene, etc.
In most programs, it is possible to assign any one colour to one, two or three planes and display
the composite of the planes (see Chapter 3). As 256 grey levels are available for each principal colour,
i.e., blue, green and red, 256 x 256 x 256 or 16,777,216 different colours can be composed and
displayed on the monitor. In fact, the eye cannot perceive all of them.

7.2.1 Comparison between colours and spectral characteristics

of objects
The colours can be Interpreted when the significance of each of the basic colours for the objects
represented on the image is known. With regard to spectral behaviour of objects, the following points
should be recognised;
128 Processing of Remote Sensing Data

1. There is no direct relation between reflectance of an object in a spectral band and the value of
digital number corresponding to the band since it is the reflected energy that is measured by
satellites and not reflectance;
2 . The calibrations made between the energy received by satellite sensors and the 8-bit digitisation
of the signal (256 levels) are not the same for different bands (see gain of the SPOT system on
the CD);
3. The dynamic-range enhancements of different bands are not the same.
There are at least three models available for converting the spectral characteristics into colours.
More models can be used if colours are compared with field measured reflectance data since in this
case other factors such as atmospheric modifications, calibration of diverse field targets, drift of satellite
sensors, etc. need to be considered.
Consequently, the same value of digital number in different bands does not correspond to the
same value of radiance. As it is possible to obtain three values of digital numbers for each pixel, a
multiple line of ‘digital characteristic’ of a pixel, corresponding to a given object can be constructed.
However, we should not search for an absolute relation with the spectral characteristics curves obtained
from field measurements of reflectance for an object of the same nature. Depending on the situation,
the digital characteristic may be very dose to the spectral characteristic, but often it is quite different.
The relative positions of digital characteristics of objects for a given band are comparable to the
relative variations of reflectance for the same objects. Consequently, linear dynamic-range
enhancements for each band would facilitate such an analysis.
General models can also be applied but one cannot go too far in interpretation unless much
trouble is taken for inversion of the three transformation models.

7.2.2 Choice of band combinations

Various types of colour composites and in particular Infrared colour (IRC) Images can be prepared
based on the models mentioned above. For this, red Is assigned to the channel corresponding to the
near infrared band that varies between 750 and 1100 nm depending on the sensor, green to the
channel corresponding to the band between 600 and 700 nm and blue to the band between 500 and
600 nm. In such a colour composite, water appears dark blue, soils more or less bright blue or cyan,
vegetation more or less intense red and shadows very dark.
We can thus produce another colour composite similar to the colours that our eyes can normally
detect by assigning green to the near infrared band, red to the 600-700 nm band and blue to the 500-
600 nm band. Vegetation then appears green, soils brownish yellow and water violet-blue. This is a
pseudo-true colour composite.
Each spectral band can be made positive or negative and it may be given one of the three basic
colours. Then there would be 2 x 3 ways of colouring the first band, 2 x 2 ways of colouring the second
and finally 2 for colouring the last one. Thus, there are 48 ways of preparing colour composites from
the raw data (CD 7.7). If it Is possible to prepare colour composites with only two bands or to assign
different colours to the same band or to modify the transformation functions, the number of combinations
further increases. It is to be noted that all the images thus obtained comprise exactly the same
information but our eye perceives different organisations of the ground features on different composites.
This is due to colour contrasts which the eye cannot interpret identically.
When more than three spectral bands are available, the best possible combination can be studied.
The latter evidently depends on what one desires to detect.
In the case of LANDSAT TM, a large number of combinations, as well as detection of varied
themes are possible, with the seven bands ranging from visible to thermal infrared.
For example, the higher reflectance of water in the visible than in the infrared leads to preferential
usage of bands 1,2 and 3 when differences in depth or turbidity of water are to be determined.
 Processing and Interpretation 129

On the other hand, to study vegetation bands 2, 3 and 4 or bands 3 ,4 and 5 (or 7) could be used
for Infrared colour composites, in fact, reflective middle Infrared holds important information on water
content of vegetation covers, complementary to that provided by visible and near infrared bands. The
thermal infrared band can be used but it is beset with problems due to Its lower geometric resolution
compared to other bands (120 x 120 m instead of 30 x 30 m) and difficult to Interpret. In fact, for
France acquisition of scenes early morning makes interpretation of thermal phenomena difficult.
Depending on the visual effect sought, the three colours can be assigned to three bands variedly
(table 7.2). It is hence possible to identify certain themes or phenomena by using on the one hand, the
differences in spectral characteristics of objects in various spectral bands and, on the other, the
differences in sensitivity of the eye to colours and colour contrasts.
The visual effects and possibilities of differentiation between various themes for the codes given
in Table 7.2 are summarised in Table 7.3.

Table 7.2: Examples of colour composites prepared using LANDSATTM bands

Code Colours/ Bands Characteristics of major land cover themes

Blue Green Red Water Vegetation Bare soils Buildings

A 1 2 3 Blue-green — Chlorophyllian: More or less White

(brightness green bright beige
increases — Dry: maroon, more
with turbidity) or less greenish

B 2 3 4 Dark to light blue — Chlorophyllian: More or less Bright blue to

(brightness magenta— Dry: bright blue or white
increases with brownish green cyan
turbidity)

C 5 3 4 Green (brightness — Chlorophyllian: More or less Bright blue

increases with magenta— Dry: violet bright mauve
turbidity); non-
turbid: black

D 3 4 5 Black to sea blue — Chlorophyllian: Mauve Mauve

(for a low to high bright green to dark
turbidity) green according to
nature of plants—
Dry: pale pink

E 4 5 7 Black (turbidity — Chlorophyllian: Yellow to Orange to pink

differences bright to dark blue orange
cannot be according to nature of
distinguished) plants— Dry: green
to yellow

F 6 5 4 More or less dark — Chlorophyllian: Different Blue

blue depending various hues of shades of
on currents (no colours: pink, orange, green to
turbidity purple— Dry: almond- yellow
differences) green to yellowish-
green
130 Processing of Remote Sensing Data

Table 7.3: Detection of themes using different colour composites

Code Comments
Risks of confusion between water and vegetation, differences in water, differentiation between green
vegetation and bare soils but no difference within these themes. Confusion between some bare soils
and buildings.
Clear distinction between water and other themes, differences within the water, differentiation between
vegetation and bare soils as well as within each of these themes. Confusion between some bare soils
and buildings.
Clear distinction between water and other themes, differences within the water, risk of confusion
between some classes of vegetation and bare soils, differences within each of these themes. Confusion
between some bare soils and buildings.
Clear distinction between water and other themes, differences within the water, good separation
between some classes of vegetation and bare soils, differences within each of these themes. Confusion
between some bare soils and buildings.
Clear distinction between water and other themes, differences within the water, good separation
between vegetation and bare soils, differences within each of these themes. Confusion between
some bare soils and buildings.
Possible confusions between water and some elements of constructions, very little differences within
water, risk of confusion between dry vegetation and bare soils, differences within each of these themes.
Good separation between bare soils and constructions, detection of dip and orientation of slopes.

7.2.3 Two-dimensional or three-dimensional histograms

A plane defined by two perpendicular axes, such as the red (675 nm) and infrared bands for example,
on which all the pixels of a satellite image are located, is called a two-dimensional histogram. For each
point of the plane marked by the co-ordinates of the two bands, the frequency of pixels corresponding
to these two specific values is given. The frequency may be coded by a colour, for example for Increasing
frequencies ranging from black to blue, green, yellow, red and white, or by dots of size proportional to
the frequency.
It Is interesting to analyse various shapes of two-dimensional histograms. In fact. It is possible to
thus Identify the relationships existing between the bands. In most cases, strong correlation exists
between successive bands, which may not always be the case, in particular between visible bands
and the near-infrared band.
Correlation between bands can be observed when the points are more or less aligned in a two-
dimensional histogram. For two highly correlated bands two new channels can be created formed by
the sum and difference of the two correlated bands. In this case, differences can be clearly distinguished
and the two-dimensional histogram is much more stretched but this poses difficulty in interpreting
objects using digital numbers.This difficulty occurs In any arithmetic combination between two different
bands.
When the three conventional spectral bands, viz., green, red and near infrared of the Earth
Observation Satellites are used and If the zone under investigation comprises green vegetation, more
or less bare soils, water and shadows, often a very large deviation is observed in the ‘red/infrared’ two-
dimensional histogram. However, this evidently varies with the zones under study and does not constitute
a norm. If the area consists of only one object such as water, soil or vegetation, the results are different.
The image of Brienne was divided Into nine subimages and correlation between the three spectral
bands computed for each. The most important results (Table 7.4) show that:
— the strongest correlation is between b1 and b2;
— the lowest correlation is observed between b2 and b3 (the entire image and the ninth situated
in the south-east) or between b1 and b3 (for the ninth In the south);
 Processing and Interpretation 131

Table 7.4: Correlation between various spectral bands of the SPOT image of Brienne

Situation Correlation coefficient: Correlation coefficient: Correlation coefficient:

b1-b2 b1-b3 b2-b3

Entire image 0.981 0.147 0.079

South-east 0.988 0.198 0.147
South 0.963 0.089 0.225
East 0.975 ■ 0.00 •0.137

— there may also be no correlation between b3 and b1 (ninth in the east where no water or almost
no forest is present);
— sometimes, correlation may be negative (b2 and b3 for the ninth situated in the east).
The values given above correspond to the most common cases: they clearly Indicate the difference
between the near infrared and the visible bands and the strong correlation between the two visible
bands.
When a greater number of bands are analysed as in the Thematic Mapper (TM) (Table 7.5),
similar very strong correlation is observed between the three visible bands. Correlation between the
near infrared and the visible bands is low. Correlation between bands 5 and 7 of the mid-infrared is
high and these two bands also show strong correlation with visible bands but weak correlation with the
near infrared band TM4. Lastly, the TM6 band of thermal infrared exhibits a special behaviour. It
should be taken into consideration that the spatial resolution of this band Is not the same as that of
other bands.

Table 7.5: Coefficients of correlation between the TM bands for the image of Brienne

c1 c2 c3 c4 c5 c7

cl 1.000
c2 0.974 1.000
c3 0.957 0.981 1.000
c4 0.098 0.113 0.012 1.000
c5 0.877 0.900 0.916 0.176 1.000
c7 0.889 0.918 0.954 - 0.024 0.961 1.000
c6 0.340 0.381 0.453 - 0.031 0.549 0.543

■ Model of spectral behaviour for interpretation of digital numbers

It is interesting to note that if the red band data are plotted on the abscissa and the near infrared data
on the ordinate, the form of the histogram is always almost of the same type (Fig. 7.4).
The points on such a diagram are distributed Inside a triangle for which the three vertices are
represented by:
— water or shadow (W),
— highly reflecting soils: bright or very smooth (S),
— most chlorophyllian vegetation (V).
Going from W to S, various states of the soil surface are observed, constituting a cluster of
soils ranging from dark or rough to very bright-coloured or smooth. From S to V, soils partly covered
132 Processing of Remote Sensing Data

Fig. 7.4: Digital model of interpretation of soils, water and vegetation on a R/IR plane.

with increasing chlorophyllian vegetation are observed. From V to W, we find vegetation of

decreasing chlorophyll content, with decreasing coverage and with changing nature of vegetative
structure. For example, conifers are situated between W and V while hardwoods occur towards the
pole V.
The features are thus distributed in the same way as on a graph with reflectance values as co­
ordinates instead of digital numbers. Hence, objects can be detected from their relative positions in an
histogram. However, this does not always suffice. It is therefore necessary to display three two-
dimensional histograms, i.e., orthogonal projections of three planes of the cube representing three
dim ensions of the sem antic inform ation in a C artesian co-ordinate system. Some image
processing programs provide this facility. Others such as TeraVue are more advanced and enable
projection of three-dimensional composition on any arbitrary plane. Thus, the cube can be rotated and
the points can be projected on a Cartesian plane for which the equation of the axes is automatically
given. This offers the advantage of defining any arithmetic combination of two axes and determining
the position of points in three dimensions. This may be termed as three-dimensional analysis of
histograms.
 Processing and Interpretation 133

■ Study of spectral characteristics of image features

The next operation emerges from the analysis of histograms. It seeks to establish a relationship between
the semantic information (digital numbers) seen on the histogram and the graphic information (positions
of the corresponding pixels) observed on the image.
It is essential to ‘navigate’ through the image, i.e., to move onto any pixel and identify the digital
numbers of this pixel in the three dimensions. This enables estimation of the type of object the pixel
represents. For deeper exploitation of the histogram, it is necessary to determine the position of every
pixel in the histogram and from it, the significance of the histograms on which the pixels are represented
by a point of the same colour as that on the image. Frequencies in this case are indicated by circles of
proportionate diameter. With such a function, it becomes possible to identify the digital numbers of a
small zone on the Image. Pixels pertaining to the small zone studied are marked successively and the
values of each as well as its position in the general histogram of the Image are noted.
This investigation of radiometric spectra of image features together with their geographic distribution
aids in understanding the spectral characteristics of objects and in choosing the boundaries of classes.
This constitutes the basis of any classification using chorological laws. These laws determine the
relationships existing between semantic or thematic units (defined by intrinsic factors) and their
geographic distribution in three-dimensional landscape (defined by extrinsic factors).

7.2.4 Segmentation of an histogram

Another method of analysing an histogram consists of defining a class in the histogram, assigning a
given colour (preferably bright) and representing all the pixels belonging to this histogram class by the
same colour on the image. The spatial distribution and organisation of all the pixels having the same
value of digital numbers on the Image can thus be analysed. The same procedure can be followed for
a neighbouring class or for a group of classes, which aids in relating the geographic position of objects
with their spectral characteristics on the image.
This phase constitutes the theoretical basis of classification to be ultimately carried out since it
enables establishing the relationship between spectral and geographic aspects and hence defining
the chorological laws. This would facilitate choice of radiometric boundaries for classifications or tracing
of boundaries of training zones, which serve as nuclei for classifications.
All these operations can be carried out rapidly using image-processing software and information
can be acquired w ithout conserving it. When it is desired to protect this information on a
geographic plane, a new band is constituted. The latter carries over the same geographic space the
value of every pixel after its transformation. This new band can be used again in a subsequent stage
of processing.

7.2.5 Arithmetic combination of bands

As the values of digital numbers of various bands (between 0 and 255) are geographically attached to
a pixel, it is mathematically possible to make any arithmetic combination of bands and geographically
assign it to the same pixel. However, it must be ensured that It corresponds to something. It is commonly
used for several objectives:
— to construct indices by associating a theme with the value (on the basis of radiometry or
diachronism),
— to Improve the representation of information and to reduce the amount of Information,
— to superpose or integrate the results of various classifications.
134 Processing of Remote Sensing Data

H Computation o f indices
Vegetation indices were developed from investigations on reflectance. Caution needs to be exercised
when these are to be used on images, i.e., no longer from reflectance data but from digital numbers.
In fact, unlike reflectance which when acquired in normal experimental conditions varies from 0 to
100% in each band, the same is not true for digital numbers. The dynamic range of each band is
independent and consequently the same value of digital number does not necessarily have the same
significance in each band. Thus, before interpreting the result of an index, it is necessary to spectrally
calibrate the two bands relative to each other.
In 8"bit image processing, all the results of algebraic operations should be between 0 and 255 to
enable display on the monitor. If the two bands are not spectrally calibrated, it is quite possible for
example that the difference IR-R would not be positive. When negative values are obtained, depending
on the program, they are either set to zero. In which case they appear on the histogram as a class with
a high frequency, or the negative values such as -1 , -2 , -3 , etc., are re-coded as 255, 254, 253, etc.
To avoid this inconvenience, an arbitrary value such as 128 for example can be added. This aids in
adjusting the result of computation of the index. If the results contain decimal values, they are rounded
off. In such cases, the result is multiplied by a coefficient such as 10, 20, 100, etc., if they are to be
interpreted.
After every computation of arithmetic functions, it Is necessary to view the histogram for evaluating
the distribution obtained. The image obtained is in black and white and can be used to make classes
and hence to determine the limits. This phase is most difficult. It necessitates good field knowledge for
obtaining reliable information. We can proceed as in the preceding case for identifying some classes
by composing a Look Up Table.

M Vegetation indices
Various vegetation indices (see Chap. 4) are most often computed based on an arithmetic combination
of digital numbers of the near infrared (NIR) and the 675 nm (R) bands (CD 7.8). The difference NIR-
R was computed for the image of Brienne. As this difference was small, a constant equal to 100 was
added so as to obtain distinct grey levels on the monitor. Six classes were determined which could be
coloured as indicated in Table 7.6. Using this colour composite, large forest blocks and riverine forests
were identified as green, bare soils as black and orange, soils bearing chlorophyllian crops as yellow
and mixed zones between soil and vegetation as blue and magenta.
The NIR/R index was also computed. If the digital numbers are high, vegetation exhibits a greater
chlorophyllian activity. This index is not applicable for low digital numbers as interpretation is difficult,
except when field-training data are available. Another Index or another method has to be used.
The Normalised Difference Vegetation Index (NDVI) is computed from the same NIR and R data
as (NIR - R)/(NIR + R).This Index can also be described as [(NIR/R)-1]/[(NIR/R)+1].Then it is similar

Table 7.6: Interpretation of NIR-R index for image of Brienne

Digital numbers of classes Colour on image Legend

100 Black Very bright to dark soils and water

101 to 121 Orange Very dark soils
122 to 133 Magenta Soils with vegetation
regrowth or seedlings
134 to 152 Blue Chlorophyllian vegetation and shadows
153 to 186 Bright green Chlorophyllian vegetation (trees)
187 to 255 Yellow Very chlorophyllian vegetation
 Processing and Interpretation 135

to the first and is based on the same type of interpretation. However, as it lies between 0 and 1, its
dynamic range is much smaller and Its interpretation is still difficult.
These three indices show good correlation (Table 7.7).

Table 7.7; Correlation between three vegetation indices

Index N IR -R NIR/R

NIR/R 0.961 1
NDVI 0.941 0.962

Three units, viz., grasslands, non-chlorophyllian vegetation and zones of sparse vegetation cover,
were identified on the image of Brienne using Bayesian classification (see Chap. 9).The NDVI indices
are practically the same. They vary respectively in the ranges 128 to 144,131 to 140 and 125 to 135.
Hence, it is impossible to differentiate between these three units on the basis of index, whereas it is
possible using a two-dimensional red/infrared histogram. These indices do not provide finer
discriminations. On the other hand, It Is possible to distinguish between water, bare soils and vegetation
(Fig. 7.5), for which the NDVI indices for the Brienne Image are 108 to 120,121 to 134 and 135 to 151
respectively.

■ Brightness index
Some authors have defined a brightness index as the square root of the sum of squares of the two
bands, NIR and R:

Bl = Vnir^ -R^
This amounts to calculating a value equivalent to albedo, limited to a wavelength Interval of 500
nm to 1000 nm. Interpretation of this index is not always easy but it clearly differentiates shadow zones
from illuminated zones. It gives an Image that differs little from a panchromatic image.
Various soils could be distinguished on the Brienne image using this index; however, cultivated
soils could not be Identified (CD 7.9). Rivers are differentiated better on this image than with the NIR
band and much better than the NDVI image.

Fig. 7.5: NDVI index histogram for image of Brienne. Water (108 to 120), bare soils
(121 to 134) and vegetation (135 to 151) can be distinguished.
136 Processing of Remote Sensing Data

■ Modification of representation of information

One can readily compute the sum IR+R and the difference IR-R of the two bands. The colour composite
of these two new bands often gives a greater dynamic range than that observed in the case of the
initial bands and the differentiation is much better. In a number of cases, the visual effect Is hardly
different from the more elegant solution, which consists of computing linear equations corresponding
to the principal components (see below).

■ Integration of results of various classifications

When several classifications that are mutually exclusive are made in different parts of the image (see
masks below), arithmetic combinations can be used to superpose them one over the other. Thus, let
us assume that a classification Is made by dividing the forests Into five classes ( F I, F2, ..., F5) and
another classification by separating the cultivated soils Into 12 classes (C l, C2, ..., C12). In order to
obtain on a single document (I) Interpretation of forest (F) and cultivated zones (C) from the entire
scene, the band comprising cultivation data may be transformed by multiplying it by 10. The classes
then vary from 10 to 120. The classification of forests Is next combined with the latter and we obtain on
a new band the entire information divided into 17 classes (see Chap. 8): Image = 10 x cultivation +
forest.
In this application, arithmetic combination forms only an intermediate tool for preparing a final
representation of information.

7.2.6 Statistical analysis of bands

Bands can be analysed pixel by pixel but nowadays we process images containing 1000 x 1000 pixels
to 3000 X 3000 pixels or more, i.e., from one to ten million pixels. It is hence Imperative to study the
statistics of these images for characterising their principal components before undertaking classification.
Almost all image processing software Includes statistical programs but in a more or less precise
manner. These programs provide the shape of histogram, mean and standard deviation, but they are
not necessarily the best parameters to define the distribution curves since the populations In image
data are not Gaussian. Nevertheless, when diachronic bands are compared, these parameters may
be initially useful to study radiometric variations of objects between two different dates.
Thus it was observed for TM images of Brienne (© ESA (1990 and 1992) Acquired by Fucino
Station, Distributed by Eurimage-Geosys, by courtesy of Eurimage) that the means and standard
deviations were greater by 2 to 3 values of digital numbers for the image of May 1992 (CM) than for the
image of April 1990 (Table 7.8). However, this difference is much greater for band 4 since it changes
from 58.9 to 91. It was verified that the vegetation has a higher chlorophyllian activity in the month of
May (CD 7.10).
It is essential to study the histograms closely and preferably directly using the digital numbers.
It is difficult to give rules applicable to all cases for statistical interpretation of bands since several
factors Interfere. In fact, the values of digital numbers are not directly obtained from radiance received
by the satellite sensor (see Chapter 2). In the case of SPOT images, gain may significantly modify the
position of the histogram between its limits 0 and 255 (see CD: SPOT Satellite).
The respective positions of the histograms of various images are not necessarily the same as
those obtained from reflectance curves. The histogram of the green band (b1) often shows a higher
mean than that of the red band (b2). Values corresponding to the Infrared band are usually higher (see
Table 7.1). Thus, for classifying images of a given date the dynamic range of each band for each
image has to be reduced to a common dynamic range for all. Reciprocally, when an image is divided
Into image segments, the histograms change.
 Processing and Interpretation 137
Table 7.8: Mean, standard deviation and correlation between the bands of two Thematic Mapper images.
M: Mean; SD: Standard deviation; c l to c7: TM bands for the month of April; CM1 to CMS: TM bands for the
month of May. Correlation coefficients between bands of the same image are shown in bold.

M SD cl c2 c3 c4 c5 c7 CM1 CM2 CM3 CM4 CMS

cl 70.S1 10.04 1.000

c2 31.89 7.63 0.974 1.000
c3 34.S2 13.S4 0.957 0.981 1.000
c4 S8.90 17.22 0.098 0.113 0.012 1.000
cS 63.37 23.86 0.877 0.900 0.916 0.176 1.000
c7 30.14 17.12 0.889 0.918 0.954-0.024 0.961 1.000
CM1 73.70 12.84 0.371 0.361 0.333 0.177 0.297 0.291 1.000
CM2 34.97 9.66 0.343 0.342 0.321 0.200 0.305 0.290 0.968 1.000
CM3 35.75 16.67 0.323 0.326 0.312 0.190 0.301 0.289 0.960 0.982 1.000
CM4 91.00 23.70 0.144 0.151 0.129 0.514 0.262 0.142-0.184 -0.140 -0.196 1.000
CMS 66.65 27.26 0.304 0.309 0.301 0.302 0.367 0.312 0.837 0.848 0.875 0.025 1.000
CM7 29.47 19.93 0.289 0.296 0.290 0.193 0.309 0.291 0.894 0.900 0.940 -0.193 0.954

A method of image segmentation is based on this principle. An Image is divided into n (4 or 9 for
example) segments of equal area. Histograms of each segment are compared with those of the complete
image (see Table 7.4). If the difference is greater than a given threshold, the segments are distributed
into different units. Then, classification is continued for each unit and the function is iterated.
Using the histograms in a similar manner, we can also compare two images, one complete and
the other subsampled, and determine whether information is lost or not. This enables us to know
whether noise at the spatial level Is high or low: If the histograms are similar, this indicates that nothing
is lost by subsampling and hence there is little useful information from the spatial point of view and
thus ‘noise’ exists.
It Is also advisable to examine correlations between bands of the same image. Some constants
exist when the visible, near infrared and middle Infrared bands are studied. In most cases, visible
bands correlate very well with one another (Table 7.8) and fairly well with those of middle infrared. For
the SPOT Image of Brienne the correlation coefficient between the two bands b1 and b2 is 0.981 (CD
7.11). The near Infrared band usually correlates less well and minimum correlation is observed for red
band, even with the nearer blue band. These remarks evidently apply for an image comprising areas
of water, soil and vegetation.
In the case of diachronic images (Table 7.8), when bands of images acquired in different seasons
are compared, since the areas covered by vegetation change considerably the correlation coefficients
also change but in most cases the separation mentioned above between visible and infrared bands
persists.

7.2.7 Principal component analysis

As seen above, satellite image data exhibit different degrees of correlation. It is interesting to investigate
a representation of the same information in another configuration, for example by projecting all the
pixels on axes that do not correlate. The principal axes of principal component analysis derived from
138 Processing of Remote Sensing Data

linear combinations of initial bands do not correlate. Thus, when three SPOT bands, b1, b2 and b3 are
used, we obtain three principal components PCA1, PCA2 and PCA3 whose equations as a function of
the initial bands are as follows:

General case Example for SPOT image of Brienne

PCA1 =ab1 + b b 2 + cb3 PCA1 = 0.701 b1 + 0.696 b2 + 0.157 b3

PCA2 = db1 +e b 2 + fb 3 PCA2 = - 0.076 b1 - 0.146 b2 + 0.986 b3
PCA3 = gb1 +/7b2 + /b3 PCA3 = 0.709 b1 - 0.703 b2 - 0.049 b3

These axes also exhibit the following interesting property: information on axis 1 is most widely
stretched and its value is the highest. Hence It is the most important axis since it carries maximum
information. Axes 2 and 3 follow successively.
If the Information carried by each of the bands, b1, b2 and b3. Is considered equal to 1, the total
information would be 3 for a SPOT image. After a principal component analysis, it is common to obtain
the information (actual values) distributed In a manner similar to that of Brienne region:
— for PCA1:2.007;
— for PCA2: 0.977;
— for PCA3: 0.016.
This may be Interpreted as follows:
— if the PCA1 axis alone is retained, nearly two-thirds of the information contained in the entire
image is preserved;
— If the first two axes PCA1 and PCA2 are retained, almost the entire information is preserved
and hardly a few parts per thousand are lost.
It follows from the above that two axes are often sufficient to represent the essential content of an
image in three dimensions. The three-dimensional representation of all the pixels of an image resembles
more a cutlet than a balloon!
It Is hence necessary to analyse what the three axes of the principal component analysis signify.
This Is done by studying the linear equations that define the three PCA axes (vectors). In most cases,
the following orders of magnitude are observed.
For PCA1, a and b are of the same sign and high, while c is of the same sign but small. Consequently,
the first component Is quite close to the sum of digital numbers for a given pixel. This is similar to the
result that would be obtained for a panchromatic band. In other words, two-thirds of the Information is
contained in the sum of three bands.
For PCA2, f Is very high and opposite in sign to c/and e, which are smaller. As a consequence,
the second component Indicates the contrast between the visible and infrared bands. This is quite
similar to a vegetation index, which could be roughly formulated as IR-R. Hence, diverse types of
vegetation are detected better on this band. One-third of the information is contained in the difference
between visible and near Infrared bands.
In the case of PCA3, grand h are quite high and of opposite sign, while / is most often very small.
This component indicates differences between the two visible bands since the Infrared band plays
almost no part. Noise is observed in this component, which represents only one- or two-hundredths or
less of information. This band aids In detecting linear features of the image. The latter are visible in
particular over water, which is understandable since it Is on this component that the information relative
to visible bands Is maximised, in which water exhibits maximum dynamic range. It is also in this
component that some elements not noticed in the other components are observed, viz., very bright
zones of small extent such as open quarries. However, shadow zones are detected from this component.
Hence the band derived from this component can be utilised to make arithmetic combinations that
tend to highlight a particular point, such as a shadow zone, or contrarily suppress the noise.
 Processing and Interpretation 139

When a PCA is carried out using the six bands of a TM image, with visible, near infrared and
reflective middle infrared bands, the results are as follows. The first axis corresponds to the sum of all
bands except the near infrared (TM4). The second corresponds to the near infrared band (TM4). The
third axis represents the contrast between the visible bands (TM1, TM2, TM3) and those of the middle
infrared band (TM5, TM7).The fourth shows the contrast of the blue band (TM1) against the green and
red bands (TM2, TM3). The fifth axis indicates the contrast between the two middle infrared bands
TM5 and TM7. The sixth mainly compares green (TM2) against red (TM3). Depending on the images,
the composition of the axes varies, but often the same type of composition Is identified.
Once the behaviour of the axes of principal component analysis is understood, it Is useful to
visualise the information of these axes in its spatial form by creating a new band for each PCA axis.
Each band is constructed in black and white and used to study the geographic distribution of the
various objects to be Identified, in fact, the value of each pixel cannot be readily interpreted since a
linear combination of the initial bands has to be analysed. The dynamic range of the signal of each
PCA axis differs considerably from that of bands of the satellite image. It extends over the entire 256
possible values (Fig. 7.6). In some cases, the contrast of each new band can be improved by studying
every histogram, as indicated earlier. If the three bands are represented on a three-dimensional
histogram, it can be seen that point distribution Is the widest possible. This is due to the fact that the
bands are uncorrelated.
Using the first PCA axis, a new band of PCA1 can be constructed (CD 7.11), which can be
interpreted visually or digitally in the same way as done for a panchromatic image. All geographic
details (including roads) become apparent on this new image and hence identification of features
becomes easy if the ‘square-roof function is used for enhancement, as in the case of the Brienne
image (Fig. 7.6).
In the case of PCA2, major land cover zones are clearly identified since the information on this
axis corresponds to a contrast between visible and Infrared bands. Hence It highlights various types of
chlorophyllian vegetation, bare soils and water bodies.
On the image of Brienne (CD 7.11) cultivated zones can be seen as white hues and forests and
riverine forests as grey. Bare soils are Identified as dark grey areas and water as black. The histogram
of PCA2 (Fig. 7.7) is similar in its distribution to that of band b3 (Fig. 7.1). The difference lies in the fact
that the histogram of PCA2 is stretched over 256 levels instead of 190 levels for band 3.
In the case of PCA3, If there is no particular feature and if the noise is distributed over the entire
image, everything appears a medium grey. Lineage effects are observed in this band. In some cases,
this Image that contains only a few per cents of information, can be visually interpreted quite well. In
the image of Brienne, even the small amount of Information enabled a correct interpretation and no

Fig. 7.6; Histogram of image of Brienne projected on PCA1.

140 Processing of Remote Sensing Data

evidence of noise was found. If high digital numbers of the histogram (Fig. 7.8) are carefully observed,
a specific form of noise can be noticed between 128 and 180. This Is spatially discernible In areas of
reservoir waters (CD 7.11).

Fig. 7.7: Histogram of image of Brienne projected on PCA2 axis.

Digital numbers

Fig. 7.8: Histogram of image of Brienne projected on PCA3 axis.

■ Colour composites
The importance of using the new bands constructed from the principal component analysis lies In the
fact that the three components are independent, which Is recommended for some classification methods.
Once the visual significance that can be drawn from an image derived from projection of each
axis of principal component analysis is understood, a colour image can be constructed using two or
three bands.The choice of colours for the final visualisation is Important since the quantity of information
differs for each component.
If such a component is to be evaluated, it is better to assign it a red colour, which offers the
advantage of several shades. On the other hand, very few shades can be differentiated in blue; hence
it can be used for a band whose visual impact is to be reduced.
A colour composite can also be constructed by choosing colours of each component in such a
way that a given object Is displayed in a predefined colour. In such a case. It Is necessary to determine
the characteristics of this object in each of the three components, which is not simple, and to apply
rules of colour composition.
 Processing and Interpretation 141

Comparison of new bands

The first component can be used to represent the entire image. This allows use of results of classification
or arithmetic combinations (for example, an index) for the second or third band. By assigning a different
colour to each of these three bands, the results obtained can be compared by analysing the colours
on the monitor.

■ Diachronic comparison
When images of the same region acquired on different dates are to be compared by means of colour
composites for a diachronic study, analysis should be limited to three bands. Consequently, one solution
is to take for each date one of the principal components, the one considered as best representing the
entire image of the given date. Once constructed, a study of colours on the monitor provides a diachronic
analysis (CD 7.12).

■ Comparison of multiple processing methods

The results of several treatments, each of which constitutes a new band, can be compared by interpreting
correlation coefficients, intrinsic values and vectors of principal component analysis. For this purpose,
we Identify the processing methods that play the same role (these are correlated with one another)
from those that are dissimilar (these are not correlated) and determine the weight of each. Obviously,
the comparison, statistical and textural, is carried out pixel by pixel and does not integrate the geographic
approach (for the method of neighbourhood, see Chapter 11).

7.3 MASKS
Masking consists of hiding part of the image and preserving the other part Intact. A mask may be
geometric or radiometric.
Commonly, only part of the image is of Interest for a given thematic application, such as land or
marine region, plateaus, slopes or valleys, swamps or well-drained zones, etc. Moreover, in order to
avoid confusion in classification of certain features, it Is important to preliminarily segment the Image
into major land cover classes. Thus, we can avoid confusion between diverse bare agricultural soils
and urban zones, or between crops and permanent grasslands. In montane zones. It Is desirable to
separate northern exposed slopes from southern; this facilitates land cover classification without
application of radiometric corrections required for compensating solar illumination differences. Lastly,
when a cloud cover shadows part of a scene, it is preferable to apply a mask so that pixels situated in
the shadow zone can be classified separately from those outside this zone.
In one case, only part of the image is normally analysed and consequently it is preferable to
suppress the geographic area that is not processed. For this purpose, a geometric type of segmentation
is carried out. The boundaries of the zone of interest are traced and the remaining area is eliminated.
This is known as geometric masking.
In the other case, it may be desired to eliminate a theme that does not form part of the study and
which corresponds to a radiometric group. This is referred to as radiometric masking.

7.3.1 Radiometric masking

Radiometric masking is designed to suppress part of the objects in one or several bands based on
digital numbers. In this case, the spectral significance of the image Is analysed. As mentioned earlier,
It is possible to prepare a band according to the ultimate theme of interest by suppressing all other
objects that do not correspond to the theme of investigation. These transformations should be taken
into consideration during the final processing of images. Suppression of some digital numbers is
142 Processing of Remote Sensing Data

significant only when it is ensured that a two-way connection exists between the latter and the object
under study. Very often, we may eliminate other pixels that belong to the object of study but In the
conditions of image acquisition, have the same digital numbers as those of the pixels outside the
theme. In such a case, geographic or logical masking should be applied.
An ascendant hierarchic classification (see below) is used for digital numbers of pixels. This
classification provides a grouping of pixels on a unique statistical basis that ensures distinction between
classes; on the other hand, grouping of objects of different nature but with identical digital numbers
should be controlled. One way of doing this is to use the two- or three-dimensional histogram and
ensure the nature of pixels by controlling them on the image. By this method we could separate
permanent grasslands from other land cover classes on an end-winter scene of Lorraine, using visible
and near infrared bands. After necessary geometric corrections, this classification served as a mask
for scenes acquired in different seasons covering the same region.

7.3.2 Geographic masking

Two types of geographic masks may be distinguished, viz., manual and thematic.

■ Manual masks
These masks are obtained by manual delineation of the boundary of zones not included in the study.
When the region of study is not identical with that of the satellite Image, it is essential to mask
from the beginning of processing the part of the image not involved in the study. When a digital
information layer corresponding to the boundaries of study Is not available (otherwise a logical mask
is applied, see below), boundaries of the region of study can be drawn manually. This can be defined
as geographic masking. Once the boundaries are drawn, it operates like a logical mask.
In an investigation concerned only with agricultural areas, the boundary of the urban zones was
determined by computer-aided visual Interpretation on a SPOT scene in the west of Paris. The work of
the interpreter was aided by preparation of a plane (LP), corresponding to a Laplacian filtering technique
(see Chapter 12), of the near infrared band (b3) of the SPOT scene. Tricolour combination of this new
plane (LP) with bands b3 and b2 enabled us to more readily identify urban expansion zones. In fact,
the plane resulting from filtering (LP) showed a distinct contrast between urban zones, characterised
by a very high density of sinuous curves (similar to spaghetti), and agricultural or forest zones of less
marked contrast, represented by polygons of varied dimensions.

■ Thematic masks
These masks are normally developed from exogenous data and require a prior geometric calibration
between these data and remote sensing data.
A digital terrain model for example, enables extraction of some slopes for isolating hill tops or
valley floors. A slope map can be used to select the slopes included In the range of interest. If the
variable of exposure is added to this map, we get a method of classifying land cover units exposed to
different illumination conditions without applying radiometric corrections. Similarly, any thematic layer
of a geographic information system emerging from a map or an earlier classification can be used to
create a mask.

7.3.3 Logical masking

Depending on the software, several techniques of logical masking exist. In all cases, we start with an
Initial Image I (characterised by digital numbers), make use of an image M to mask the unwanted part
(characterised by the value 0) and by a logical operation generate the final Image F. In this image, all
 Processing and Interpretation 143

the pixels that are not to be considered for the study are represented by the same value, for example
zero (Table 7.9).
The logical expression used in masking is of the form:
IF (mask Image M = 1) THEN put the pixel value of the initial Image
IF NOT put the pixel value 0
A mask can also be prepared (Table 7.10) by considering that class 3, for example, of a classified
image C constitutes the mask for the initial image I, and by obtaining the value of 50 for the pixels
eliminated.

Table 7.9: Design of a logical mask, starting with an initial image I, to obtain the final image F using an
intermediate image of masking M.The values indicated in the image are digital numbers.

Initial image I Mask M Final image F

12 24 25 17 58 69 64 0 0 0 0 1 1 1 0 0 0 0 58 69 64
11 28 31 67 98 85 79 0 0 0 1 1 1 1 0 0 0 67 98 8579
17 21 35 62 78 29 27 0 0 1 1 10 0 0 0 35 62 78 0 0

Table 7.10: Design of a logical mask, starting with an initial Image I, to obtain the final image F using an
intermediate classified image C. Values indicated in the images are digital numbers.

Initial image I Classified image C Final image F

12 24 25 17 58 69 64 3 3 3 3 1 2 4 50 50 50 50 58 69 64
11 28 31 67 98 85 79 3 3 3 1 2 2 4 50 50 50 67 98 85 79
17 21 35 62 78 29 27 3 3 2 2 1 3 3 50 50 35 62 78 50 50

The mask for this case can be written as follows:

(C = 3) take the class 3 of image C
True: 50 if it is true that pixel p of image C is 3 then put 50
False: I if it is false that pixel p of image C is 3, then put the value of the pixel of the initial image I
In this case, the final Image F generated comprises all the values of pixels of the initial image
irrespective of the classes to which they belonged in the image C, except for class 3 in which the pixels
take the value of 50.
Masks offer the advantage of dividing the image into several segments each of which can be
processed differently (see CD 8.4). Subsequently the entire image is reconstructed by arithmetic or
logical operations (see sec. 7.2.5). This Is used for detailed analysis of dark soils and vegetation in
studies by ascendant hierarchic classification.
A geographic zone corresponding to a civil parish for example can be delineated at the beginning
of processing and the remaining area masked. In such a case, the statistics of different groups of
classification of the civil parish can be defined.

7.4 CONCLUSION
The various techniques of preliminary processing enable construction of Images suitable for visual
Interpretation (see Chapter 5). Even at this level of processing, some questions can be answered.
Finally, it is necessary to understand well all these preliminary processing techniques since they are
used not only before, but also after classification of images. The latter will be discussed in the following
chapters.
 8
Unsupervised Classification
Image processing methods can be categorised in two groups:
— Unsupervised classification methods in which the data are classified according to their structure.
The interpreter gives no a priori information about the objects to be determined. However, as mentioned
in Chapter 6, he (she) interacts by choosing the number of groups, thresholds, etc.
— Supervised classification methods based on searching for features that are similar to reference
objects. The latter can be defined radiometrically on a multidimensional histogram or marked
geographically on an Image (nuclei, reference zones, etc.).

8.1 ASCENDANT HIERARCHIC CLASSIFICATION

8.1.1 Principles
Ascendant hierarchic classification is based on the measure of a distance, such as Euclidean, computed
for each of the N pixels having different spectral characteristics. On an image comprising ten million
pixels, the order of magnitude of N is only about ten thousand.
We search in a table of N (N - 1) distances two pixels that have the shortest distance between
them and decide to group them together. These two pixels are then converted into a single fictitious
entity, viz., a mean pixel.
A dendrogram is drawn representing various phases of classification, placing pixels on the abscissa
and a value similar to an intergroup distance on the ordinate. The two pixels are joined on this diagram
by a segment equivalent to their distance.
Distances between N - 1 pixels (real or fictitious) are recomputed and pixels that have the shortest
distance between them identified. These may be two real pixels or one real and one fictitious,
representing a group. These two new units are combined.The procedure is thus continued until all the
pixels are grouped. This involves N - 1 phases of grouping of all the pixels or groups of pixels into a
single large group (CD 8.1, Fig. 8.1, Tables 8.1 and 8.2).
Classifying all pixels Into a single group amounts to obtaining a single group for the entire image.
Such a classification is certainly correct since there Is only one group and the intragroup distance is
the shortest possible, since it corresponds to that of the entire Image, but this does not constitute the
objective of classification. Similarly, it is possible to classify every pixel into a separate group.Thus the
intragroup (intrapixel) distance is the minimum, namely zero, but this likewise is not the aim of
classification. Moreover, this information is always available since It Is strictly comparable to the initial
histogram.
 Processing and Interpretation 145

8.1.2 Groups and legend

A proposed classification of all pixels Is represented in a dendrogram. The number of groups desired
must necessarily be selected, which constitutes external intervention.
This classification is based only on a mathematical model of classification and hence provides no
information on the nature of the groups. One has to guess what the groups represent in order to
establish a meaningful legend. Also, the choice of groups ought to be determined by the possibility of
thematically defining each group identified.

■ Choice of groups
Choice of groups primarily depends on the four types of objectives pursued.

□ Reconnaissance
A primary aim could be preparation of a preliminary map of a region for which no information on the
theme understudy is available. In such a case, the main objective could be establishing a segmentation
of the region from processing of satellite images. It follows that a priori all the pixels in the entire zone
under investigation should be classified. It Is not necessary to define a precise legend for each group
since this would be done subsequently during field studies (CD 8.1).

□ General classification of units

A second objective could be to give a general classification of the most characteristic units of the
region in a scale that does not indicate too small spatial units. As very small zones cannot be represented,
they need to be combined with others so as to obtain an areal extent that is compatible with the
minimum acceptable area for representation on the map or greater than a given threshold. Thus for
example, the minimum size of objects represented on the CORINE Land Cover databank of 1/100,000
scale is 25 ha. In the multispectral SPOT image this corresponds to 625 pixels. Consequently, objects
that are too small are combined with those of the semantically nearest group.

□ Definition of major thematic zones

A third objective could be to establish just a very general segmentation of the region under investigation
to identify only a few major zones for a given thematic study. Unsupervised classification is very well
suited for this objective. Often we are not interested in all the possible themes in a region and sort only
the important zones. If, for example, only forests are being investigated, all bare soils need to be
eliminated. This very rapid method of classification enables us to select on the dendrogram a minimum
number of groups relevant to the theme ‘bare soils’. This is done based on a radiometric model. Then
these groups are masked and an image is thereby obtained on which all pixels corresponding to
forests are preserved. In this process. It is better to retain pixels that might ultimately prove useless
rather than eliminate those that might form part of the theme.

□ Precise thematic study

A fourth objective could be to study a precise theme without considering its size or spatial distribution.
Areal extent of the zones retained does not constitute a criterion. As it is very fast, unsupervised
classification may be very useful in such cases if it Is observed that the unit under investigation appears
in one of the proposed groups. Then only this group is retained and through masking an image obtained
wherein the sizes and spatial distribution of this unit can be readily analysed, as was done for lakes on
the Image of Brienne.

■ Legend
The purpose of a legend is to give meaning to the classification created. Each group is assigned a
name with reference to a semantic or a geographic model.
146 Processing of Remote Sensing Data

irrespective of the objective of image segmentation by this method, a legend ought to be given for
each group identified. The legend cannot be beyond this mode of classification since no a priori
hypothesis is made about radiometric values. In order to define themes that might cover different
groups, a histogram of digital numbers Is determined for each group so as to characterise the digital
behaviour of the theme relative to the group. For this purpose the interpretation model of digital numbers
on a three-dimensional histogram is employed (Fig. 7.4).
Analysis of spatial distribution of pixels belonging to a single group may also give interpretable
geographic information. In such a case, a geographic approach is adopted and the positions and
forms of zones of each group obtained are analysed, which may give information about the objects
combined into a given group. Rivers and talwegs are readily recognised by their linear and curved
forms and by their hierarchic relationships such as joining of a watercourse with another to form a
higher level of flow. Unlike roads, they do not criss-cross one another. However, there can always be
exceptions as in the case of channels or ditches that more or less follow valleys or cut across hills or
even may in some cases intersect rivers.

■ Analysis of a dendrogram
A dendrogram can be analysed systematically. Two groups are chosen and the geographic spatial
distribution of each is observed on the image. As a matter of fact, the dendrogram gives only semantic
Information and the number of pixels pertaining to a group is not ascertainable. It varies from a few
tens to several hundred thousand. This aids in judging the aptness of preserving a particular group: a
group cannot be preserved because it represents a very small number of pixels or because the pixels
it contains are scattered over the entire image and do not constitute sufficiently large zones.
The same analysis is repeated by successively increasing the number of groups and thereby the
optimum number of groups for the objective of investigation is determined (CD 8.1).

H Interpretation of a dendrogram
A dendrogram is provided in some programs such as TeraVue (see Fig. 8.1). This is essential for
understanding the statistical approach of segmentation into groups.
A dendrogram Is an expression of the results of classification. For interpretation, it needs to be
considered similar to a mobile system. This means that the order of pixels and regroupings on the
dendrogram constitute only a projection of a group, which in fact is a three-dimensional unit, onto a
plane. Thus the following two representations (Table 8.1) are identical. It would be a serious error to
consider that according to the first scheme group 18 is closer to group 9 than group 15, or that groups
5 and 15 are farther apart than groups 27 and 15.

Table 8.1: Two forms of representation of a dendrogram derived from an ascendant hierarchic classification

27 18 15 18 15 27

Number of groups
 Processing and Interpretation 147

The dendrogram given as an example indicates that the two nearest groups are 5 and 27. They
constitute a very homogeneous unit because they are separated by a small distance (grey level threshold
expressed on the dendrogram). These two groups are connected to group 9 also by a small intergroup
distance. On the other hand, groups 3 and 18 are also connected but with a larger distance than the
group (5,27,9) since they necessitate three levels on the dendrogram. Group 15 is very different from
the groups (5, 27, 9) and (3, 18). Nevertheless, it combines with group (3, 18).
Mathematically, division of a dendrogram is carried out at the same level. On proceeding through
It the following segmentations will be obtained (Table 8.2):
1. Two groups: One a part of group A: 3 ,18,15 and the other part of group B: 5,27,9. The intragroup
distance of A Is greater than that of B; group A is thus more heterogeneous than group B.
2 . Three groups: These are groups C: 15, D: 3,18, and E: 9, 27, 5. Group C Is thereby isolated and
hence becomes very homogeneous since Its intragroup distance Is zero. Group D becomes more
homogeneous than the preceding group A, since group 15, which significantly contributes to the
high value of the intragroup distance of A, is detached from it. However, the intragroup distance of
D is greater than that of E. Group E is similar to group B and its intergroup distance Is the same.
It should be noted that when we move to the lowest level in the dendrogram, the same three
groups are found, and the same holds to true when we go down to three levels. This indicates that
division Into three levels Is steadfast.
3. Four groups: These are groups F: 15, G: 3, H: 18, K: 9, 27, 5. Two groups are not affected by this
segmentation. Group F retains the same characteristics as found in group C and group K those of
groups E and B. Thus the former group D is divided into two groups, G and H, which become
homogeneous since each intragroup distance is zero.
4. Five or more groups: When there are five groups, segmentation results in division of group K into
two: on one side L: 9 and on the other M: 5, 27. At this level, intragroup distances are all very small
or zero and on the other hand, the number of groups has considerably Increased.

Table 8.2: Analysis of segmentation of a single dendrogram according to the choice of number of groups

Number of groups 18 15 27

It is also possible to define 6 groups; but this amounts to further dividing each group of pixels:
each intragroup distance Is zero and the intergroup distances are maximal. Theoretically speaking,
such a classification Is the best; however, since no regrouping was carried out it is no longer a
classification!

■ Interpretation of classification
We cannot determine on a dendrogram the area corresponding to each group nor the geographic
distribution of pixels that occur in the same group. The possibility offered by image processing to
148 Processing of Remote Sensing Data

immediately and interactively see the dendrogram, the number of groups chosen and the corresponding
spatial result is of great significance in interpretation of classification (CD 8.1). It is thus possible to
choose the desired number of groups best suited to the objective pursued.
When the cross-section of the dendrogram that is most suitable to the objective of interest is
chosen, the groups can be combined based not only on statistical criteria, but also on thematic
considerations. It suffices to allot a single colour and a single attribute to various groups that we desire
to combine. Thus groups corresponding to different types of water (CD 8.1) or to different vegetation
units can be combined. A new image is thereby obtained, which may serve as a mask in subsequent
processing.
A new ascendant hierarchic classification can be applied to this image to be processed as done
earlier. As it contains fewer pixels, a different and more diversified dendrogram would be obtained for
themes that are not masked and would be retained.

8.2 EXAMPLE: IMAGE OF BRIENNE REGION

8.2.1 Statistical interpretation
S Analysis
Various images obtained from division of a dendrogram at different levels that define different groups
(Fig. 8.1) can be analysed. For each group, the displayed colour for each band is the mean of digital
numbers of pixels belonging to this group (in TeraVue program).
The first division (Table 8.3) gives two groups, viz., very bright bare soils (group code No. 31) and
ail other themes. The second division displays water bodies. The third division Indicates chlorophyllian
crops which In the month of September could be beetroot and potato. The groups thus formed are
derived from a division of groups previously detected.
Thus groups 5 and 6 were obtained from division of unit 4 which comprised slightly dark bare
soils. Group 7 is segmented Into many groups, viz., 8, 10, 11 and 13 (Table 8.3 and Fig. 8.1).
The first 11 groups are well separated from those that follow and hence the tenth division is
robust. A new robust division can be made into 17 groups (Fig. 8.1).

I
5 classes

J 1 classes

17 classes

4 3 2 1 Numbers of 5 classes

---------------------\<— >I-------->-«H
11 8 13 17 10 7 14 4 15 12 6 5 3 2 1 16 9 Numbers of 17 classes

Fig. 8.1: Seventeen-group dendrogram obtained by ascendant hierarchic classification.

 Processing and Interpretation 149

Table 8.3: Interpretation of dendrogram of image of Brienne

Division no. Identification of groups appearing in Group code when

division 32 of them are required

0 Entire image
1 Very bright bare soils 31
2 Water 25
3 Chlorophyllian crops 24
4 Slightly dark bare soils 16
5 Slightly clear water 23
6 Unclear water 2 2

7 Fairly chlorophyllian vegetation (including forests) 7

8 Vegetation less chlorophyllian than 7 5
9 Sparse chlorophyllian vegetation on very bright soils 32
1 0 Vegetation and shadows or conifers 11

11 Highly chlorophyllian vegetation 1

1 2 Bright bare soils 19

13 Vegetation more chlorophyllian than 11 6

14 Very dark bare soils 14

15 Borders of ponds: soil and water 18
16 Almost white soils 30

When all the 32 groups are completed, the following groups can be differentiated:
— 12 groups for vegetation, the first of which (Table 8.3, No. 24) Is separated very early;
— 6 groups for water, 2 pertaining to turbid water;
— 3 groups constituting a mixture of water and soil;
— 9 groups for bare soils, 4 of which are very bright to white soils (In the Munsell code the value
is greater or equal to 5) and five darker soils;
— 2 groups represent soils sparsely covered by vegetation (winter crops already sown).
Clouds, being white, are confused with bright soils and their shadows with clear water. Only their
shape is useful in differentiating them.

■ Characterisation of groups
The groups need to be characterised by their effective pixel number which Is directly related to their
readability on the Image. As an approximate guide, the mean number of pixels can be taken as 36,000
(or 31.25 per thousand). Below this value, a group is under-represented.
Here, five units can be distinguished (Table 8.4) by taking the threshold limits as:
— twice the mean value or 72,000 pixels, division between A and B,
— mean value or 36,000 pixels, division between B and C,
— one per cent of the pixel population or 11,520 pixels, division between C and D,
— one per thousand or 1152 pixels, division between D and E.
The groups differ in effective number of pixels. In fact, the nine groups of units A and B, which
have a mean frequency of more than 100,000 pixels, represent nearly 85% of the image processed.
150 Processing of Remote Sensing Data

Table 8.4: Regrouping of 32 groups of ascendant hierarchic classification (ARC)

Unit AHC groups Sum of pixels for Value per Mean number of
the unit thousand pixels per group

A 12,16, 7,10, 8 , 6 823,904 715.9 137,317

B 19,9, 24 151,601 130.7 50,533
C 14, 25,5, 27, 20, 11, 15 147,048 127.6 2 1 ,0 0 0

D 4, 30, 26, 13, 31 25,273 21.9 5,054

E 1,3, 29,18, 2, 23, 28,17, 22, 21 4,174 3.6 379

■ Statistical interpretation according to objectives

If the objectives enumerated at the beginning of the study of ascendant hierarchic classification (sec.
8.1.2) are recalled, the following decisions can be taken:
For the first objective, viz., ‘segmentation of the image prior to field verification’, only groups with
a sufficiently high frequency of pixels are taken to enable ready detection in the field. Hence, only
groups A, B, C and D are used. Groups of unit E are too small as their mean area is only 15 ha. If pixels
are not combined into a single zone, it is difficult to mark them in the field and establish sufficiently
wide zones for making measurements. This also depends on the form and scattering of various map
zones contained in each of the groups.
For the second objective, viz., ‘general classification of the region’, statistical analysis alone does
not suffice; radiometric analysis is also necessary since the pixel frequency of a group is not the only
criterion for retaining a given segmentation. If it is decided to take only groups of more than 625 pixels
for a single zone, which corresponds to a resolution of 25 ha (as In the case of CORINE Land Cover),
only groups A, B, C and D can be retained. It then becomes necessary to combine all other groups into
the units retained. This necessitates radiometric analysis of the groups in order to give a name, at
least approximate, to each group.
For the other two objectives, viz., ‘thematic study of large zones’ and ‘precise study of objects’,
obviously the thematic significance of each previously defined group should be available.

8.2.2 Digital interpretation

One solution would be to define the groups from a field study. However, this involves a supervised
method of classification. Here, the objective is to make a semantic analysis of groups from their digital
characteristics.
Each group is defined by a statistical population for each band. Their spectral characteristics can
hence be approximately described by means of digital numbers recorded In different bands of the
Image (three in the present case). These may be termed digital characteristics.
While comparing digital numbers it should be remembered that the pixel frequencies of various
groups are extremely different. According to the radiometric model adapted to the values of digital
numbers of an image (see Chap. 4 and Chap. 7, Fig. 7.4), the groups can be categorised Into four
themes (CD 8.1), viz., water, soil, vegetation and soil and vegetation.
For precise classification of objects from their digital numbers, it should be remembered that this
approach is not effective unless the position of objects is studied band by band (see Chap. 9).
The following ten features may be defined based on their digital characteristics In the three spectral
bands of SPOT (Fig. 8.2):
 Processing and Interpretation 151

— clear water (group 25) 31,802 pixels,

— turbid water (group 13) 2,649 pixels,
— dark soil (group 15) 11,809 pixels,
— slightly dark soil (group 16) 198,786 pixels,
— bright soil (group 27) 17,194 pixels,
— white soil (group 31) 2,395 pixels,
— vegetation and shadows or conifers (group 11) 11,880 pixels,
— forest vegetation (group 10) 107,044 pixels,
— cultivated vegetation (group 24) 41,118 pixels,
— vegetation and soil (group 5) 26,983 pixels.

Fig. 8 .2 ; Digital characteristics of ten objects (image of Brienne).

Numbers given to groups are defined in the text.

Other groups are combined into four themes defined by their digital characteristics, as shown In
Fig. 8.3 (CD 8.1). For clarity of reading, the groups corresponding to vegetation have been distributed
in two figures.

■ Water
The digital characteristics of water can be recognised from the fact that digital numbers are the smallest
in band 3, but not in 1 or 2. The clearer the water (groups 25 and 26), the smaller the digital numbers.
Moreover, the greater the concentration of suspended particles and hence more turbid (groups 23,
22, 21,17, 28), the higher the values (Fig. 8.3, water).

■ Soils
Digital characteristics of soils exhibit little variation from band to band. However, digital values increase
with Increasing brightness of soils (Fig. 8.3, soils). In the case of white soils (groups 19, 27, 30), the
colour due mainly to chalk or gravel material (groups 31, 29), a convexity is observed in the digital
characteristic curve, which is distinct in band 2. For soils of medium brightness (group 16) the digital
characteristic can be approximated to a straight line. Lastly, in the case of darker soils (groups 12,14,
15) a concavity is observed in the digital characteristic, which is marked for band 2.
152 Processing of Remote Sensing Data

H Chlorophyllian vegetation
In the example presented above, the digital characteristics of various types of vegetation are marked
by low digital numbers In band 2, lower than those in band 1 and band 3, the latter DN values being
very high. Two units can be identified. One consists of vegetation cover with the lowest values in band
2 (Fig. 8.3, forests and others), which corresponds to forests and riverine forests (groups 9 and 10) as
well as much darker vegetation (group 11). The other unit comprises vegetation with the lowest values
in band 2 and the highest values in band 3 (Fig. 8.3, crops and others). This unit represents plants for
which chlorophyllian activity Is the highest (group 24).

Fig. 8.3; Digital characteristics of 32 groups derived from AHC of image of Brienne.

■ Mix of soil and vegetation

After isolating major groups of objects. It was observed that areas exist which exhibit digital
characteristics of vegetation (Fig. 8.3, vegetation and soils) but have a small concavity (groups 32, 2
and 5). Evidently these correspond to a mix of vegetation and soil, which has the effect of relatively
decreasing the concavity of the curve of band 2. Groups 2 and 32 have very high values and on an
IRC image a pink colour is perceived, indicating zones of sparse vegetation that may correspond to
regrowth after harvesting or to germination after sowing.

B Shadows and clouds

The IRC image shows some clouds (for example, row 725, column 605; row 728, column 450) as well
as their shadows (row 660, column 580; row 660, column 430). Two clouds are represented by pixels
 Processing and Interpretation 153

belonging to groups 31, 29, 27, 20, 19 and 16. Hence, they are spectrally homogeneous and their
digital characteristic is close to that of soils.
Shadows of clouds have very low values of digital numbers pertaining to groups 25, 26 and 15,
Interpreted as water or very dark soils.

■ Towns
The town of Brienne does not exhibit a specific digital characteristic but a mixed one (Fig. 8.3, town),
as described below:
— digital numbers of bright zones (groups 18,19, 27 and 30) which correspond to roofs of large
industrial and commercial complexes are similar to those of bright soils;
— digital numbers of darker zones (groups 13, 14, 15 and 16) such as roads, shadows, roofs
(tiles, slates) are similar to those of dark soils;
— zones of dense vegetation (groups 5, 6, 7, 8, 9 and 11) such as ‘green belts’ of the town
(gardens, row trees etc.) exhibit digital characteristics of vegetation, partly mixed with shadows or soil.
The digital behaviour is characteristic of most urban agglomerations which, depending on their
organisation, combine some pixels having a characteristic of vegetation and a large majority of pixels
with the characteristics of soil. This heterogeneity characterises the digital behaviour of towns (CD
7.13). If the pixels are examined not by ascendant hierarchic classification, but directly in an IRC band,
a still greater diversity of pixels will be obtained.
An urban agglomeration can hence be identified by Its position in the landscape and its shape;
digital heterogeneity confirms this assumption. It is therefore necessary to carry out spatial analysis of
images.

8.2.3 Spatial interpretation

For spatial interpretation of groups derived from ascendant hierarchic classification there should be
several Images, each of which displays a single group. If all the groups are seen together the eye
cannot perceive the spatial organisation of each image segment. The most useful tool for spatial
analysis is an Image in which the group under study appears in colour and all others appear black.
Visual spatial analysis of these groups is based on several factors which aid in analysing different
image zones pertaining to a single group.

H Number of pixels in a group

The number of pixels in a group determines the organisation of various groups. When the groups have
a very small number of pixels, three cases are possible. A fourth case exists when contrarily the
number of pixels is very high.
1. This is a group little represented on the image but which might include many more pixels if a
slightly wider spatial field was investigated. Variation in the pixel frequency of this group may be
very large if the zone of study is changed. Delineation of such a group is fully justified. It is
recommended to expand the field of study for better characterising the group by a larger population
to facilitate better evaluation.
This may be the case of group 29 in the image of Brienne, which is always associated with group
30 or 31. If the image field of view comprises more bare soils on chalk (Chalky Champagne), it
may be considered that this group has a higher pixel frequency.
Such a situation occurs when image borders artificially intersect a unit such as a water body or
when a new landscape is observed on the border.
2. A group may have a small number of pixels since it represents by nature a small surface area.
Such is the case for example, of various types of water in ponds and lakes In the image of
154 Processing of Remote Sensing Data

Brienne. These can be combined into a single thematic unit if that is the objective (see third
objective defined earlier). However, such a unit would evidently have a higher standard deviation
that would affect subsequent processing. Groups 21,22 or 23 correspond to this situation,
3. A group may have a small pixel frequency since the pixels combined are in fact a combination of
different objects; they represent mixels. These only constitute a variant of a larger group (mixels
of combination). On the other hand, the group may have spatial significance. It corresponds to a
transition and its analysis provides an estimate of contrast between two spatial units (boundary
mixels, see Chap. 15).
In the image of Brienne, the pixels of group 3, which are contiguous to those of group 24, represent
this case (the latter corresponds to chlorophyllian crops developed in fields of regular geometric
form). When the zone is surrounded by bare soils, pixels of group 3 are observed on either side of
the boundary between the two. This leads to their interpretation as mixels comprising a mix of
bare soil and highly chlorophyllian vegetation. Groups 17 and 18 are mixels of a mix of water and
very bright soil.
This may also Indicate a variation In land use of soils such as starting growth of some crops which
on a given date have not yet sufficiently covered the soil. The same scene acquired 15 days
earlier would not have shown this group since vegetation coverage was not yet significant (for
example, less than 20%). The same scene acquired 15 days later would show a group of more
numerous pixels corresponding to an ensemble of fields containing the same crop, since the
coverage would be more than 40%. This group may be either combined with the group having the
closest value of digital number or retained as the indicator of a temporal variation. Groups 2 and
32 in the Brienne Image are of this type.
4. Contrarily, when a group has too high a pixel number It may be considered that the number of
groups established by classification Is Insufficient. Groups 12 and 16 which represent bare soils
of medium brightness are of this type. In such a case the number of groups needs to be increased.
For this purpose all other groups are masked and a new ascendant hierarchic classification is
started on the new Image.

H Form of image zones

Shape is one of the elements to which the human eye and brain are most sensitive. Analysis of forms
constitutes an integral part of image processing. Unfortunately, it is not included in most image
processing programs for remote sensing. It Is still in the domain of research. It Is true that the human
brain is still much faster than processing algorithms that are now in full development and are based on
fuzzy logic and on concepts (neural networks, etc.) and techniques of artificial intelligence.
In interpretation, form of image zones substitutes for the Inadequacy of radiometric information.
Forms can be readily analysed from simple values of perimeter and area, which can be directly obtained
from a geographic information system. A greater or smaller consistency of image zones can be derived
from it which, together with size, provides a typology of forms (see Chap. 11, Fig. 11.4).
The results of an ascendant hierarchic classification can be readily interpreted by separating
image zones that are more or less large, are more or less compact and elongated in shape.
One of the common difficulties in interpretation of classification results is that a very large number
of image zones of a single group covers a very small area. On the image of Brienne the 56 pixels of
gravel pit water (group 28) are distributed in more than 30 zones! The same is the case with very bright
soils (group 30) comprising more than 500 zones for 5626 pixels. A group (No. 21 ) with a small number
of pixels (26 pixels) may also consist of only a single zone (pond In southern Brienne). Except in the
last case, it is hardly possible to interpret the forms of very small Image zones.
When image zones are large. It is possible to determine their shapes. The ponds and lakes in
southern Brienne (group 25) correspond to compact and large forms. It is rare to observe this type of
zone on an image after a precise interpretation.
 Processing and Interpretation 155

Very large zones of the order of 1000 pixels (40 ha) for example are observed for group 19 (CD
8.1) but they are highly divided and distorted. As these are included in agricultural plots, they are
inferred to be due to heterogeneities in soil and not to cultivation practices. Hence, similar forms ought
to be observed on Images acquired in different years. They can also be Interpreted as erosional tracks
or indices related to soil depth.
In some cases highly geometric forms of almost all image zones clearly indicate that they belong
to agricultural plots, as in the case of group 24 (CD 8.1).
in some other cases highly tapered, elongated or distorted forms are noticed, as in the case of
group 9 (CD 8.1) which follows the Aube River. It can be concluded that this group is associated with
the river. This may be a land cover type related to alluvial zones.

■ Localisation of image zones

Localisation of Image features can be studied by various criteria, viz., geographic position, dispersion,
synthesis of forms and neighbourhood.

□ Geographic position
Geographic position forms a source of information about Image features of a group under study. First
of all, image regions where features are observed and not observed are determined.Thus bright bare
soils are found only In the north-west, which corresponds to the chalky Champagne region. In the
south-east all the soils are more or less dark (humid Champagne). Position of image features of a
given group can be combined with other sources of information to facilitate interpretation of the group.
For example if clear water is spectrally confused with shadows, then the topographic data Indicate that
the zone is on a northern exposed slope, prompting the conclusion that the group represents shadows.

□ Dispersion
Another analysis Is based on dispersion of image zones of a group. Group 5 comprises a large number
of small image features that are distributed throughout the cultivated zone (CD 8.1) and hence it ought
to correspond to a particular vegetation cover. On the other hand, group 19 a soil of medium brightness
Is not scattered. It is almost absent in valley and forest zones and is mostly situated in the north­
western half of the image (CD 8.1).

□ Synthesis of forms
The eye Is able to Identify on the monitor associations of image zones which indicate a preliminary
hypothesis of interpretation. Piles, patchiness due to uncontrolled urban sprawling and alignments are
determined in this way. Piles are combinations of image features close to one another giving an
impression of compactness.
When an image is enlarged, an image zone earlier considered to be compact may reveal a large
number of pixels pertaining to other groups, giving the Impression of patchiness like a moth-eaten
fabric. A somewhat similar situation Is observed in the forest of Brienne (groups 9 and 10) where the
effect is more apparent when all the groups other than 9 and 10 are depicted in black (CD 8.1).
Alignments not made up of a single digital zone but of many straightline segments are also
determined in this way since the eye interprets them through integration as a straight line. Various
image features in group 11 are not contiguous and correspond to vegetation with shadows, giving the
impression of the Aube course (CD 8.1).
Image shapes of urban agglomerations are interesting. They appear as a unit of piled up pixels if
taken group by group. However, they also exhibit alignments due to roads and patchiness associated
with gardens (CD 7.3).

□ Neighbourhood
Neighbourhood of image features are very important for Interpretation of objects. Evidently, groups 9
and 10 (CD 8.1, band 38) can be interpreted one knowing the other. Both mainly represent riverine
156 Processing of Remote Sensing Data

forests developing in the Aube valley. In the case of crops, groups 1 ,2, 4 and 5 commonly form the
edges of cultivated plots of group 24.
It is also on the basis of unassociated and more distant features that a shadow is interpreted as
that of a cloud. It was computed (see above) that shadows for the two clouds are offset roughly by the
same distance, viz., 85 rows or 1.6 km northwards and 20 columns or 400 m westwards.This evidently
depends on the time (here, 10 h 51 min 13 s) and date (16 September 1996) of acquisition of images,
which determine the solar angle (43.1° solar height with an azimuth of 164°, the angle of incidence
being 0.9°).
When more detailed analysis of image features is desired, structural classification methods need
to be employed (see Chap. 11).

8.2.4 General interpretation

A general Interpretation of results of ascendant hierarchic classification can be carried out using these
elements. Groups are examined one after the other using three sources of information, viz., statistical,
spectral and spatial. They are presented through major thematic groups.
The units obtained from this preliminary ascendant hierarchic classification can be categorised
into the following groups: crops, forest and riverine forests, water, water and soil mix, dark soils,
medium soils, bright soils and vegetation and soil mix.

■ Crops
Chlorophyllian crops are primarily represented by group 24 which corresponds to agricultural plots
clearly identifiable by their linear boundaries. Considering the date of image acquisition, these crops
may be beetroot or potato for example. This group comprises pixels corresponding to the chlorophyll
pole on the two-dimensional histogram (Fig. 8.4). Having a medium pixel frequency, it represents a
typical group that can be clearly identified In the image and on the ground.
This group may be combined with groups 1,2 and 3 which are characterised by low frequencies
and whose pixels are close to those of group 24. On the ‘red-infrared’ two-dimensional histogram

oo
O 24 O
O 24 0
oo 4 4 3 3 2 o 32 oo 31
ooo 4 4 o 30 30
o 6 6 oo 5 o o o 20 O 27 27 o
o 10 7 o o o o O 20 o 19 19 o 28
o 10 7 o o o o O 16 O 19
o 9 8 o oo o 16 o O
o o o 11 o o 12 O O 18 17 17 N u m b e rs in d ic a te
11 o o o o o o G ro u p s o f a s c e n d a n t
11 11 15 14 o o H ie ra rc h ic c la s s ific a tio n
oo 14 13
13 23 22 21

O 26
25 26

R e d b a n d b2

Fig. 8.4: Distribution of 32 groups defined by AHC on a red-infrared plane.

 Processing and Interpretation 157

these groups are situated in the proximity of soil cluster and chlorophyll pole and hence are intermediate
(Fig. 8.4).
This group can also be associated with groups 20 and 32 corresponding to germinating crops.
Group 2, represented by higher values of digital numbers in the red band can be further
differentiated. It is probable that vegetation cover Is less and hence this group comprises mixels
pertaining to a bare soil and a cultivated plot.
The pixels of group 4, with a greater frequency than the preceding, correspond to heterogeneities
in the agricultural plots of group 24. Group 4 can hence be interpreted as a crop that is relatively less
chlorophyllian than that of group 24, due to earlier ripening for example.
Group 5 sometimes represents mixels surrounding the agricultural plots of group 24, sometimes
heterogeneities in agricultural plots and sometimes a crop when it occupies several agricultural plots.
This is explained by its position on the two-dimensional histogram (Fig. 8.4). Groups of this type are
difficult to classify without ground information. They add to the difficulties of classifying all the pixels
from a single image.
The groups of this unit are situated in the north-west half of the image in chalky Champagne and
are absent in valleys. They are also observed a little north and south of Brienne. The unit of ‘crop’
groups Is practically absent in the forest zone in the south.

■ Forests, riverine forests and grasslands

Other chlorophyllian units of the image belong to groups 6, 7, 8 ,9 ,1 0 and 11 which are very close to
one another on the ‘red-infrared’ two-dimensional histogram where they are situated in the centre and
towards the base of the chlorophyllian zone (Fig. 8.4). The frequencies are high in every group.
Group 6 is the most chlorophyllian of this unit and is observed in valleys. Interestingly, this group
marks an ancient meander of the Aube valley (In a square situated between the points defined by ‘row
105-column 225’ and ‘row 125-column 245’). It is also found in large forests in a disseminated manner
except in the south-west where it constitutes compact forest plots. It is also observed In some agricultural
plots. Hence this group does not correspond exclusively to forest zones but also to chlorophyllian
herbaceous communities.
Group 7 Is most often located In valleys or in the south-east part in the vicinity of darker soils. It
ought to represent grasslands.
Group 8 is geographically closely associated with group 7. Based on its digital characteristics It
can be considered as belonging to another group of grasslands.
Group 10 constitutes the major part of forest cover. In the image, it is recognised from the effect
of relief due to different heights of the summits that generate shadows. This is the most chlorophyllian
group constituting forests. It covers a major part of the valleys.
Group 9 is highly scattered but occurs only In valleys and forests. It is situated between groups 10
and 11 on the red-infrared two-dimensional histogram.
Group 11 has the lowest values in the infrared band. It can hence be concluded that it comprises
either chlorophyllian vegetation and shadows or conifers for which the digital numbers In band 3 of
SPOT satellite are always less than those of broad leaves. In the Brienne forest a compact zone is
Interpreted as a conifer plot. Image features that follow streams in the midst of riverine forests are
Interpreted as trees and shadows.
The unit of these groups corresponds to riverine forests developing in secondary valleys and
around the Aube. Its full extension is observed in the south-east part of the Image In humid Champagne
which comprises grasslands and forests.

■ Water
The various types of water can be quickly separated while analysing the classification dendrogram.
They are recognised from their digital characteristics represented by low values in the infrared band
158 Processing of Remote Sensing Data

(band 3). Their position in the red-infrared histogram is distinct (Fig. 8.4). Clear water has the lowest
values (25 and 26) and more or less turbid water (21 and 22) or shallow water (17 and 18) occurs
below the soil cluster.
Very shallow waters in which the chalky bottom can be seen are characterised by high digital
numbers (28). It should also be noted that all the groups, except numbers 25 and 26 have very low
frequencies. These correspond to various ponds.
Ponds or gravel pits in north-western Brienne are associated with various groups, which indicates
that the waters contain highly variable quantities of suspended particles. Groups with low frequencies
cannot be subjected to detailed spatial analysis. At best, one can state which pond contains what type
of water, as in the case of groups 17, 18, 21 and 22.
Group 23 is essentially observed In a gravel pit in north-eastern Brienne. This gravel pit extends
over an area of 20 ha (500 pixels) and exhibits a large variation since it comprises pixels belonging to
6 groups of water and 6 groups of soil.
The otherthree groups, viz., 13,26 and 25, constitute water types that are Increasingly transparent
(Fig. 8.4) and exhibit a spatial organisation. Group 25 corresponds to two reservoirs of the Aube
south-south-west of Brienne. Group 26 is situated on the borders of the latter and in particular In the
concave edges where water Is shallow and where suspended matter could also be higher. Lastly,
pixels of groups 13, 16 and 12 are observed on the shores of reservoirs.
Surprisingly, no difference exists between waters of the two reservoirs.
It is observed that these groups do not facilitate tracing the Aube River. To do this, group 15
should be combined with group 11 but still there would be no satisfactory continuity.
When all these groups are combined Into a single unit and displayed by a single colour, the
envelope encompassing the water bodies fairly forms the boundary of the humid Champagne region.
Consequently, for a total interpretation of the theme ‘water’ in this image groups 13, 15, 17, 18,
22, 23, 25 and 26 ought to be preserved and others masked, and a new ascendant hierarchic
classification attempted on the new image.
As mentioned above, a shadow of clouds is found to be associated with group 25. If this group
was not connected to clouds, it would have been readily mistaken for a pond because of Its shape.
Flence, the possibility of confusion between water bodies and cloud shadows should be anticipated
during classification.

■ Soils
Soils are represented by separate groups, except some which correspond to a mix of soil and vegetation
or clouds. Three major classes can be identified in this soil cluster depending on their brightness.
Four groups show convexity in digital characteristic and hence represent bright soils. These are
groups 31 and 30 whose frequency is several thousand pixels, and group 27 which consists of 17,000
pixels. Group 19 comprising more than 60,000 pixels also belongs to this class.
The classic chorological law is observed in this image relative to soils (see parallelepiped
classification. Chap. 9): brightest soils (31,30) are surrounded by darker soils (27, 19) (CD 8.1).
Groups 14 and 15, dark soils, are at the bottom of the soil cluster (Fig. 8.4). They are mainly
situated on the boundary between humid Champagne and chalky Champagne. They are fairly scattered
in small and compact zones constituting agricultural plots and boundaries of cloud shadows. Group 15
is also observed on the reservoir shore, it may represent moist soil.
Groups 12 and 16 are located between these two extremities. Their frequency is very high and
these two together represent nearly 38% of the image area. Geographically, these are found everywhere.
In chalky Champagne (where group 12 is predominant) as well as in humid Champagne (where group
16 is predominant).They do not have specific shapes and occur as scattered zones as well as compact
agricultural plots.
These two groups situated in the central zone of the two-dimensional histogram (Fig. 8.4) have no
neighbours and ensure ‘filling’ of the image. All the pixels that were not classified satisfactorily in the
 Processing and Interpretation 159

preceding groups are observed in these two groups. For a better understanding of the content of these
groups, only these two should be retained in the colour infrared image by masking all the others, and
a new ascendant hierarchic classification carried out.

8.2.5 Classification quality assessment

The quality of any method of classification, including the ascendant hierarchic, ought to be verified.
The principle of this method is to classify all the pixels in an image. A pixel belongs to only one group
and consequently division between groups is perfect. However, all pixels belonging to a single group
do not have a single digital characteristic and it is not always possible to identify chorological laws that
explain their spatial organisation.
The quality of ascendant hierarchic classification can be assessed by computing the distances of
all the pixels in the centres of groups taken as reference (called nuclei) (CD 8.2). At the end of this
computation a three-dimensional histogram is obtained from which the probability of each pixel belonging
to the group into which It is classified can be ascertained (Fig. 8.5). Groups 12 and 16 contain a very
large number of poorly classified pixels.

O Well-classified pixels

O Acceptably classified pixels

O Poorly classified pixels

m Very poorly classified pixels

Green band b1

Fig. 8.5: Distribution of 32 groups defined by AHC on a green-infrared plane,

indicating quality of classification.

8.2.6 Response to identified objectives

■ General classification of units
To achieve the second objective, viz., general classification of most characteristic units of the region,
the following eight units (CD 8.3) are retained by reorganising the 32 Initial groups (Table 8.5).
The radiometric regroupings carried out during the creation of these units can be seen from
Fig. 8.4.
160 Processing of Rennote Sensing Data

Table 8.5: Reorganisation of 32 AHC groups into 8 units

Unit No. Legend Groups combined Colour Area (km^)

10 Cultivated plots 1,2, 3, 4, 5, 24 Yellow 37

20 Forest and riverine forest 6 , 9, 10, 11 Dark green 98
30 Grasslands 7, 8 Bright green 80
40 Soils and vegetation 1 2 , 16 Bright maroon 175
50 Clear water 13,25,26 Dark blue 15.5
60 Very dark soils 14, 15 Dark maroon 18.5
70 Turbid water 17, 18,21,22,23, 28 Bright blue 0 .2

80 Bright soils 19, 27, 29,30, 31 Beige 36

The groups obtained are of sufficient areal extent to be used for regional analysis. If necessary
the two types of water can be combined.

■ Reconnaissance
To achieve the first objective, viz., preparation of a reconnaissance map from the preliminary
segmentation of the region, it is imperative to proceed up to 32 groups and regroup them as mentioned
above (CD 8.3). Groups 12 and 16 constituting unit 40 (Table 8.5) have a very high frequency— more
than 400,000 pixels (38% of the image)— and hence further analysis is necessary. For this, after
masking (CD 8.4), a new ascendant hierarchic classification of unit 40 alone is carried out in 32
groups. These are then reorganised into 5 new units (Table 8.6 and CD 8.4).

Table 8 .6 : Legend of second AHC of poorly defined groups of the first AHC

Unit number Legend Band colour Area (km^)

1 Very bright soils Beige 44

2 Soils of medium brightness Bright brown 55
3 Sparse chlorophyllian vegetation Green-blue 45
4 Very dark soils Maroon 31
5 Masked groups Black 286

Group 5 corresponds to pixels already classified earlier and masked (units 10,20,30,50, 60,70,
80).
The results of the two classifications should then be integrated into a single image (CD 8.4) by an
appropriate arithmetic combination. Groups of the first classification carry numbers in steps of ten
from 10 to 80. By combining the two bands, unit numbers are obtained as 15, 25, 35, 41,42, 43, 44,
55,65,75 and 85. All the pixels pertaining to units 1 to 4 of the second classification correspond in fact
to only unit 40 of the first classification and hence groups 4 1,4 2,43 and 44 are obtained by addition.
All pixels of unit 5 of the second classification belong to units 10, 20, 30, 50, 60, 70 and 80 of the first.

H Delineation of major thematic zones

The third objective of classification was to determine major zones of a given theme. Our interest was
riverine forests, recognisable from their shape that follows streams. The image classified Into 32 groups
 Processing and Interpretation 161

was geographically investigated for all groups situated in riverine areas (CD 8.5).The following groups
were thus identified:
— 4, 5 and 24 corresponding to crops;
— 6, 7, 8, 9,10 and 11 corresponding to riverine forests and forests (which include practically all
the groups in which chlorophyllian biomass is found);
— group 13 (which represents water) and groups 12,14,15,16 and 19 corresponding to bright to
dark soils.
Thus 15 of the 32 groups determined are situated in the riverine zone, of which 9 are major. This
indicates the heterogeneity of this group and the diversity of these media, which constitutes a preliminary
conclusion. If a new image was generated comprising all these groups, three-quarters of the image
would be covered. Hence there is no unambiguous relationship. All these groups are necessary to
delineate riverine but combinations of these groups clearly describe units other than riverine. The
latter corresponds to a specific spatial organisation of these groups. Other types of classification such
as OASIS are employed to determine these spatial organisations (see Chap. 11).
A similar approach can be followed for the theme ‘chalky soil’. In this case, 11 groups would be
needed, of which the most Important 6 practically correspond to all bare soils and the others to a mix
of chlorophyllian vegetation and bare soil and to fields in which crops still exist.
It is hence possible to characterise various landscapes with the same groups of objects.
Differentiation occurs by the degree of abundance of the groups retained and their relative geographic
positions. This leads to structural investigation of remote sensing images (see Chap. 11).

■ Precise thematic study

The fourth objective is the precise investigation of a feature such as water. It was seen in the second
classification that clear water is readily separated from the rest of the image. It also comprises cloud,
detected only when more groups are identified. A fifth classification is necessary for detecting a new
group corresponding to turbid water. In the sixth classification the latter is subdivided Into two. However,
an eighteenth division is required for obtaining all the groups of water. A large number of subdivisions
is necessary for precise identification of an object. This is achieved by a combination of 9 groups
followed by masking and a new classification.
This detailed analysis led to identification of 21 groups Illustrating the large diversity of pond
waters situated In the Brienne region at the border between chalky and humid Champagne. Three
groups characterise the three reservoirs (CD 8.6). The banks of reservoirs are represented by a large
number of pixel groups, which also attests to their wide heterogeneity and diversity. Cloud shadow is
still mixed with some types of water.
We can proceed further in this way by making more detailed groups. However, we then have to
use bands b1, b2 and b3 directly, employing parallelepiped classification and limiting the analysis only
to the reservoir zone on the Image.

8.2.7 Conclusion
It follows from this study that ascendant hierarchic classification is very useful when no field spectral
data or measurements are available. Computationally this is a very fast method. However, If the software
provides choice, the distance and ultrametric parameter should be chosen carefully since results may
be relatively different. This choice Is not very simple if the objective is not clearly defined and if sufficient
information on statistical methods is not available. Unfortunately, this is frequently the case in programs
In which descriptions of statistical methods are often discrete.
While this method is rapid for computations. Involving a few seconds, much time Is needed for a
correct interpretation. In fact, the statistical data, spatial (geographic) data and reference spectral data
in the interpretation model (established on the basis of reflectance measurements) should be studied
162 Processing of Remote Sensing Data

together. It should be remembered though that a serious and minute study ensures satisfactory results
for all the four objectives proposed earlier.

8.3 CLASSIFICATION BY MOBILE CENTRES (OR

“K-MEANS”)
8.3.1 Principle
The method of classification by mobile centres is based on measurement of distance and an initial
choice of the number of groups or ‘nuclei’. For each, a certain number of pixels are randomly picked to
statistically determine the radiometric centre of gravity of each nucleus (Fig. 8.6A).Then the distance
of every pixel of the Image to each of the nuclei is computed. Each pixel is attributed to the nucleus
closest to it. From this, a kind of segmentation is obtained, constituting as many groups as there are
nuclei. However, in spite of the fact that all the pixels are classified, they are not necessarily well
classified. In fact, there is no certainty that the nuclei drawn at random are the most discriminant.

• Pixels to be classified
® Centres
< 2 ^ Resultant class

Fig. 8 .6 : Scheme of classification by mobile centres. Each of the five iterations (A to E) shows pixels to be classified,
centres of nuclei, results obtained at the end of iteration (zone around the pixels placed in a single group).
 Processing and Interpretation 163

Hence several iterations are carried out. After the first iteration, the centre of gravity is computed
for each nucleus using all the pixels grouped in it, resulting in the definition of a new nucleus (Fig.
8.6B). If the nuclei of the first iteration include pixels far from the centre of a nucleus, the new centre
will be quite different (nucleus x) and hence moves in the second iteration. This method is therefore
known as classification by mobile centres. Other criteria can also be used to characterise new nuclei.
Some nuclei may be removed if their frequency is too small, for example 1 per 10,000. A nucleus may
be split into two if its variance Is very high (Fig. 8.6C: nucleus y was subdivided into two, s and
y; similarly, in the next iteration x was divided into t and u). Lastly, nuclei whose distance is very
small can be combined into one (Fig. 8.6E: nuclei y and z). New nuclei can also be added (Fig. 8.6E,
nucleus v).
All the pixels from these new nuclei are reclassified and the result of the next Iteration obtained.
The result of nth iteration Is then compared with that of n+t. If they are still different, iteration Is
continued; if they are similar, the process is stopped. Various criteria are employed to decide whether
to continue the process or not; most common is the criterion of convergence of the sum of distances
of pixels to the nuclei to which they are allotted or the criterion of the ratio of mean intragroup variance
to the intergroup variance.
It is possible that during an iteration no pixel is classified into one of the nuclei (Fig. 8.6E, nucleus
v).Thus if originally A/groups are required, at the end of classification only AA-1 or even less number
of groups may be obtained.
The choice of number of groups is very important. This is more difficult to decide in this method
than in ascendant hierarchic classification if a precise objective is lacking. One solution is to examine
the results obtained when a successively different number of groups is required.

8.3.2 Method of interpretation

■ Number of iterations
When an image is classified Into a given number of groups, it is necessary to ensure that the number
of iterations applied to the system Is greater than the number of iterations required for converging.
Otherwise, there would be risk of an Incorrect result since convergence might not be reached. The
effects of various iterations can be verified by analysing as earlier the consequences in spatial and
spectral domains using a three-dimensional histogram.
Thus in the Image of Brienne, if a single iteration is employed for two nuclei the result differentiates
bright soils and shallow waters (G2) from the rest of the image (G1). The separation corresponds to a
more or less high value of digital number In band b2 (Fig. 8.7A and CD 8.7). These groups do not
represent the same area (Table 8.7). With two iterations, separation occurs at lower values (Fig. 8.7B
and CD 8.7). With three iterations (Fig. 8.7C and CD 8.7), the boundary between the two groups
becomes less distinct. This represents the case of mobile centres which is difficult to interpret
thematically. A similar situation exists In the case of 4 and 5 iterations in which adjustment modifications
occur. In iteration 6 (Fig. 8.7D and CD 8.7), pixels with low values in the red band (chlorophyllian
vegetation and clear water, very dark soils) are combined Into a group. Between iterations 7 and 10
(last) separations occur between soils more or less covered by vegetation (situated at the centre of
the two-dimensional histogram) and more or less clear waters (Fig, 8.7E and CD 8.7).
Variation of nuclei with the number of iterations can hence be exactly traced in digital space (Fig.
8.7) as well as in geographic space (Table 8.7 and CD 8.7).
If the number of iterations is small, this means convergence is fast. The resultant nuclei have
better a chance of being well-separated radiometrically relative to one another.
164 Processing of Remote Sensing Data

••• •• « • • • • • • € ••••• •• •
•••••••
• •• •• •• •• •• •• •• ••• •• •• ••••••••#
• ••
• ••••« • • ••••
• • •• • ••• ••••••
••• • ••••
• • •• • ••••• • • . •••• •

• • • •

Red band b2 Red band b2 Red band b2

A Iteration 1 B Iteration 2 C Iteration 3

• •••• •• « JD • • • • • •• <
• •• Group 1 pixels
••••••••• T3
•••••••••
• ••
*••••••• • •••••••• • Group 2 pixels
•••••• ••••••
••••• • ••••
•••••• • •••••• •
••

Red band b2
Red band b2
D Iteration 6 E Iteration 10 (last)

Fig. 8.7: Radiometric representation of two groups for different iterations.

Table 8.7: Variation in areas of two groups with number of iterations

Iteration Group 2 area (ha) Percentage Group 1 area (ha) Percentage

1 40,775 88.5 5,305 11.5

2 37,076 80.5 9,004 19.5
3 34,257 74.3 11,823 25.7
6 30,266 65.7 15,814 24.3
1 0 28,387 61.6 17,693 38.4

H Number of groups
The number of iterations necessary for convergence and the number of groups actually Identified can
be analysed by demanding an increasing number of nuclei (Fig. 8.8).
Obviously, a single group is obtained in a single Iteration.This represents the entire Image whose
characteristics correspond to the mean of all the pixels.
From 2 to 4 nuclei, 2 to 4 groups are obtained. As the number of groups increases, vegetation,
water bodies, dark soils and bright soils are gradually differentiated. The number of Iterations is only 5
for 4 groups and they are distinctly isolated. A satisfactory segmentation is achieved with 4 groups.
When 5 groups are required, 11 Iterations are necessary, which is too high for such a small
number of groups. When 6 nuclei are demanded, only 5 groups are identified. With 8 nuclei, 24 iterations
 Processing and Interpretation 165

Fig. 8 .8 : Relationship between number of iterations and number of groups actually identified for a number of
nuclei varying from 1 to 32.

are needed for convergence. Further on, a new stability is observed for 10 groups with 11 Iterations
when 10 to 14 nuclei are demanded. With 10 groups the following features are distinguished;
— cultivated plots,
— riverine forests and forests, not separated,
— grasslands,
— ^four surface states of bare soils,
— ^three types of water.
Subsequently, if the number of nuclei Is Increased the number of iterations increases with the
number of groups. At 22 nuclei the number of iterations remains at 9 wherein a new stability Is observed.
One more step occurs at 26 groups and the next stage at 31 groups. Thus several points of stability
are observed In the number of iterations for certain values of the number of groups.
Using this model, the optimum number of groups to be retained can be chosen. In the present
case, we can take 4,10, 22 or 31 groups (CD 8.7). Choosing an Intermediate number of groups does
not lead to vigorous interpretation.
As In the case of ascendant hierarchic classification, a legend has to be given to every group. If
some groups have very low frequency, they can be combined with others; this depends on the objective
pursued, as seen earlier. The 31 groups of the last classification have been regrouped into 11, similar
to the case of ascendant hierarchic classification and with the same colours (CD 8.7).
Groups situated in the outer part of the two-dimensional histogram are classified quite fast whereas
at the centre of the histogram with similar values between IR and R the pixels are poorly differentiated.
Hence it is necessary to apply the same method as that used in ascendant hierarchic classification,
viz., masking the classified and stable pixels, recomposing an IRC image based on the centre of
three-dimensional histogram and making a new classification.
If the quality is assessed by estimating the probability of correct classification (see maximum-
likelihood classification), it may be concluded that this method of classification is less accurate than
the preceding.

8.3.3 Comparison with ascendant hierarchic classification

Validity of a classification cannot be determined without field data. However, the results of the two
classifications, viz., ascendant hierarchic (AHC into 11 groups) and mobile centres (MCC into 31
166 Processing of Remote Sensing Data

groups, regrouped into 11 units), can be compared. The result (CD 8.7) can be broadly analysed by
comparing the areas of each of the 11 groups (Table 8.8).
Table 8.8: Comparison of results of ascendant hierarchic (AHC) and mobile centres (MCC) classifications

Legend Colour Area (km^) AHC Area (km^) MCC

Clear water Dark blue 15.5 15.2

Turbid water Light blue 0.2 0.2
Very bright soils Creamy 36 31
Bright soils Beige 44 68
Fairly bright soils Light brown 55 51
Dark soils Maroon 31 44
Very dark soils Deep maroon 18.5 17
Sparse vegetation Bluish-green 45 53
Grasslands Light green 80 63
Forest and riverine forests Dark green 98 78
Crops Yellow 37 40

The difference In the area of pixels between the two classifications is about 88 km^, which
represents 19% of the area studied. This difference is not very high but the two classifications cannot
be considered to be giving identical result. It is nevertheless interesting to visually compare the results
of these two classifications. It was observed that major units, viz., the Aube valley and secondary
valleys, large forests, bright soils of chalky Champagne, dark soils of humid Champagne and water
bodies are identified without ambiguity. However, in a more detailed analysis many differences are
observed in the boundaries. Hence a theme-by-theme analysis Is necessary.
Zones classified as water (and shadow) hardly differ between the two classifications. The cultivated
plots are more compact in the classification by mobile centres: no pixels belonging to forest-riverine
forests are observed here. The forest group is less extensive in the mobile-centres classification since
some of the border pixels are classified in the grasslands group. This in fact represents a modification
in the boundary values of digital number between the two groups, values situated at the centre of the
three-dimensional histogram which will have to be specifically reclassified. The forest-riverine forests
zones become “nibbled” with the grasslands group. On the other hand, grasslands are converted in
the MCC to sparse vegetation. In this case also the same boundary effect and the same position of
pixels, viz., centre of the three-dimensional histogram, are observed.
Very bright soils, less significant in the classification by mobile centres, are partly integrated in the
group of ‘bright soils’ of the ascendant hierarchic classification. The same effect extends to other bare
soils; transfers occur from one group to another, which also takes place for pixels located at the centre
of the radiometric histogram.

8.3.4 Conclusion
The method of mobile centres is very fast for computations. Its importance lies in the fact that an
approximate idea of the stability of groups is obtained from the number of iterations. It is especially
useful when neither field data nor spectral characteristics of objects In the area of investigation are
available. However, Its thematic interpretation is not easy and demands much time. It serves as a rapid
method for reconnaissance study of an unknown region, it should be combined with other methods in
order to extend interpretation up to characterisation of the various groups Identified.
 9
Supervised Classification

9.1 PARALLELEPIPED CLASSIFICATION

Parallelepiped classification is based on a radiometric model but not on measurement of distance or
probability. Every pixel can be represented by its radiometric characteristic on a Cartesian diagram
with as many axes as bands. On such a diagram every pixel is situated in an />dimensional hyperspace.
Most often we operate in a three-dimensional space to represent the image using three basic colours.
However, this classification develops on a two-dimensional space, viz., the monitor screen. Hence the
software employed should be able to support three-dimensional information.
If field measured reflectance values are available, they are converted to digital numbers by means
of models that use various satellite parameters. These are used as references on a three-dimensional
histogram from which parallelepiped classification can be constructed. If field radiometric data are not
available, important objects to be interpreted may be defined in the field. Digital numbers of these
features can be determined from their geographic location and used as references on the three-
dimensional histogram. When no field data are available, the classic radiometric model is employed
(Chap. 7, Fig. 7.4).
The classic radiometric model facilitates differentiation between soils, water bodies and
chlorophyllian vegetation. Obviously, depending on the regions under investigation, climatic zones,
etc., several other features can be added to these three basic objects. Examples are clouds and their
shadows, shadows of large topographic elements, bush fire ashes, etc. Towns whose centres often
appear as very dark soils are characterised by a mix of various vegetations, bare soils, water, etc.
Their identification Is based more on structural analysis than textural. It should be remembered that it
Is the surface of objects nearest to the satellite which is interpreted, viz., surface of bare soils or upper
part of canopy in the case of vegetation. In the latter, obviously the tallest vegetation is interpreted.
Interpretation of sub-layers of vegetation is possible only when coverage of higher layers Is incomplete,
as in the case of deciduous forests during fall and winter.

9.1.1 Segmentation of radiometric scatter diagram

Classification amounts to segmentation of the radiometric scatter diagram. Segmentation can be
carried out in several ways. In all cases, the radiometric space Is divided Into volumes by defining
cubes or, more often, rectangular parallelepipeds (hence the name of the method). Classes can thus
be defined for each band. Their intersection leads to correct segmentation if the various objects we
wish to separate do not overlap in the radiometric space (scatter diagram) of the image, something
which rarely occurs.
Division of radiometric scatter diagram can be improved by working directly on three-dimensional
histogram In which each point represents a specific spectral characteristic in hyperspace. The latter is
defined as a parallelepiped that can be isolated, coloured and assigned to the desired class. Thus,
168 Processing of Remote Sensing Data

parallelepipeds with similar spectral characteristics are combined. It Is possible to determine their
spatial relationships since they are coloured over a geographic space constituting the image. The
classification hence becomes an association of elementary parallelepipeds connected by at least a
face, an edge or an apex in three-dimensional space (Fig. 9.1).

Fig. 9.1: Model of parallelepiped classification.

C: Crops; G: Grasslands; F: Forests; Cf: Conifers; VDS: Vegetation and dark soils; DS: Dark soils; MS: Medium
bright soils; VMS: Vegetation and medium bright soils; CS: Bright soils; VWS: Vegetation and white soils; WS: white
soils; CW: Clear water; FCW: Fairly clear water; TW: Turbid water.

9.1.2 Single-band segmentation

The plane corresponding to each band can be segmented. However, this is not always easy if histograms
are not unimodal nor Is it easy to determine the dividing lines for close digital numbers. It is hence
necessary to study the histogram of digital numbers of every band (see Chap. 7, sec. 7.1).

■ Construction of classes
A simple division of band b1 of the image of Bhenne into three units (dividing boundaries at values 58
and 72) clearly reveals three types of objects, viz., broadleaved forests at non-chlorophyllian stage
and water, chlorophylllan vegetation canopies and bare soils (Fig. 9.2 and CD 9.1).
Band b2 of the Brienne image also shows a trimodal histogram. Division can be at values 35 and
62 (Fig. 9.2 and CD 9.1). This image can also be segmented into four parts since an overlapping is
observed in high values around 106. The first part corresponds to forests and water bodies, the second
to cultivated plots, the third to more or less dark bare soils and the fourth to bare soils.
 Processing and Interpretation 169

0 32 64 96 128 160 192 224 256

Fig. 9.2: Divisions of b1, b2 and b3 histograms of the Brienne image.

b1: FW: Forest and water; V: Chlorophyllian vegetation; BS: Bare soils. b2: FW: Forest and water; C: Cultivated
plots; DBS: Dark bare soils; BS: Bare soils. b3: W: Water; DSV: Dark soils and vegetation; CSV: More or less bright
soils and vegetation.

Band b3 of the Brienne image also shows trimodal distribution. The first Is separated from the
second by a very wide gap. Choice of a specific value of digital number as the dividing point between
the two modes is not obvious. Nor can the boundary between the second and the third be readily
traced (Fig. 9.2 and CD 9.1). Values from 9 to 34 represent water, 136 to 198 the most chlorophyllian
plots and for rest of the Image values between 35 and 135 correspond to more or fess bright soils and
vegetation with different degrees of chlorophyllian activity and coverage. One more class can be
Identified around value 83. However, interpretation of two classes is not easy since there is a mix of
dark soils and vegetation on the one hand, and of more or less bright soils and sparse vegetation on
the other.

■ Results
When every band is segmented into 3 or 4 classes, an image can be composed using 3 bands with
3 X 3 X 4 = 36 classes. In fact, only 23 classes are found to contain pixels. Thus redundancy exists
between classes and that top despite taking fewer classes In each band. This of course is due to very
strong correlation between bands 1 and 2 (CD 9.1).
Of the 23 classes, 13 comprise less than 5000 pixels, of which 6 contain even less than 1000
pixels. Classes with very small frequency arise from segmentation of histograms. If the segmentation
limits of the three bands were modified by one value of digital number, the results would change little.
Hence they should not be taken into consideration too formally.
170 Processing of Remote Sensing Data

The major part (99.8%) of the image consists of 10 classes. However, 4 classes are found to
contain 100,000 to 400,000 pixels. They correspond to the central part of the histogram which obviously
combines several objects that cannot be differentiated in the continuum of the histogram. This
demonstrates the limitations of this method based on a single-dimensional approach for deciding the
choice of classes.
This method, very fast, can be used to roughly divide the Image according to various themes, the
latter being very general. In the Image of Brienne the chalky Champagne region is distinguished from
humid Champagne and similarly highly chlorophyllian crops, forests, grasslands and three types of
water bodies are clearly differentiated. Such a method can be useful for a very general analysis.

9.1.3 Multispectral segmentation

The three-dimensional histogram can be directly segmented making use of only the simplified
radiometric model comprising the objects bare soil, chlorophyllian vegetation, water and shadow (Chap.
7, Fig. 7.4).

■ Soil
In the b3-b1 plane of radiometric values, soils are represented by a linear cluster. For the image of
Brienne, the linear soil cluster (LSC) Is defined by the equation:

LSC = 0.56 b3 + 0.83 b1 (Fig. 9.3)

All the pixels situated on both sides of this straight line and pertaining to the cluster can be
coloured as a single unit. If more detailed analysis is desired, the cluster thus obtained can be further
divided into several parts that will determine soil groups, from very bright to very dark (Fig. 9.4 A to C),
such as the three groups in the Brienne image (CD 9.2). This depends on the objective identified. In
the following example of Brienne, six groups, viz., 6, 0, 3 ,1 1 ,1 6 and 22, are used (Fig. 9.5).
Two hundred and forty-four classes of pixels or different types of spectral characteristics exist in
the three-dimensional histogram of the IRC image. Of these, the soil cluster comprises 90, which

@ Dark soils

© Fairly bright soils

® Bright soils
O Non-soils

Soil cluster

Fig. 9.3: Segmentation of soil cluster in b3-b1 histogram.

 Processing and Interpretation 171

Fig. 9.4: Multispectral segmentation of image of Brienne (representation on axes b2 and b3).
A: Bright soils; B: Soils of medium bright; C: Dark soils; D: Clear water; E: Fairly clear water; F: Turbid water; G:
Boundaries of water bodies; H: Crops; I: Grasslands; J: Broadleaved forests; K: Conifers; L: Sparse vegetation; M:
Vegetation and bright soils; N: Vegetation and soils of medium bright; O: Vegetation and dark soils.

indicates that soils represent 37% of spectral types. Geographically these 90 classes of pixels
correspond to 41.5% of the area under investigation (CD 9.3).

■ Water
In the b2~b3 plane the radiometric values of different types of water always occur below those of the
soil cluster (CD 9.3). The digital number value of the near infrared band is primarily related to clarity or
depth of water layer. Different groups of water can hence be defined as a function of this value (Fig. 9.4
D to F).
Cloud shadows are very often confused with clear or deep water. A very accurate analysis is
needed to differentiate them. In this case, shadows and clear water overlap each other.
Pixels are also observed between soils and water. These are mixed pixels (mixels) in the boundaries
of water bodies (Fig. 9.4 G).

■ Vegetation
Radiometric values of chlorophyllian vegetation are situated above the soil cluster towards higher
values of the near infrared band (CD 9.3). Two units can be differentiated, one for complete vegetation
cover or at least more than 40% and the other for sparse coverage.
The first unit can be further classified according to the magnitude of digital number in the infrared
band. The following groups can thus be distinguished:
— Vegetation with high chlorophyllian activity and large biomass characterised by high digital
numbers in Infrared and low In red (Fig. 9.4 H). In the Brienne example these are crops of beetroot,
potato or maize (as the image pertains to the month of September).
— Vegetation consisting of a mix of more or less chlorophyllian plants and low biomass such as
grasslands and partly broadleaved trees (Fig. 9 .4 1).
172 Processing of Remote Sensing Data

— Broadleaved forests comprising highly chlorophyllian tall trees and associated shadows, mixed
within a single pixel and giving rise to a mixel. This results in a smaller value In near infrared and a
generally low value in band b2 (red) (Fig. 9.4 J).
— Coniferous forests whose canopy structure leads to smaller values in near infrared (Fig. 9.4 K).
The second unit Is made up of pixels whose spectral characteristics Indicate a mix of more or less
chlorophyllian vegetation and bare soil. Several groups are Identifiable:
— One group comprising both highly chlorophyllian vegetation and sparse covers. The digital
numbers are equally high in the near infrared and the red bands due to the contribution of soils for
which the values are greater in band b2 (Fig. 9.4 L).
— A series of groups situated close to the soil cluster, which indicate a mix of spectral characteristics
of chlorophyllian vegetation and soils, the latter exerting a strong Influence. Several subgroups can be
identified depending on whether the vegetation is situated on bright soils or darker ones (Fig. 9.4 M, N
and O).

9.1.4 Chorological segmentation

A three-dimensional histogram can be segmented using the radiometric model described above,
together with reasoning based on spatial relationships between pixels belonging to a given group, in
other words on the basis of chorological laws. More precise classes are defined which make radiometric
relationships compatible with spatial relationships.
For tracing boundaries between soil groups that differ in luminance, we can use relationships
between radiometric vicinity (on three-dimensional histogram) and geographic neighbourhood. A spatial
organisation of soils is discovered, represented by a geographic transition from ‘white’ soils to very
bright soils, bright soils and fairly bright soils. This variation Is explained by the fact that the soils are on
chalk (of Champagne) and their colour varies inversely with thickness. Thus they vary from LITHOSOLS
to RENDOSOLS and CALCISOLS (Baize and Girard Référentiel pédologique, 1995). Dark soils
correspond to decarbonatisatlon clays or clayey and moist soils pertaining to the humid Champagne
region. Flence there Is a correlation between soils and their surface states. These chorological laws of
soil can be fairly well determined, based in particular on images in which soils are bare. Such soli
organisations have been detected in other regions also, Caux county for example (Burlot, 1995).
Plant canopies in agricultural plots reveal soil heterogeneities, viz., more or less thick soils,
occurrence of microtalwegs, etc. Different water bodies have different spectral characteristics which
ought to be related to their mode of utilisation. In the ‘reservoir-ponds’ of the Aube various spectral
units parallel to the border are observed on a flank, which are related to depth and nature of the
bottom, whereas the water In the central part exhibits a single spectral characteristic. However,
differences are observed between various ponds, which may be due to differences In their management
and utilisation.
It was decided to retain only three groups for agricultural plots corresponding to different
chlorophyllian states. In fact, if they pertained to different crops, the three classes would have been
distributed according to plots, but they are often inside a single plot. This hence corresponds to Intraplot
heterogeneities associated with differences In soils and not due to agricultural practices. Forms of
these intraplot heterogeneities. In fact, resemble microtalwegs. This is often confirmed since a given
heterogeneity extends across several plots. Irrespective of whether they are crops or bare soils.
A total of 27 groups were chosen in the image of Brienne (CD 9.4).

9.1.5 Spectral characteristics of various groups

Results of this classification can be expressed by digital characteristics of each group (Fig. 9.5). However,
it should be remembered that this procedure cannot be used rigorously unless the position of objects
 Processing and Interpretation 173

Fig. 9.5: Digital characteristics of 27 units obtained from parallelepiped classification by chorological analysis.

is studied band by band (also see Chap. 8, Figs. 8.1 and 8.2). If position of objects in two- or three-
dimensional histograms is studied, relative positions with respect to one another should be used.
These objects can be marked by a combination of three values of digital numbers that characterise
them fora given image but one should be cautious.Thus, it was observed In the Image around Brienne
that the curves representing the objects In the three bands (Fig. 9.5, Chap. 8, Figs. 8.1 and 8.2) do not
directly correspond to those of reflectance (see Chaps. 4, 23 and 25).
In fact, bright soils which have a continuous, convex and ascendant reflectance curve exhibit a
convex but descendant curve of digital numbers. Water likewise has no zero values in the near infrared
band. The form of digital number curves of groups corresponding to vegetation covers, which Is close
to that of their spectral reflectance curve, cannot be considered a general case.
Six soil groups are clearly distinguished (Fig. 9.5, soils): white soils (22), very bright soils (16),
bright soils (11), medium bright soils (3), dark soils (0) and very dark soils (6). All the digital characteristic
curves are convex except for the very dark soils, which are concave.
Of the six types of water (Fig. 9.5, water), very clear (12), clear (8) and fairly clear (15 and 19)
waters are similar to one another. Slightly turbid (23) and very turbid (26) waters differ from others. If
high precision of classification is not required, only two or three groups can be kept.
All the six groups of chlorophyllian vegetation (Fig. 9.5, vegetation) show small values of digital
numbers in band 2, representing absorption of chlorophyll. They are distributed in two groups
differentiated by band 1 values. Digital numbers of band 3 indicate a regular decrease from one group
to another. The first group comprises highly chlorophyllian cultivated plots (14 and 13) and permanent
174 Processing of Remote Sensing Data

grasslands or forest clearings (10). The second group is represented by a complex combination of
less chlorophyllian permanent grasslands and trees (5), forests with shadows (2) and conifers and
plants associated with shadows and water, such as on river banks (17).
The other eight groups consist of vegetation and soil mixes (Fig. 9.5, vegetation and soils).
Three groups, 4, 18 and 21, exhibit characteristics of a plant canopy. Group 21 is relatively the
most chlorophyllian and densest since the difference between b3 and b2 is the highest. This type is
always associated with cultivated plots either on the borders or throughout the area. It corresponds to
heterogeneity in a crop that is less chlorophyllian because in an advanced stage of development or
less dense due to boundary effects or soil differences. Type 18 is situated on bright soil and type 4
corresponds to tall trees in forests or heterogeneous grasslands with green and dry plants.
The two types 1 and 20 are intermediate between soils and vegetation: convexity of spectral
characteristic of soils (Fig. 9.5, soils) is compensated by concavity of vegetation spectral curves (Fig.
9.5, vegetation), resulting in a straight line. Other types are more or less influenced by plant canopy
but types 20 and 25 correspond to bright soils, while types 7, 1 and 9 correspond to Increasingly
darker soils. Type 7 comprises more chlorophyllian or denser vegetation than type 1.
Type 24 corresponding to cloud shadows is not shown in Fig. 9.5.

9.1.6 Statistics of various groups

If all the 27 groups covered equal areas, there would be an average of 42,666 pixels per group (CD
9.4). This is used as reference for statistical analysis.
Groups with the smallest surface areas, excluding cloud shadows, represent various water bodies
and conifers, which constitutes a characteristic feature of the region under investigation.
Next in size are white and very bright soils and dark soils. These groups can evidently be combined
with their neighbours and hence classified into three or four units instead of six. This means that bright
soils are relatively less numerous than others in the image. However, this Is specific to the portion of
the Image studied: if the portion farther west, in chalky Champagne Is analysed, coverage by bright
soils would be greater and in the portion farther east in humid Champagne almost completely absent.
This shows that for correct Interpretation well-defined landscape units should be used (see Chap. 18)
if statistical results are to be analysed in terms of areas. However, this also indicates that with fairly
simple land use types, landscape units can be readily defined. In the present case, the chalky
Champagne region can be characterised by an association of bright soils and the humid Champagne
region by darker soils.
The two groups representing agricultural plots not yet covered by chlorophyllian crops in September,
occupy less area. Obviously, processing of an image of the same region acquired a little later would
show these two as a single group. The influence of date of acquisition on differentiation of groups can
be assessed from the LANDSATTM images of April and May (CD 7.12). This shows the Importance of
the choice of acquisition date depending on the objective of investigation (see Chap. 16).
Lastly, several groups comprising bare soils with dense or sparse vegetation cover, mature or
seedling, mostly have small areal extents. Obviously they vary depending on the climate of subsequent
days: for some vegetation will dry and for others develop. A new class could be identified but it would
represent a different feature. It is certain that in an Image acquired a little later, plots already sown for
the next year can be readily separated from plots which will be tilled later and are not yet sown.

9.1.7 Visual interpretation

The preceding interpretation is facilitated by choice of established colours according to the functions
identified. Large spatial groupings perceived by colour associations are thus readily Identified. If colours
are distributed to various groups randomly, spatial organisation is no longer easy to interpret (CD 9.5).
 Processing and Interpretation 175

In order to discover or evaluate chorological laws that define spatial organisation of a region,
correlation should exist between semantic characterisation of objects and their geographic distribution.
Study of reciprocal spatial positions of different groups allows refinement of interpretation.
The number of groups can be increased or decreased according to the objective of classification.
The result of 27 classes may be considered too precise for a preliminary investigation of the region.
The number of classes may hence be reduced to 5, viz., bright soils, dark soils, vegetation, water and
shadows (CD 9.5).
Contrarily, for accurate identification of a theme a group may be taken, isolated by masking and
again subjected to parallelepiped classification.
Thus the groups chosen by parallelepiped classification can be precisely defined by means of
radiometric differentiation and analysis of distribution and spatial organisation, i.e., by constructing a
set of chorological laws.

9.1.8 Quality of parallelepiped classification

Parallelepiped classification is based on radiometric models and eventually chorological laws. Three
levels of models were identified.
The simplest is based on segmentation of each of the three bands before combining them. Results
are fairly general and may suffice if the aim is to identify major groups. However, this method cannot
be employed unless histograms of each band can be segmented with no serious errors. Such is not
always the case. If the processing tools available are limited in performance, the single-band method
is still useful. Multispectral segmentation is preferable, however.
The model based on multispectral segmentation is more efficient than the preceding but assumes
knowledge of the principal spectral characteristics of objects of the region under study. The latter is not
always possible, for example when this method Is employed just before conducting a detailed field
study of the region. The results will hence depend on how effective the model Is for the zone under
analysis. Care should be taken to choose the groups according to adaptability level of the radiometric
model to the conditions of the region. In a poorly known area It is desirable to choose a relatively
limited number of groups.
The model based on radiometric and chorological segmentation Is evidently the most efficient,
but It assumes that spectral behaviour of major objects of the image is known and that certain
chorological laws can be applied. At least a general knowledge of the region is necessary in the
absence of accurate information. The more certain we are of the chorological laws, the more precise
analysis. This certainly is the most suitable method for a thematic specialist having experience in
remote sensing or in the region, since he (she) can derive a combined advantage.This method provides
the best use of expertise. For the expert, it Is the most reliable, fast and least expensive.
It is possible to assess the quality of parallelepiped classification. As in ascendant hierarchic
classification (Chap. 8, Fig. 8.5), a modal radiometric value is defined for each class and compared
with the values of all pixels belonging to it. Thus pixels classified well or poorly are determined.

9.1.9 Conclusion
Parallelepiped classification ought to be developed based on a correct radiometric model applicable
to the region under study. As it is totally supervised, the analyst can build up his (her) interpretation
step by step. In this respect, it is very efficient for the expert. It also aids In rapidly defining a classification
based on a conventional radiometric model. Lastly, it enables classification of only those groups which
correspond to the objective of investigation and grouping of all other components of the image that are
not useful for the objective into a single class (for example, coloured black).
176 Processing of Remote Sensing Data

9.2 MAXIMUM-LIKELIHOOD CLASSIFICATION

9.2.1 Probabilistic spectral behaviour
Spectral behaviour of an object is neither unique nor specific and changes with time. This results in an
uncertainty in discrimination level. In fact, remote sensing data always contain random errors which
reduce contrasts between objects. The position of an object cannot be readily marked on an image
and locating on the image a region known on the ground is also not easy and may lead to errors (see
Chap. 10). Finally, pixel size and positioning (variable depending on the satellite) make it difficult to
avoid mixels (see Chap. 13) which, by definition, are poorly discriminated spectrally. Consequently,
spectral behaviour of an object varies depending on images and changes with time. It is hence not
possible to correlate the n values of digital numbers (n being the number of bands) assigned to a pixel
with a definite type of object. In fact, if it is considered that a signature Is characteristic of an object and
that it does not vary in space and time, an object cannot be said to have a spectral signature at ail
since digital numbers vary in time and space.
We can, however, consider that a pixel has a ‘spectral signature’ which is often represented in
software programs (Including TeraVue) by a set of lines that indicate a digital characteristic. Contrarily,
an object which is represented by a pixel population, has no spectral signature. It is dangerous to
speak of a spectral signature for corn, for example, since its spectral behaviour varies depending on
the region in which it is cultivated and the season.
The probabilistic approach. In fact, replaces the spectral signature of a pixel, described by a
simple set of lines, by spectral behaviour described by a probability distribution centred on a mean
spectral characteristic.
So it is interesting to study spectral characterisation of objects in terms of probabilities.

9.2.2 Principles of classification

The maximum-likelihood classification is thus based on probabilistic methods. We compute for each
pixel its probability to be assigned to a given class rather than to another. The law of affectation thus
derived enables minimisation of risk of error by a better utilisation of probabilities pertaining to one
class or another. In this method a pixel belongs to a class with a certain probability and not In a binary
manner, viz., ‘belongs'or ‘does not belong', as in the preceding methods of classification.
Consequently, in order to determine these probabilities known zones should be chosen as a
priori references and the final result discussed on the basis of appurtenance of a given probability.

H Bayes decision rule and maximum-likelihood rule

When reference zones are chosen in an image, digital number histograms of pixels contained in each
zone can be computed. Each histogram defines a density of conditional probability of a pixel p pertaining
to a class C/.This probability is denoted by P(p/C/).W e can also determine the probability of occurrence
of a pixel of a class Cj, indicated by P{Cj). However, this requires that the pixel frequency of class Cj
be known before undertaking classification.
If during classification pixel p is associated with class Cl when it actually belongs to class C2, an
affectation error occurs. The rule of affectation of a pixel p to a class C is known as maximum likelihood
when it minimises the mean error of affectation.
When the error of affectation does not depend on class C, the decision rule is called ‘Bayesian’
and may be described as follows:
If Ci is the class of appurtenance of pixel p and Cj every other class,
p belongs to class Ci, is equivalent to P(p/Ci) ■ P(Ci) > P(p/CJ) • P(Cj).
 Processing and Interpretation 177

!Gaussian parametrisation
The conditional probabilities of appurtenance based on the histogram of a reference zone are called
empirical probabilities. Their determination is restricted to sampling of classes present in reference
zones. Obviously an attempt is made to ensure that the geographic extent of the final classes far
exceeds that of reference zones. Hence it is sometimes necessary to determine the parameters of
these classes by assuming that every distribution corresponding to the classes under investigation is
Gaussian. The probability P(p/Ci) in this case is defined by the mean and variance-covariance matrix
of pixels of each class. The latter are estimated from the values computed for reference zones.
In remote sensing it Is often assumed that every class (defined by reference zones) is a priori
équiprobable, i.e.,

P(C1) = P(C2) = ... = P(C/).

The Gaussian discrimination function FDi{p), necessitating computation of variance-covariance

matrix, can thus be calculated and simplifies the decision rule expressed as:
p belongs to class Ci, is equivalent to FDi (p) > FDj (p) irrespective of / and j.
Hence it is necessary to choose reference zones in such a way that they represent the entire
object and that the number of pixels of such a zone suffices to correctly estimate the parameters of the
Gaussian statistical population which it is assumed to represent

9.2.3 Rejection threshold

The discrimination function used allots every pixel to one of the classes defined by the reference
zones. This may be accepted if the objective of classification is to classify all the pixels even when
poorly categorised. However, the importance of this method lies In detecting the validity of classification
of every pixel into the class to which It is assigned.
It is hence natural to create a rejection class which receives pixels that cannot be assigned to one
of the classes defined by the reference zones. For this, a rejection threshold Is defined which corresponds
to the admissible lower limit of conditional probability of a pixel p pertaining to the class Cj. Pixels
below this threshold will be assigned to a rejection class.
The rejection threshold can be expressed as percentage. While the threshold can be readily
determined In the case of one-dimensional analysis, it is much more difficult for multidimensional
situations, the most common In remote sensing. Based on the analytical expression of the Gaussian
discrimination function PD/(p), Monget (1986) showed that a p% rejection threshold, fixed for Gaussian
densities of conditional probability, is manifested by a threshold of p% applied only once on the %^-law
with n degrees of freedom (n being the number of bands used for classification). Hence for ‘Bayesian'
classification this threshold should also be defined.

9.2.4 Classification operations

The maximum-likelihood classification requires: 1— a choice of reference zones or nuclei, 2— ^statistical
and radiometric analyses of these nuclei, 3— study of separability table, 4— classification proper, with
accurate evaluation of classes by probability study and 5- analysis of performance table after
classification is completed.

■ Preparation of nuclei
Nuclei are used to define populations based on which classification can be carried out. A large
number of pixels and a homogeneous population are necessary for obtaining the best possible
178 Processing of Remote Sensing Data

approximation to a Gaussian law. Every nucleus can be defined by several zones separated from one
another.
Each nucleus ought to have a frequency of more than 200 pixels, which is not difficult to get in a
remote sensing image with millions of pixels, except in some cases where the theme is very precise.
It is preferable to choose sample zones in several parts of the Image so as to avoid the situation
wherein the nucleus represents only one type of region. If the image features to be traced are very
small, the software program used should enable their delineation through zooming and for this the
zones should be multiplied.
It Is also preferable that the population chosen be very homogeneous and defined correctly. The
image features retained should be largest possible and in any case the most compact. The zones
should also be delineated accurately If the boundaries are sinuous. This is very difficult In the case of
clouds and linear features such as roads and watercourses.
As errors are always possible when delineation of image zones is difficult, facility should exist In
the software to return to the traces already drawn and hence to erase easily.
Nuclei are defined geographically. This assumes knowledge that such a place exists, established
by even minimal field study whereby major objects of special attention can be identified on the ground.
However, they may have been forgotten or not properly evaluated due to the perspective effect (see
Chap. 18, Fig. 18.1). Definition of nuclei hence depends on the geographic approach. It is useful to
take the help of existing thematic maps for a better definition of image zones comprising the nuclei. In
fact, in most cases field radiometric measurements are not made.
Another approach is based on combining field observations with radiometric measurements and
constructing a local radiometric model that facilitates choice of nuclei, since locations Identified in the
field and radiometric differentiation are used simultaneously.
In all cases, it is recommended to visually interpret colours observed In the infrared band for
delineating image zones of the nuclei. One can always refer to the interpretation model of IRC images
(Chap. 7, Fig. 7.4).
It is common in a real study to use several types of classification for a given object. Thus ascendant
hierarchic classification or classification by mobile centres can be employed to chose the nuclei. It is
better to construct more nuclei than necessary and regroup them, rather than taking a small number
of heterogeneous nuclei.
Lastly, the visual or thematic evaluation of objects should be followed by their spectral
characterisation based at least on their digital numbers in the image under study. This statistical study
is mandatory.

M Statistical and radiometric analysis of nuclei

For correct interpretation of the nuclei selected, which constitute the basis for Bayesian classification,
statistical analysis is necessary. Means, standard deviations and variance-covariance matrix of the
bands used are essential for computations in classification (see above). Consequently, access to
statistical programs during the process of interpretation is imperative.
It Is possible to reconstruct the digital characteristic of each nucleus from the mean values of
digital numbers of each band through a set of lines, as mentioned in the example of parallelepiped
classification (Fig. 9.5) or ascendant hierarchic classification (Chap. 8, Figs. 8.2 and 8.3).
Thus, in the 20 nuclei (characterised by 91 elementary image zones) chosen for the Image of
Brienne (CD 9.6), the gradual variation from a convex curve to a concave one represents soils with
different levels of luminance. In some cases, the standard deviations are not distinct. The standard
deviation of nucleus ‘white soil’ Is slightly higher relative to others; this can be corrected by removing
the darkest parts of the image zones of this nucleus, which reduces the heterogeneity.
Various types of water exhibit a concave curve of digital characteristic. They are differentiated
mainly from digital number values of band 1. The most transparent water has the lowest value while
 Processing and Interpretation 179

turbid waters containing suspended matter or shallow waters, or waters located on a chalky surface
are represented by higher values.
Vegetation Is characterised by low values of digital number in band 2. This value is very low for
chlorophyllian crops growing in September in this region (beetroot for example), a little less for forests
and riverine forests, and becomes higher for ripe crops or for sparse vegetation. The latter two cases
represent mixes of soil and vegetation. It may be noted that the spectral characteristics for the nuclei
‘forest-riverine forest’ and ‘dark forest’ are very similar. These should be combined and the zone of
conifers and shadows defined by a new, more representative nucleus.

■ Separability
A separability table shows the degree of differentiation of pixels used for defining the nuclei. In this
table, nuclei are given in rows and the thematic groups defined by various nuclei listed in columns. The
table is read column by column. Each column, defined by a number that refers to the name of the
nucleus and its pixel frequency, shows the number of pixels of nuclei which have values identical to
pixels pertaining to another nucleus. This corresponds to pixels for which histograms of different nuclei
overlap.
Thus, for Brienne (Table 9.1 ), 1,864 pixels constitute the nucleus (no. 3) ‘bright soil’. 1,730 pixels
are common with the nucleus ‘very bright soil’, 7 with ‘bright soil’, 62 with ‘fairly bright soil’, 1,603 with
‘sparse vegetation’, 1,682 with ‘dry crops’, 332 with ‘riverine forests and forests’ and 159 with ‘clouds’.
In total, 5,575 pixels have at least a dual appurtenance; as there are only 1,864 pixels, many pixels
have multiple identification.
This table hence gives an approximate idea about the accuracy of nuclei selection. If pixels of a
nucleus Interfere with no other nucleus (in the corresponding column there are only blanks or zeros),
Its histogram is perfectly distinct from the histograms of other nuclei. Separability is then said to be
perfect.
In the image of Brienne, a nucleus such as no. 10 (crop) is well separated from others. In fact, only
a very few pixels (151), compared to the number of pixels used for defining it (frequency: 3,593), are
common to other nuclei. A similar situation is observed for nucleus 15 (very clear water) which has
only 20 pixels common with nucleus 16 (clear water), for a frequency of 3,554 pixels.
Contrarily, a nucleus is less separable if a large number of pixels used to define it are common to
other nuclei. Such is the case especially with nuclei 13 (riverine forests and forests), 20 (clouds), 8
(sparse vegetation), 9 (dry crops), 11 (green grasslands) and 12 (dry grasslands).
The separability table shows that in the case of bare soils, many pixels simultaneously belong to
different soil types: brighter or less bright. Hence it is necessary to improve the test zones of these
different nuclei so that they become better separated.
For the nucleus ‘cloud’, of 171 pixels 130 are not separable from the nucleus ‘white soil’, 139 from
‘dry crop’ and 21 from ‘very bright soil’. In total, the 171 pixels can be distributed In 7 nuclei, which
represents 301 possible cases.
Many pixels indicate a dual or triple appurtenance. This shows that the nucleus ‘clouds’ is not
satisfactory. Therefore It is not possible to identify clouds by their radiometric characteristics. This is a
common case in images (CD 4.2).
It can be seen that total confusion exists between the nuclei ‘green grasslands’ (no. 11) and ‘dry
grasslands’ (no. 12). The curves defining spectral characteristics from digital numbers are practically
parallel. They also show confusion with ‘riverine forests and forests’ (no. 13) and partly with ‘sparse
vegetation’ (no. 8) and ‘dark soil’ (no. 5). In the latter case, it is necessary to combine these two nuclei
and take them as only one.
One way of deciding which of the two nuclei is to be removed is to mask all features other than
these two In the infrared image. It can be seen that the nucleus of green grasslands lies in the centre
of the three-dimensional histogram, which is significant since there are few nuclei in the centre. It does
not confound with various nuclei pertaining to crops. Contrarily, the nucleus of ‘dry grasslands’ is much
 Table 9.1: Separability between 2 0 nuclei chosen for maximum-likelihood classification

Nuclei 1 2 3 4 5 6 7 8 9 1 0 11 1 2 13 14 15 16 17 18 19

Pixel frequency 1371 1483 1864 1755 1751 885 848 1667 1478 3593 2691 1419 3703 4433 3554 1490 1471 383 433
1 White soil 603 7 2 714
2 Very bright soil 49 1730 1 2 1 377 931 1 2 4 1

3 Bright soil 1 2 1157 52 64 1

4 Fairly bright soil 1 62 638 544 27 2 2 2

5 Fairly dark soil 1 1306 236 3 393 2 1421 270 133 1

6 Dark soil 923 607 91 1 35 404 1030 79 16

7 Very dark soil 89 855 1 1 6 50 161 109
8 Sparse
vegetation 7 1031 1603 1395 768 205 2 2 0 2 2 357 1004 60
9 Dry crops 81 1682 382
1 2 0 0 60 24 1 574 3 3 3
1 0 Crops 1 6 2 205 5 44
11 Green
grasslands 1 1 0 615 2 0 1 1034 1 1380 2289 4275 2 11

12 Dry grasslands 1 1 158 44 2 0 891 2 2640 1527 2

13 Riverine forests
and forests 397 332 2 133 276 726 621 9 108 2677 1405 4433 1 2 116
14 Dark forests 8 1244 6

15 Very clear water 1480

16 Clearwater 2 0 935
17 Fairly clear water 51 361 1 1 0 160 23
18 Turbid water 1 1 555 75 83 29
19 Cloud shadows 52 489 3 89 82 288
20 Clouds 171 368 159 204 326 842 1 4 281
Number of pixels 320 4761 5575 3312 3712 1769 2 2 1 0 4582 3001 151 8080 5223 5717 8790 2 0 1490 1253 171 281
with multiple
appurtenances
 Processing and Interpretation 181

more similar to those of soils. Consequently, all image zones that define the nucleus of ‘dry grasslands’
are simply removed. Pixels belonging to this nucleus are redistributed between the nucleus of ‘green
grasslands’ and various nuclei of soils.
A similar analysis is followed for various nuclei that are poorly classified.

■ Analysis of conditional probabilities of appurtenance

Every pixel has a greater or smaller probability of belonging to a statistical population defined by the
pixels of each nucleus. To avoid getting too many poorly classified pixels in the thematic groups of
interest, a probability threshold of rejection is defined. Any pixel classified with a probability lower than
the threshold is eliminated. It Is possible to make the value of threshold variable. The consequences
are examined on the three-dimensional histogram and the image.

□ Analysis of three-dimensional histogram giving classification probabilities

The three-dimensional histogram represents the probability of appurtenance of every pixel to a given
nucleus. The probabilities of appurtenance are coloured in the decreasing order from white to black
(Fig. 9.6) or In TeraVue, from white to black, passing through yellow, orange, red and maroon. Pixels
with probabilities lower than the threshold defined by %^-parameter are not indicated and are rejected.

o ^ oo
Green band b1
Rejection threshoid=0

O Very well-classified pixels Poorly classified pixels

(§)
^ Well-classified pixels Very poorly classified pixels
Acceptably classified pixels Rejected pixels

Fig. 9.6: Histograms for bands b1 and b3 giving probabilities of appurtenance of pixels to nuclei,
with a rejection threshold of 0 and 1 0 %.

□ Rejects
The rejection threshold can be varied from 0 to 100%. When the threshold is 0, almost all the points
are classified and most of them poorly, hence the large number of black or dark pixels in Fig. 9.6.
When the threshold is higher (10% in Fig. 9.6), the number of rejected pixels is very large and those
classified have a much higher probability.
In practice, the probability level of rejection should not exceed a few per cent.
182 Processing of Remote Sensing Data

□ Analysis of probability images

In a probability image the colour of every pixel is dependent on its probability pertaining to one or the
other nucleus. The whiter or yellower a pixel, the greater its probability, and hence the better it is
classified (CD 9.6). Thus, for a given rejection level, the accuracy of classification of every pixel in the
image can be evaluated. This information can also be obtained pixel by pixel. The latter enables
identification of pixels assigned to a given nucleus and determination of whether the pixels situated at
a given place are well classified.
We can also determine for a given area (a cultivated or a forest zone for example), at what
rejection threshold a pixel becomes black, i.e., is rejected. The higher the rejection threshold, the
better it is classified.
It can be readily seen that there are regions that are poorly classified or unclassified, using the
nuclei defined. It Is hence useful to add some new nuclei to represent them. For this, digital spectra of
the unclassified zones in the image are investigated. A new nucleus can be defined by choosing it
geographically at the place indicated (in black) in the probability Image.

□ Example of image of Brienne

‘Cloud’ is characteristic of a heterogeneous nucleus. During classification with a preliminary set of
nuclei it contains 11,790 pixels, which largely exceed the surface area of clouds present in the image.
The pixels classified as cloud belong to the nucleus of white soil for the brightest and to the nucleus of
dry crops for the darkest. Study of the probability histogram indicates that in the image there are many
points which can be allotted to the nucleus ‘cloud’ (this can also be seen by reading the line ‘cloud’ of
Table 9.1, the separability table). Hence, this nucleus is heterogeneous and overlaps on the neighbouring
nuclei. It introduces confusion during allotment of nuclei to various other nuclei. It is better to remove
this nucleus in order to reduce confusion. As the clouds observed In the image do not exhibit a specific
spectra! characteristic it is not possible to specifically Identify them.
When all the nuclei and the probability image are analysed systematically with a rejection threshold
of 0 (Fig. 9.6), it can be seen that unclassified or poorly classified pixels exist in the centre of the three-
dimensional histogram. On the other hand, unclassified zones are distinguished in the image (coloured
black— CD 9.6). This Indicates that the nuclei retained do not suffice to classify them. If some nuclei
are added in the unclassified segments, by choosing the zones that characterise them, classification
can certainly be Improved.
The nucleus ‘dry crops’ (no. 9 in Table 9.1) confounds with too many other nuclei. This nucleus is
redefined by associating it with the nucleus ‘crops’. Hence, a mask was applied isolating these two
nuclei in the band of NDVI (CD 9.6). The histogram of the latter band is distinctly found to be trimodal
(Fig. 9.7). This nucleus can hence be segmented into three units; the earlier nuclei ‘crops’ and ‘dry
crops’, as well as a new one ‘little chlorophyllian crops’.

® Quality in nuclei selection

The probability image, the separability table and the spectral characteristics of every class constitute
tools for evaluating the quality of classification. They aid in selecting the nuclei and Improving them
through several iterations. At the end of an iteration several modifications of nuclei may exist
corresponding to the following:
1. The chosen nuclei exhibit good quality and are retained. However, modifications of other nuclei
may lead to modification of the quality of the first nuclei if the new nuclei created are quite close
to the former. The nucleus ‘crop’ is a typical example.
2. The nuclei are close to one another but spectral characteristics are quite different. They are
retained for several reasons, such as thematic. Their quality needs to be re-examined by means
of a performance table at the end of classification (see later). This is the case of various levels of
bright of bare soils.
 Processing and Interpretation 183

Digital number

Fig. 9.7: Values of NDVI for the two nuclei ‘crops’ and ‘dry crops’, which are regrouped into three classes.

3. The nuclei indicate poor separability and similar spectral characteristics. They may be preserved
if necessary but it is prudent to use new image features In defining them to achieve the best
possible semantic separation. Such is the case of the nuclei ‘dry crops’ and ‘sparse vegetation’.
If two nuclei are very close and their separability negligible, they can be combined into a single
nucleus. Such is the case of the nuclei ‘dry grasslands’ and ‘green grasslands’.
5. Some nuclei may be least separable. This can be seen in the line corresponding to incriminated
nucleus. It is better to remove them. This is the case of the nucleus ‘cloud’ for example.
6. It can be seen by observing the probability image that nuclei for which rejected pixels are minimal
are absent. ‘Vegetation with fair coverage’, ‘vegetation and dark soils’ and ‘dense vegetation’ are
examples.

B Performance table
Before accepting the final classification, it is important to consider the last quality indicator of this
classification, viz., the performance table (Table 9.2).
It may be recalled that a nucleus is a group of pixels that serves to a priori characterise an object
we wish to use for designing a classification procedure and that a thematic group is the population of
pixels which at the end of classification, hence a posteriori, is combined into a single group. After
classification, the thematic groups carry the same name as those of the nuclei used to constitute them
during classification.
If all the pixels of a nucleus are found in the corresponding thematic group, it is considered that
the pixels of the nucleus are well chosen and that performance is very good. Contrarlly, it often happens
that a certain number of pixels chosen to constitute a nucleus are found classified in another thematic
group that does not correspond to the nucleus. This indicates that the pixels of the nucleus In question
are not correctly chosen. Choice of pixels constituting the nucleus can be modified or nuclei that are
too close can be changed.
The performance table is useful for evaluating the homogeneity of every thematic group and
hence of nuclei chosen before classification (Table 9.2). Rows in the table represent the nuclei and
columns the thematic groups. The distribution of pixels used to characterise the nuclei In various
thematic groups can be read from each line. The sum over a line Is hence equal to the number of
pixels that represent the nucleus.
Classification errors in this table are far fewer than in the separability table. The sum of pixels of
nuclei classified into the corresponding thematic groups (first diagonal) is 33,236 or 91.2%. Classification
quality is hence acceptable. Each line also gives the percentage of well-classified pixels. Pixels
 Table 9.2: Performance table for 20 nuclei of maximum-likelihood classification of the Brienne image

Nuclei 1 2 3 4 5 6 7 8 9 1 0 11 1 2 13 14 15 16 17 18 19 2 0

Thematic
groups

1 White soil 131626 4 25

2 Very bright soil 4 1404 36 1 1 34 1 2

3 Bright soil 337 1505 9 2 6 5

4 Fairly bright soil 34 1702 1 0 8 1

5 Fairly dark soil 42 1664 39 1 2 3

6 Dark soil 34 775 57 1 2 1 13 1 1

7 Very dark soil 1 33 800 2 1 11

8 Sparse 4 19 109 31 1 1432 15 55 1

vegetation
9 Dry crops 2 2 2 15 7 1 1 0 1281 140
1 0 Crops 2 3 3587 1

11 Green 3 4 21 2133 377 143 7 3

grasslands
1 2 Dry grasslands 4 8 25 136 1230 16
13 Riverine forests 1 1 8 45 8 15 250 47 2703 571 54
and forests
14 Dark forests 78 4353 2

15 Very clear water 3534 2 0

16 Clear water 1480 1 0

17 Fairly clear water 1 1441 29

18 Turbid water 2 381
19 Cloud shadows 30 2 8 393
2 0 Clouds 19 1 1 1 1 26 1 2 2

Number of pixels 1341179515101341 1795 1610 945 1509 1370 3604 2524 1728 2944 4933 3534 1500 1459 410 464 2943
of thematic groups
 Processing and Interpretation 185

corresponding to the nuclei of clouds, riverine forests and forests and bright soil are poorly classified.
Test zones of these nuclei may hence be improved.

9.2.5 Iterations: a h euristic approach

After an iteration is completed and balance is achieved, it is very often necessary to carry out other
iterations.
When a small number of nuclei is taken, separability is generally good but rejections would be
many.To overcome this, it is necessary to increase the number of nuclei and avoid ultimately grouping
them in a single colour, if several groups have the same function for the objective envisaged.
We can also combine multiple groups, apply a mask and reclassify them. This procedure may be
followed in the case of bright soils, clouds and riverine forests and forest zones in the image of Brienne
region, since confusion exists between these features.
In the example of Brienne, after three iterations (CD 9.6) for the same image with 22 nuclei, the
following results are obtained: some are the same as in the preceding iteration, others have a modified
composition, some have been suppressed and some new nuclei have been created. Hence it
is Imperative to analyse the modifications produced by various Iterations In terms of quality of
classification.
At the end of the third iteration, the image obtained is coloured so as to better reveal thematic
units (CD 9.6). The major units earlier detected from other classifications are again observed. There is
no thematic unit with less than 3,600 pixels. The largest class, viz., ‘green grasslands’ has more than
140,000 pixels. As mentioned earlier, this class was poorly defined since all the pixels having a digital
characteristic of chlorophyllian vegetation and not being in a specific group are classified in this group.
This group could be redefined if necessary to study a particular theme in detail. However, this would
not be simple since this group replaces the concerned boundaries already existing between the
‘chlorophyllian’ groups. Hence, this group can be isolated by masking and reclassification carried out
only on the corresponding pixels. Fifteen different groups were thus obtained (CD 9.7).
For a more detailed analysis of soil bright, which is very useful for estimating susceptibility to
erosion, the thematic groups ‘fairly bright soils’ and ‘fairly dark soils’ can be reclassified. Each of these
comprises more than 100,000 pixels.
Various thematic groups can be readily combined for obtaining a general picture of the region.
The combined groups are coloured with a single colour. For example, only seven themes may be
retained (CD 9.8): 1— bright bare soils, 2— dark bare soils, 3— in-sltu crops, 4— grasslands, 5— forests
and woods, 6— water bodies and 7— shadows and very dark soils. The major regions of chalky
Champagne with very bright soils and large plots drained by valleys can be readily recognised. Very
dark, clayey soils form a transition zone with humid Champagne which comprises grasslands and
forests. The Brienne plain consists of bright soils and, further south, lakes and reservoirs.

9«2.6 C lassification quality assessm ent

To analyse the results of maximum-likelihood classification its quality needs to be determined. This is
achieved by successively examining the following aspects:
— spectral characteristics of each nucleus,
— separability,
— probability image with rejection level,
— performance table.
Three iterations are carried out, designated ML1, ML2 and ML3, using the maximum-likelihood
method, each time improving the nuclei chosen and accuracy of classification (performance table).
186 Processing of Remote Sensing Data

The quality of classification was estimated by the same computations as earlier (sum of diagonals)
and 94.2% of the pixels defining the nuclei found to be well-classified In their respective thematic
groups for MLS (Table 9.3) compared to 91.2 for ML1 (values used for this computation were taken
from a performance table with zero rejection threshold). This Is satisfactory inasmuch as the number
of thematic groups has increased from 20 to 22.

Table 9.3: Performance evaluation of various classifications.

Values of well-classified pixels for ML1 and MLS are given in bold; values mentioned in the text in italic.

Classi­ Rejection Well Poorly Rejected Sum Well Poorly Rejected

fication threshold classified classified pixels classified classified pixels
pixels pixels pixels (%) pixels (%) (%)

ML1 0 33,236 3,207 0 36,443 91.2 8 .8 0

ML1 1 32,589 2,970 884 36,443 89.4 8 .2 2.4

ML1 3 30,954 2,625 2,864 36,443 84.9 7.2 7.9
ML1 5 29,697 2,505 4,241 36,443 81.5 6.9 11.6
ML1 1 0 26,522 2,293 7,628 36,443 72.8 6.3 20.9
MLS 0 33,295 2,061 0 35,356 94.2 5.8 0

MLS 5 29,990 1,628 3,738 35,356 84.8 4.6 10.6

For the third iteration (MLS), if the rejection threshold is taken as 5 for example, the performance
table Indicates 10.6% rejection pixels and 84.8% well-classified pixels, whereas for the first (ML1)
iteration rejects are more (11.6%) and well-classified are less (81.5%). The percentage of poorly
classified pixels decreases from 6.9 to 4.6. Hence, performances of the third iteration are better than
the first.
Let us now examine the objective of classification. Are accurate thematic groups preferred even if
all pixels are not classified or classification of the entire image preferred even if the thematic groups
are not very accurate? Obviously, the more the thematic groups, the closer they are to one another.
Consequently, separability is poorer. If it Is desired to improve separability, more homogeneous nuclei
and hence of smaller frequency should be taken. This makes the method more difficult to apply since
it Is not easy to correctly adjust the statistical populations using a nucleus having too small a number
of pixels.
It is also possible to use the three axes of principal component analysis instead of the raw bands,
in order to reduce the dimensionality of the problem.

9.2.7 Conclusion
Maximum-likelihood classification is a mathematically satisfactory method since the pixels are classified
using a probability; this approach Is extremely desirable in remote sensing. Obviously, the choice of
nuclei and judgement of the expert and interpreter influence the classification. It should be noted,
however, that the method, in particular its operation in the TeraVue program, provides for monitoring
the choice of nuclei during the process by means of statistics and the separability tables. The
performance table likewise enables evaluation of the final result. Lastly, the image obtained with various
rejection levels gives a numerical value for the quality of results. Classification can be improved by
carrying out several iterations. The quality of this classification can thus be determined, which is
essential in image processing. However, it Is necessary in this case also that a true confusion matrix
be established, which allows only comparison of classification results with field data.
 10
Image Processing Methodology
Any classification in image processing results in reduction of the quantity of existing information and
definition of less number of units known as thematic groups or classes by combining several pixels.
Most often these groups are constrained with reference to their number, thematic significance,
statistical coherence, etc.

10.1 OBJECTIVES
In order to evaluate the constraints mentioned above, the objectives need to be defined before
processing the images. The objectives determine the characterisation of the field of study, fineness of
analysis desired, homogeneity Investigated, accuracy required, reliability of the methods, need for
diachronic images and the number of groups to be studied. The most common objectives are:
— Preparation of an integrated map and general investigation of the most characteristic units in
the field of study. This may correspond to the question: what are the major spatial units observed in the
region under investigation?
— Segmentation of the field of investigation contributing to a mission of study. This corresponds to
the question: where should we look on the ground for the major thematic units of this region?
— Investigation of a simple segmentation for some specific themes only. This may be represented
by the question: where are such objects, for example grasslands or uncultivated lands, located in this
region?
— Precise analysis of a single type of object, avoiding others. This may be represented by the
question: what are the various types of bare soil identifiable by their hue, in order to determine the
types of inputs they receive?
— Diachronic monitoring of an object and its modelling to enable prognosis.

10.2 METHOD
To achieve any of the several objectives the study has to be divided Into three phases as in any
Information processing system:
— status prior to processing: data available and objectives;
— proper processing and choice of methods;
— results of processing: evaluation and presentation.

10.2.1 Input
For a processing technique to be applicable to any investigation, the best possible images should be
chosen according to the objectives, environment and theme. An image processing not suited to the
188 Processing of Remote Sensing Data

objective pursued does not provide correct answers. However, most often images not necessarily
optimal for the given objective are analysed for reasons of cost. In such cases it is necessary to take
into consideration the accuracy required (which in turn determines the pixel resolution), image periodicity,
acquisition date and spectral characteristics of the phenomenon under investigation (which determine
choice of most suitable wavelength bands available with the existing sensors) (see Chap. 16).
Basic models appropriate for processing have to be defined. Radiometric or chorological models
are mainly used for this purpose.
Radiometric models are established from field measurements in the zone of study or in a zone
considered equivalent. If field measurements are not available, a general radiometric model (Chap. 7,
Fig. 7.4) can be used. Care has to be taken to adapt the model to the region or theme under investigation.
Chorological models are developed by thematic experts and geographically tested in various
sectors of the region or in equivalent regions. If precise chorological models are to be employed, it
should be remembered that they are in most cases applicable only for specific zones and cannot be
generalised.

10.2.2 Processing
The processing methods to be used are limited by the following technical considerations:
— capacity of the computer to analyse a large number of pixels and speed of execution of
classification;
— number of persons and work stations to be employed to obtain the answer, when faster analysis
is required (see Chap. 18).
The processing techniques employed are often those given in available softwares. Hence, it is
particularly important when purchasing the software to check that it is compatible with the problems to
be solved. Costly programs do not necessarily give better results. It Is imperative to recognise the
‘default’ options provided in each and to acquire the correct method pertaining to programmed
classification. Thus, principal component analysis rarely gives the same results for the same image
with different softwares.
The processing method used should be amenable to evaluation for quality. So it is necessary to
be equipped with a statistical tool that provides such evaluation; this is not always easy.
Generally speaking, thematic experts ought to Intervene during classification. This intervention
depends on various phases of classification and its impact on the results Is more or less significant, as
has been seen in the preceding chapters. In all cases It is desirable to most explicitly specify the
decision rules and phases in which they are applied. This Is a condition of repeatability of the method
and hence of technology transfer.
It Is often necessary to use several processing methods in a single work sequence in order to
adapt them to the specific problem posed during analysis of results or during preparation of images for
a subsequent processing. As mentioned earlier, a given processing method is Incorporated within a
more general procedure. Some examples are: return to parallelepiped analysis for better identification
of a thematic group, use of a vegetation index to subdivide or to reformulate a group of very similar
nuclei, succession of ascendant hierarchic classifications after masking a given theme, etc.

10.2.3 Output
Before utilising the result of a classification, it should be assessed for quality (see Chap. 17). This is
done with confusion matrices, usually represented as performance tables. The classification result is
compared with external data corresponding to ground observations or reference data acquired from
aerial photos or thematic maps (see Chap. 19).
 Processing and Interpretation 189

The image processing results ought to be evaluated from the semantic (content of each spatial
unit retained) as well as graphic point of view. It is important to interpret the shape of the map zones
obtained for each landscape unit and thematic group. This is carried out by analysis of shape criteria
(perimeter and surface area), readily obtained, especially when a geographic information system is
available.
Ground reference data should also be acquired according to the objectives envisaged. Collection
of such information is often a long process but must be entirely reliable since the type of classification,
method and procedure recommended are based on it. The method of validation of results ought to be
compatible with the cost and time allotted for the study. It represents one of the main considerations,
along with acquisition of data to be processed, in the management of a project.
The results of analysis are conventionally represented as a map, often on a computer monitor
which offers many advantages vis-à-vis map presentation on paper. The number of colours in which a
map can be displayed is much greater and readily changeable and zoom is always possible. A minimum
of 100 pixels is required for representation on a paper map and for defining the smallest map units
obtained at the end of classification. This corresponds to a minimal surface area of 4 ha in the case of
SPOT-HRV, 9 ha for LANDSAT-TM and even 0.1 ha for resolutions of 2 to 3 m. If we apply the quarter
rule, viz., that any zone should have at least 1/4 cm^ on the map, 100 pixels are needed per 25 mm^,
or 10 pixels per 5 mm.The scale of map representation E = 1/Xis hence defined as:

X = 2 p x 10^

where p is the pixel resolution used for processing, expressed In metres. The following presentation
scales are obtained:
Resolution Scale
10m 1:20,000 possible utilisation for 1:25,000 maps
20 m 1:40,000 possible utilisation for 1:50,000 maps
30 m 1:60,000 possible utilisation for 1:63,600 maps

10.3 PROCEDURE
Classification procedure varies according to the theme Investigated and the concepts of thematic
experts. A new view of the Earth can be obtained from remote sensing applying the following rule:
If a unit having a specific and detectable spectral characteristic can be concomitantly defined in
space and time, then an entity corresponding to this unit actually exists.
This entity does not necessarily correspond to a theme; It may represent a specific organisation
of objects (a pattern), the latter considered very different by various thematic specialists.
It is therefore necessary to identify an object not only by its nature, but also by its spatial organisation
and diachronic behaviour.This Identification can be done before, during or after processing of images.
Thus four steps of image classification can be defined (Fig. 10.1).

10.3.1 Initialisation
Initialisation consists of extracting maximum information from Image data by means of statistical,
structural, diachronic or chorological analysis. These analyses are unsupervised and necessitate no
field study. They enable identification of zoning of image features for which a precise legend cannot be
associated. While the container can be precisely determined, the content of each zone is often poorly
defined.
190 Processing of Remote Sensing Data
 Processing and Interpretation 191

10.3.2 Correlation
This stage consists of establishing references on the ground, in the laboratory or through investigations
at the time of image acquisition, in order to compare the spectral data recorded by the sensor with
those collected on the ground. The reference zones facilitate pre-supervision processing. For example,
a radiometric interpretation model adapted to the region under study can be created. From this stage
spatial units are obtained to which we can assign a legend prepared from the reference data acquired
prior to processing.

10.3.3 Verification
This stage involves determination by means of ground control oi the content of the zones delineated
by a preliminary unsupervisedclassWicaWon. A second post-supervision classification is then applied.
Maps thus obtained are accompanied by a legend corresponding to units identified and verified on the
ground.

10.3.4 Modelling
The modelling stage utilises field references for pre-supervision classification and the units obtained
are verified for their content by ground control. Co-supervision processing is thus carried out combining
ground control and reference control in a to-and-fro movement. Radiometric and chorological models are
developed. It Is possible to determine all parameters of a general model, which enables generalisations
based on certain rules, and introducing it in a geographic information system. The general model in
some cases imposes a diachronic study for incorporating characteristics of objects in it.

10.4 INTERPRETATION OF PROCESSING

The result of image classification is not the end of image analysis. The results need to be further
interpreted relative to the Initial objectives. This Interpretation is generally based on two approaches,
viz., radiometric and chorological.

10.4.1 Radiometric approach

Each object ought to be defined by its spectral characteristics and this requires ground reference
measurements. If the latter already exist, they can be used but only after verifying that they were
acquired in conditions similar to those in which the data will be used. Thus, laboratory measurements
and field measurements should be adjusted. Generally, measurements obtained in one region are not
directly applicable to another for reasons of topography, climate, latitude, etc.
As the field data most often consist of reflectance measurements, it is necessary to verify that the
reference model thus obtained is coherent with the digital numbers of the image processed.This Is not
always simple.

10.4.2 Chorological approach

Chorology (derived from Greek ‘khoros \ country or district, and ‘logos \ logic or science) is the study of
relationships existing between the characteristics of semantic (thematic) units identified (intrinsic factors)
and their distribution in three-dimensional landscape (extrinsic factors).
192 Processing of Remote Sensing Data

Chorological analysis in the case of remote sensing thus comprises geographic-domain

investigation of distribution of objects identified in radiometric space. Objects that define the map units
obtained from image classification are most often not randomly distributed in geographic space. Some
units are located in a particular geographic position, in a particular region or contrarily scattered, but
scattering is not necessarily random. Some units show systematic spatial associations; unit A always
surrounded by unit B which per se Is connected to unit C, and so forth. Frequently the shape of a unit is
indicative of its nature, for example digitate indicating a zone of circulation and infundibular representing
a zone of accumulation. Other common forms are elongate, compact, segmented, regular, etc.
Shapes and Interzone and interunit relationships are readily detectable in images. This is one of
the advantages of the latter. Analysis is much easier on computer systems since it is possible to play
with various colour contrasts, masks, zooms, etc., which enables evaluation of various relationships.
The data derived from image processing, once interpreted, are used to prepare maps
corresponding to the designated objective. However, at present these data often constitute input to
another information system.

References
Anonymous. 1994. Guide d’utilisation deTeraVue. Éditions La Boyère. Valbonne, 220 pp.
Baize D, Girard M-C. 1995. Référentiel pédologique. INRA-AFES, 332 pp.
Belluzo G, Girard C-M. 1997. Identification et classification de l’occupation du sol à partir de scènes Thematic
mapper: application aux prairies d’une région de champagne humide. Bull. SFPT, 146:2 2 -3 2 .
Bertin J, Barbut M. 1973. Sémiologie graphique. Mouton, Gauthier-Villars, Paris, 2'^^ ed., 431 pp.
Burlot F. 1995. De l’interprétation visuelle à l’interprétation automatique des images satellitaires: application aux
pédo- et hydropaysages. Mémoire Mastère SILAT, Grignon, 65 pp.
Congalton RG. 1991. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sensing
of Environment, 37: 35-46.
Deschamps P-Y, Duhaut P, Rouquet M-C, Tanre D. 1984. Mise en évidence, analyse et correction des effets
atmosphériques sur les données multispectrales de Landsat ou de SPOT. Les colloques de l’INRA, 23:709-
722.
Diday E. 1971. Une nouvelle méthode en classification automatique et reconnaissance des formes: la méthode
des nuées dynamiques. Revue de Statistiques Appliquées, 19 (2): 283-300.
Gao BC, Goetza AFH. 1995. Retrieval of equivalent water thickness and information related to biochemical
components of vegetation canopies from AVIRIS data. Remote Sensing of Environment, 52:155-162.
Girard C-M. 1987. Spectral and botanical classification of grasslands: Auxois example. Advances in Space Research,
7: 67-70.
Girard C-M, Girard M-C. 1995. Qualité des méthodes d’interprétation: application à la caractérisation et cartographie
d’unités de paysage, qualité et validation des résultats. Bull. SFPT, 137:62-66.
Girard M-C. 1983. Recherche d’une modélisation en vue d’une représentation spatiale de la couverture pédologique.
Application à une région des plateaux jurassiques de Bourgogne. Thèse Doc. ès sciences, Université Paris 7.
SOLS, 12:414 pp.
Girard M-C, Girard C-M. 1989.Télédétection appliquée: Zones tempérées et intertropicales. Masson, Paris, 260 pp.
Jambu M. 1978. Classification automatique pour l’analyse des données. Dunod, Paris.
Monget J-M. 1986. Cours de télédétection, 1. CTAMN, octobre, Sophia-Antipolis.
Monget J-M. 1986. Cours de télédétection. École des Mines de Paris, 165 pp.
Monget J-M, Robertson Y-C. 1992.Two-variable mapping applied to remote sensing data interpretation: a software
implementation. In: Remote sensing from research to operation. Proc. 18th Ann. Conf.The Remote Sensing
Society, pp. 571-580.
Orth D. 1996. Typologies et caractérisation des prairies permanentes des marais du Cotentin, en vue de leur
cartographie par télédétection satellitaire, pour une aide à leur gestion. Thèse INA P-G, 150 pp.
Van Den Driessche R. 1965. La recherche des constellations de groupe à partir des distances généralisées de
Mahalanobis. Biométrie, 41 (1): 36-47.
Wilmet J. 1996. Télédétection aérospatiale. Méthodologie et applications. SIDES, 300 pp.
 ___________________ n
Structural Processing of Satellite
Images

11.1 INTRODUCTION
The most common techniques of satellite image processing, viz., principal component and parallelepiped
analyses and ascendant hierarchic, mobile centres and maximum likelihood methods of classification,
etc. are all based on textural analysis oi images. In these methods each pixel is classified individually
without taking its neighbourhood into consideration. Methods based on structural analysis of Images
are utilised for three reasons described below.

11.1.1 Boundaries and mixels

The first reason for using structural analysis of images is that the grid formed by pixel positions on the
ground Is entirely arbitrary since it depends on the satellite trajectory and sensor type. This is one of
the reasons why a number of mixels exist whose spectral characteristic is a combination of spectral
responses of several objects geographically included in the resolution element (see Chapter 15). This
makes delineation of a boundary between two objects difficult even when it Is distinct on the ground.
Hence, it can be practically assumed that most boundaries identified on images are erroneous. This
leads to a significant error in classification. This error will be larger if the map units Identified are small
or digitate and hence of a longer perimeter. This ‘noise’ which frequently attains a strip of 1 to 2 pixels
along the boundary of the map units should be related to the mapping quality desired. Any superposition
of various maps on images by imposing a precision greater than 1,2 or 3 pixels (or 20, 40 or 60 m for
SPOT multispectral Images) would be in vain. It is therefore necessary to take the pixel neighbourhood
into consideration for this primary reason.

11.1.2 Classification and mapping

The second reason for using structural analysis is as follows: textural methods are efficient as long as
a definite spectral characteristic can be defined for an object. In almost all cases, a classification is
obtained in which each class comprises several map units that are not compact. The result of such a
classification is similar to an impressionist’s tableau with a multitude of small patches of various colours,
seen from distance, an assemblage of large units. However, no boundaries of these units exist. Fuzzy
logic needs to be applied to generate these ‘homogeneous units from groups that are heterogeneous
by nature’ which are called patterns. In fact, for these units it is difficult to decide appurtenance of a
pixel to a class with absolute certitude.
194 Processing of Remote Sensing Data

Moreover, thematic maps exhibit a relatively limited number of zones for each map unit retained
and these zones are much more compact than those obtained through textural image processing.
These units often correspond to complex groups of objects such as juxtaposition, association, sequence,
combination, etc., which comprise several elements. It is therefore difficult to compare a remote-
sensing document derived from a textural classification with a thematic (structural) map using the
geographic information system, since the nature and number of spatial units— classes in remote sensing
and map units in thematic study— are too diverse.
It is hence necessary to find a classification method that enables enhancing compactness of
image zones.

11.1.3 Complex map units

The third reason for use of structural analysis arises from the fact that entities characterised by a
group of objects related to one another and exhibiting spatial organisation do exist. Such is the case,
for example, of a ‘valley’ which comprises the objects ‘water’, ‘trees’, ‘grasslands’, ‘alluvium’, etc. In
such cases structural Image processing has to be employed. However, only very few classification
algorithms based on image structure exist.
Two methods of structural processing of satellite Images based on neighbourhood analysis, viz.,
OASIS and VOISIN (Didier, 1992; Francoual, 1994; Girard et al., 1980, 1990 and 1991; Gilliot, 1989,
1992 and 1994; Ranaivoson, 1990), are presented below for various applications. Similar methods
have also been developed in France, viz., PAPRY (Borne, 1990) and CLARAS (Robbez-Masson,1994).
This is one of the important lines of research at present (Girard, 1995; Girard and Girard,1996; Girard
et al., 1997).

11.2 VOISIN
Classification of a pixel should also take Into consideration the pixels that surround it and hence Its
neighbourhood. This neighbourhood may be defined by a certain distance from the pixel under study.
It thus consists of investigating le Voisinage et l’Organisation des Informations Spatialisées par
Informatique et Numérisation (neighbourhood and organisation of spatial data by information technology
and digitisation), known as the VOISIN programme.
The objective of VOISIN is to determine the neighbourhood of an image pixel as it gradually
increases. Characterisation of a point effectively varies with its surroundings. Let us consider an example:
a fisherman (1 ) is situated in a boat (2), sailing in a lake (3), surrounded by a beach (4), bounded by
mountains (5) in a country (6) in Europe (7) on the planet Earth (8) In the solar system (9), galaxy
Milky Way (10). These ten levels of neighbourhood represent totally different distances varying from 1
m to 10^^ km. The fisherman can be defined In ten different manners: a man, a marine, a resident of
a country, an European, aTerrestrian, etc. All these definitions are true but important only once, when
the perception level taken into consideration Is defined (see Chapter 15). The VOISIN model enables
us to delineate, through the neighbourhood parameter, boundaries of spatial units whose heterogeneity
is defined by a pattern identified by a composition vector.

11.2.1 Neighbourhood in window: composition vector

The composition vector is a textural measure giving frequency distribution of various classes in a
given neighbourhood. It Is described by a local histogram. While this measure does not take the shape
into consideration, it is well suited to describe textures of satellite images and facilitates determination
 Processing and Interpretation 195

of patterns. For example, a valley constitutes a repetitive pattern, comprising free-water surfaces,
trees (riverine forest), some cultivated plots, grasslands and roads.
The level of neighbourhood is analysed by means of a window as follows (Fig. 1 1 .1). For a square
window FfOi size f (f is the number of pixels defining the side of the square), a ‘composition vector’
is defined as the vectorial sum of elementary vectors Q . Each of the latter represents a class Cj,
balanced by a coefficient equal to the ratio of the number of pixels in class Cj, present in the window
under consideration, to the total number of pixels in the window (or t^).

Q = Pc1 + P c 2 ¿ 2 + P c 3 ¿ 3 + ■ • • ■*“ P c i ^ i + • • •

Image in pixels of a portion of the ground

□ w □ Ig Pi d Mb Land cover classes
Composition vectors of three pixels p1, p2 and p3

p1 = 20.4 b + 30.6 d + 22.4 Ig + 26.6 w

p2 = 18.4 b + 32.6 d + 22.4 Ig + 26.6 w

p3 = 8.2 b + 32.6 d + 20.4 Ig + 38.8 w

Fig. 11.1: Definition of the neighbourhood by a composition vector.

11.2.2 Method
For neighbourhood analysis, a given point of the image constitutes the central pixel for successive
windows.The window size is gradually increased to values of f = 3, 5, 7, ...(Fig. 1 1 .2) and difference D
between the successive composition vectors q is computed as: C5- C3; D j = C j- C^; Dg = Cg-
Cj] ....The results are represented on a graph with the successive differences Dg, Dy, D g,... on the
y-axis and the size of successive windows on the x-axis. These sizes are equivalent to an increasing
geographic distance vis-à-vis the central point.
Sizes of successive windows for which differences are close to one another are detected from
this curve. Window sizes that have relative minima in D, denoted as t^ are used for subsequent
analysis. For these window sizes, the local histogram remains stable even if the window size changes
and permits determination of a pattern. It is characterised by the window vector of size f^.The magnitude
of the difference D ^ - is a measure of Its heterogeneity (Fig. 11.2). Most often, several local
minima appear In the D = f (t) curve, defining multiple patterns. The geographic distance between
these various patterns is expressed by the size difference between two windows corresponding to two
minima of D. In this way we can
196 Processing of Remote Sensing Data

Pixel image of a portion of the ground Neighbourhood curve of the central

□w Dig M 6 M b Land cover classes point of the image on the left

□ Representation of successive windows of

9 to 169 pixels

Fig. 11.2; Neighbourhood of a point in an image and representation of its neighbourhood function showing two
local minima and /r?2 for windows of 3 and 11 pixels. Minimum corresponds to a pattern consisting of light
grey or dark pixels (Ig, d) and extends north-west; the minimum m 2 represents a pattern containing more black (b)
and white (w) pixels than the preceding, it extends more southwards.

1) identify patterns,
2) characterise them by their ‘composition vector’,
3) obtain a measure of their heterogeneity by the value of D,
4) estimate the distance between two patterns that are juxtaposed; this corresponds to the minimal
distance on the ground to change from one pattern to another.

11.2.3 Example
At a given point in the space under study, the composition vector for the smallest window F3 of size
3 X 3 is determined. Composition vectors are then computed for increasing window sizes, viz., 5 x 5 ,
7 x 7 , ..., (n~2) X (n -2), n x n.
Manhattan distance between each composition vector and that of the next higher size, 1
VFr^A, is computed. We thus obtain the neighbourhood curve, which represents change In tne
neignbourhood when the field of study Is enlarged.
When the same object is analysed while increasing the field, Manhattan distance remains zero;
this represents a homogeneous pattern. If heterogeneity manifests in the form of appearance of a new
object, this distance increases; this increase is more when this object occupies a larger area or several
objects appear together. In all cases, increase In distance indicates a change in neighbourhood. If
increase in the field of investigation is continued, the Manhattan distance curve passes through a
maximum and then decreases to reach a new relative minimum, i.e., a second pattern Is observed. Its
statistical composition is determined from the composition vector.

Interpretation of neighbourhood curve (Fig. 11.3) defined by the relation

‘mathematical distance between windows’ = /[window size]
 Processing and Interpretation 197

CO
Ti- Window size
CO

i
H=0 S=6.76 ha
DC
LLi
H =803 S= 33.64 ha
<-
> p H=204 S= 88.36 ha
i.1-2 L2-3
<- ->
2 0 0 m S20 m

Fig. 1 1 .3: Neighbourhood curve of a classified image. In this curve three patterns Pcan be distinguished, having
heterogeneity /-/defined for window sizes of area S, the width of the boundary between these patterns of definite
homogeneity being L

enables identification of various patterns (P), window sizes (f) (or area S = 400 x in for
multispectral SPOT) for which they have stable composition and width of the boundary between two
stable patterns (L). For each pattern a heterogeneity factor (H) can be defined, characterised by a
value of mathematical distance between successive windows of a definite pattern.
This model permits characterisation of various patterns situated around a localised point or, by
analogy, various planes of a landscape observed on the ground from an observation point. It establishes
locations of transitions (boundaries) between patterns in which the mathematical distance changes
from one value to another (boundary width), as well as places of maximum heterogeneity. Lastly, it
gives a value of internal heterogeneity for each pattern.
VOISIN is a model for heterogeneity characterisation of spatial patterns, leading to a localised
geostatistical study using n variables. It helps in decision-making for optimal delineation of
heterogeneous groups in a given spatial field. It enables correct choice of a representative moving
window for a spatial pattern to be used for OASIS classification.
This program also constitutes a tool for making scale changes and enables ‘synthesised
generalisation’ (see Fig. 15.1).
198 Processing of Remote Sensing Data

11.3 OASIS

11.3.1 Method
Ü Definition
OASIS, Organsiation et analyse de la structure des informations spatialisées^ (organisation and
analysis of spatial data structure), Is a method of supervised classification based on the fuzzy groups
theory, which uses spectral and textural characteristics for heterogeneity analysis.
Classifying all pixels according to neighbourhood parameter is the objective of this method. Each
class is characterised by a pattern defined by several classes of objects. This classification is based
on nuclei defined by the thematic specialist and assumed to characterise the patterns being searched.
The result gives zones that are coherent with the units of a map and much more compact than those
given by textural methods. Moreover, high quality classified Image is obtained. The OASIS method can
be used for raw bands of a satellite Image as well as for a single band such as a panchromatic image
or a result of classification. It is most efficient in the latter case since the composition vectors correspond
to a known assemblage of classes of objects but not to raw spectral characteristics.
A moving window is used to cover the entire image and aids in determining the composition
vector for each pixel.

M Nuclei
Several important patterns, designated nuclei, are defined either by investigating their composition
derived from the image or by parametrising their composition vector.
It is possible to define the composition of nuclei in satellite images by marking zones corresponding
to the patterns which the thematic specialist wishes to detect. The user delineates each nucleus by
one or several polygons traced on the computer monitor.
Composition of these zones, determined by the local histogram, is computed and the corresponding
nucleus is then defined from Its composition vector. It is possible to modify the composition vector or
to directly define it In a table.

® Size of moving window

The thematic specialist defines the window size. Either the window defined by means of VOISIN,
which indicates the average size of the patterns, is taken, or it is defined as a function of the object
studied. The larger the window, the larger the neighbourhood considered for characterising the pixel
and more compact zones obtained. However, If the window is too large, smaller objects will disappear
and will be integrated with a more complex reference zone. Thus, rivers are frequently missed with
windows of 5 or 7 pixels (CD 11.5 or 11.8). In this case, they are integrated into the reference zones
that comprise a few per cent of ‘river vector’ in their vector composition, as was observed in the
CORINE Land Cover study (Chapter 19).

H Computations
Manhattan distance from each pixel to each nucleus is computed and a pixel is assigned to the
nucleus for which the distance is minimum (DIMITRI method, Girard, 1983; Girard and King, 1988). A
preliminary classification is obtained which Is evaluated by the sum of minimal distances of all pixels

■•This free software can be obtained from the website http://lacan.grignon.inra.fr/resources.

 Processing and Interpretation 199

to the nuclei they are assigned to; the smaller this sum, the better the classification. Multiple
approximations are made until stability of nuclei is obtained. Computations may be extensive since
analysis of millions of pixels is involved depending on the windows (defining neighbourhood) which
comprise hundreds of pixels.

H Iterations
Nuclei of the next iteration are redefined from the statistics of pixels grouped in each nucleus. More
nuclei can be added in each Iteration and then a new Iteration started.
The sum of distances (S D) of pixels to nuclei in which they are grouped varies for each iteration
and gradually decreases. Several iterations can thus be carried out until the grouping is optimal. The
optimum Is reached when the sum of distances (Z D) becomes constant.
The term iteration Is employed when only the nuclei change from one classification to another.
Contrarily, If the window size Is changed or if the initial image is no longer used for making a new
classification but an already classified image is used, the term approximation is employed.

■ Results and quality

We thus obtain nuclei which at the end of processing are called references. The references define the
composition of map units of the final document. Each reference may be defined by a single class or
several classes. The former is a homogeneous reference and the latter heterogeneous. The latter
corresponds to a pattern whose composition makes the heterogeneity explicit. The pattern reveals all
information of the initial classes and hence no Information loss occurs. OASIS is not a reduction
method and differs completely in this from a smoothing technique.
The result of each Iteration is displayed as a classified image (one colour for each class) to which
a black and white image of distances of each pixel to the nucleus in which it is grouped is appended
(see the example and images in GD11.2 and 11.3). The latter gives a quantitative estimate of the
quality of grouping. Its analysis enables identification of well-classified and poorly classified regions.
New nuclei more appropriate to poorly classified regions are redefined in such cases.
Compactness of zones obtained varies according to the window size used to define neighbourhood.
Form and size of zones vary with window size in a more complex manner (Fig. 11.4) but generally
towards greater compactness (expressed by a maximum area for a minimum perimeter). At the end of
the process, images with much more compact zones are obtained than in the maximum likelihood
classification, for example.
The Images can then be readily converted into vector mode for introducing in a geographic
Information system.

H Conclusion
OASIS Is a non-hierarchic structural classification system based on patterns defined by organisation
and neighbourhood of each classified pixel, which creates non-hierarchic units.
Classification by OASIS enables delineation of compact map units from the patterns chosen and
model heterogeneity. OASIS, through iteration, determines the best spatial division for a moving window
of a given size.

11.3.2 Example
Processing techniques corresponding to this example are given In the CD in Index CD 11. Each
illustration Is indicated by a number 1 to 12, which represents the Image band processed.
200 Processing of Remote Sensing Data

Area

Fig. 1 1.4: Variation in shapes of map units according to various approximations of OASIS and window size (after
Yongchalermchai, 1993).
Points represent shape of units obtained from three types of classification:
ML: Maximum likelihood; OASIS 1: First approximation of OASIS; OASIS 2: Second approximation of OASIS.
SM, SC and SR represent three types of units.
The figure shows their variation during three types of analyses,
a: shows all zones; b: represents enlargement of the preceding figure between values 0 to 4 ha and 0 to 2.5 km.

■ Characterisation of nuclei
The image in CD 11.1 was classified by the maximum likelihood method after three iterations (see
Chapter 9). Eight nuclei are defined, viz., forest (violet), water (blue), valley (dark cyan), riverine forest
(khaki), grasslands (light green), bright soils (light yellow), beige soils (salmon) and dark soils (maroon).
The composition of each nucleus is given as a function of 22 classes used for maximum likelihood
classification (Table 11.1).
The nucleus ‘forest’ thus comprises 88% pixels classified as ‘hardwood’ and 12% pixels classified
as ‘conifers and shadows’ in the maximum likelihood method. A similar situation is observed for other
nuclei. However, for better readability of the table, cells are left blank when the value Is less than 10%.
 Processing and Interpretation 201
Table 11.1: Composition of nuclei from groups of maximum likelihood classification
The values given in the table are percentages read along a row: the sum of a row is equal to 1 0 0 . Percentages
less than 1 0 are ignored.

MVS
Nuclei 1 11 14 15 16 17 22 23 26
White Very Bright Fairly Fairly Dark Grass- Hard- Very Clear Fairly Green Conifer Dense
soil bright soil bright dark soil lands wood clear water clear crops sha- vege-
soil soil soil forest water water dows tation

Riverine forest 18 18 18 10 15
Grasslands 56 35
Forest 88 12
Water 35 32 31
Valleys 14 20 10
Bright soil 2 0 38
Beige soil 15 34 17
Dark soil 2 0 28 1 2

Consequently, columns corresponding to groups of maximum likelihood classification having values

less than 10% are removed.
The nucleus ‘water’ Includes various types of water. ‘Forest’ mainly consists of deciduous trees
and conifers. The compositions of ‘valley’ and ‘riverine forest’ nuclei have common groups, viz.,
‘grasslands’ and ‘conifers and shadows’, which in the case of riverine forest correspond to shadows.
However, the nucleus ‘valley’ also comprises ‘fairly bright soils’, whereas riverine forest Includes
‘deciduous’ and other chlorophyllian plants. The nucleus ‘grasslands’ Is made up of ‘grasslands’ and
‘dense vegetation’. Various soil nuclei have compositions extending over several groups of maximum
likelihood classification, but a continuous variation exists from bright to dark soils.

■ Interpretation of results: first approximation

A first approximate classification is made using OASIS, with a 3 x 3 window of 9 pixels. An image (CD
11.2) is obtained which is less differentiating than that of maximum likelihood classification. Fairly
simple groups are obtained: all the pixels assigned to various types of water are classified in the
nucleus ‘water’. The nucleus ‘forest’ is slightly less than the sum of the two groups ‘forest’ and ‘conifers
and shadows’. A final composition after each approximation is also given, which can be compared
with the initial.
A black and white image is also given for each approximation for assessing the quality of
classification. This image gives the value of distance separating a pixel from the nucleus to which It is
attached. The smaller the distance (represented as blank), the better the quality of classification (CD
11.3) . The histogram of distance values extends from 3 to 199 (Fig. 11.5).
The channel that gives classification and the channel that gives its quality can be combined into
a single image (CD 11.4). Cbjects are represented by various colours. Their classification quality is
expressed in the same colour of higher intensity when it is good (code 4) and lower Intensity when It is
poorer (code 0). It can be seen (Table 11.2) that the best-classified patterns, relative to the initial
objects, are forest (66 of 93 per thousand well classified), grasslands (140 of 193) and beige soils (73
of 160). This Indicates that one object predominates in each pattern (see Table 11.1).
202 Processing of Remote Sensing Data

Fig. 11.5; Histogram for first approximation of distances of all pixels classified into their classification nuclei. If a
pixel is perfectly classified, its distance is zero.

Table 1 1.2 : Classification quality of 8 patterns of first approximation. Values for 4 quality grades are given per
thousand image pixels.

Quality High (4) Correct (3) Poor (2 ) Inadequate (1) Total

Water 0 4 28 3 35
Forest 6 6 17 4 6 93
Valley 0 7 0 1 0 17
Riverine forest 1 26 35 19 81
Bright soils 0 39 70 65 174
Beige soils 0 73 85 2 160
Dark soils 0 35 144 65 244
Grasslands 15 140 37 1 193
Total 82 341 403 171 997

H Change of nuclei: second approximation

After preliminary conclusions are drawn, nuclei are redefined in a second approximation either by
adjusting them, combining them or taking new ones. In the example considered, the nucleus Valley’ is
removed as it is too close to riverine forest. The new composition of nuclei for the second approximation
is shown in Table 11.3, wherein rows indicate the composition of nuclei of second approximation from
the nuclei of first approximation.
If, for example, a window with a side of 5 pixels, i.e., total 25 pixels, is chosen, the resultant image
describes a neighbourhood of 1 ha around each pixel. A second image is obtained (CD 11.5) which
consists of more compact units. It is simpler to analyse.
The pixels that were grouped In the unit Valley’ In the first approximation are now found in the
units ‘riverine forest’, ‘grasslands’ and ‘dark soil’. The pixels In ‘riverine forest’ of the second approximation
comprise pixels that belonged to the units ‘riverine forest’, ‘valley’, ‘grassland’ and ‘forest’ in the first
approximation. This is logical in view of the composition of the nucleus ‘riverine forest’.
 Processing and Interpretation 203

Table 11.3: Composition of nuclei for making second approximation (in rows) from the classes obtained in the
first approximation (in columns). Values are expressed as percentage.

Nuclei 1 Riverine Grass- Forest Water Bright Beige Dark soil Valley Number
forest lands soil soil of pixels

Riverine forest 53 29 12 100

Grasslands 6 92 3040
CM Forest 2 98 1726
o
Water 100 2047
D
z Bright soil 84 14 1 3499
Beige soil 10 77 13 4939
Dark soil 12 85 5053

The image of distances is much clearer, Indicating that the quality of classification has improved.
The entire image is fairly well classified. It Is mainly the boundaries of grasslands, water bodies,
forests, etc. which appear as poorly classified (the dark zones). They represent assemblages of more
than 7 objects used for classification over a small area (25 pixels or 1 ha). These zones, being
heterogeneous, exhibit a wide variation. The histogram of distance values extends from 1 to 186 (Fig.
11. 6 ).
Comparison with the preceding approximation (Fig. 11.5) shows that the range of distances has
decreased and hence the classification quality has improved. Comparison of quality of the two
approximations (CD 11.6) indicates that poorly classified pixels constitute 3.9% and are localised (in
blue) mainly along valleys, which Is explained by the removal of the nucleus Valley’ between the two
approximations. They are observed in the boundaries between the units ‘forest’ and ‘grasslands’.They
are associated with the pixels that have not changed the group (in red: same distance in the two
approximations) and represent 5.5% of the image pixels. Improvement in the quality of the image

Fig.l 1.6 : Histogram for second approximation of distances of pixels classified to their nuclei. If a pixel is
perfectly classified, the distance is zero.
204 Processing of Remote Sensing Data

between the two approximations (which corresponds to decrease in distances) is hence greater than
90%.
Comparison of the classification results of the first and second approximations (CD 11.7) shows
that 172,973 pixels (15%) have changed group, which is due to the removal of the unit Valley’ and also
to modification in window size. There is no change of class in 85% cases.
Lastly, one can compare the characteristics of the composition vectors of the nuclei used to make
the second approximation (Table 11.3) with the composition of reference vectors at the end of the
second approximation (Table 11.4). Comparison of the two tables shows some small modifications in
the composition of vectors. The unit ‘riverine forest’ consists more of ‘forest’— 23% instead of 12%
initially— and less of ‘grasslands’— 18% Instead of 29%. The reference unit ‘beige soil’ comprises less
‘beige soil’— 64% instead of 77%— and more ‘bright soil’— 20% instead of 10%. The reference unit
‘dark soil’ is characterised by a composition vector comprising almost all classes, each represented
slightly. However, the class ‘beige soil’, which had a weight of 12% In the nucleus, represents only 3%
in the reference unit.

Table 11.4: Characterisation of composition vectors of reference units obtained at the end of the second
approximation (row) from the land-cover classes of the first approximation (column).
Values expressed in percentage.

Nuclei 1 Riverine Grasslands Forest Water Bright Beige Dark Valley

forest soil soil soil

Riverine forest 55 18 23 4
Grasslands 12 84 1 2 1
CM Forest 4 2 93 1
O
"o Water 100
13
Z
Bright soil 82 15 3
Beige soil 20 64 15
Dark soil 2 5 1 1 3 83 5

■ Change of window: third and fourth approximations

For the third approximation, only the window size Is changed from 25 pixels (5 by 5) to 49 pixels (7 by
7). The image is little modified (CD 11.8), except that the units are more compact and units of very
small surface area have disappeared because absorbed by neighbouring units. Consequently,
characteristics of the final units have changed little.
On comparing the classifications as done earlier, it can be seen that the unit ‘riverine forest’ is still
the most heterogeneous, producing maximum changes between the two approximations (CD11.9).
This Indicates that riverine forest is made up of a mosaic of very diverse units over a small distance,
since significant modifications are observed when the window size is changed from 25 to 49 pixels.
Other modified pixels pertain to boundaries of the units and hence are boundary mixels (see Chapter
15). The modifications concern 9.3% pixels instead of the 15% earlier.
With respect to quality, the distance histogram varies from 0 to 184 (Fig. 11.7). In fact, most pixels
have slightly modified values due to change in window size.
A fourth approximation is made (CD 11.10) by significantly increasing the window size to 225
pixels (15x15). Modifications are observed between the nuclei before the approximation and the final
reference units (Table 11.5).
 Processing and Interpretation 205

96 128 160 224 256

Distance of a pixel from its nucleus

Fig. 11.7: Histogram for third approximation of distances of pixels classified to their nuclei. If a pixel is perfectly
classified, the distance is zero.

Table 11.5: Characterisation of composition vectors of reference units obtained

at the end of the fourth approximation (row) from land-cover classes of the
second approximation (column). Values rounded off to percentages

Nuclei 2 Riverine forest Grasslands Forest Water Bright soil Beige soil Dark soil

Riverine forest 6 8 1 2 2 0

Grasslands 2 2 76 1 1

Forest 18 15 67 1

Water 1 0 0

Bright soil 3 72 21 4
"o
D Beige soil 2 1 11 6 8 18
z
Dark soil 2 6 2 1 0 80

The quality of results has improved since the distance histogram varies only from 0 to 139 (Fig.
11.8 ).
Comparison of the results of classification at the end of fourth and third approximations (CD
11.11) shows that the modifications concern 2 1.4% pixels. This large number indicates that a threshold
has been exceeded between a neighbourhood estimated over 49 pixels and a neighbourhood expressed
over 225 pixels. Hence the threshold of organisation level has been changed (see Fig. 11.3 and
Chapter 15). Evidently, these computations can be made by systematically increasing the window of
the neighbourhood and a function of changes defined relative to the size of the neighbourhood. Thus
the procedure is similar to that used In VOISIN but applied to the entire image. A similar method is
described in Chapter 15.
206 Processing of Remote Sensing Data

32 64 96 128 160 192 224 256

Distance of a pixel from its nucleus

Fig. 11.8 : Histogram for fourth approximation of distances of pixels classified to their nuclei. If a pixel Is perfectly
classified, the distance is zero.

■ Conclusion
In total, modifications have occurred for 32.5% pixels (CD 11.12) between the first approximation with
a neighbourhood window of 9 pixels and the fourth approximation with a window of 225 pixels. This
indicates that more than two-thirds of the pixels are classified in a stable manner. This shows the
advantage of this type of classification when it is desired to delineate relatively compact and composite
map units from automatic processing of satellite images. This is a method for making generalisations
by synthesis (see Fig. 15.1) and scale changes.

References
Borne F. 1990. Méthodes numériques de reconnaissance de paysages, application à la région du lac Alaotra,
Madagascar. Thèse de doctorat. Université de Paris 7, 213 pp.
Didier F. 1992. Analyse structurale des images: étude de l’organisation spatiale et de l’hétérogénéité. Mémoire de
Diplôme d’Études Supérieures Spécialisées, École des Hautes Études en Informatique, Université René
Descartes, Paris, 29 pp.
Francoual T. 1994. OASIS, notice d’utilisation du logiciel. Laboratoire de Science des sols et Hydrologie de i’INA-
PG, 18 pp.
Gilllot J-M. 1989. Analyse thématique d’images par multidensité, DESS EHEI. Université Paris 5, 56 pp.
Gilllot J-M. 1992. OASIS, un système de télédétection sur station IBM RS/6000. Rapport INA-PG, 24 pp.
Gilliot J-M. 1994. Traitement et interprétation d’images satellitaires SPOT, application à l’analyse des voies de
communication. Thèse de Doctorat, Université René Descartes, Paris 5,197 pp.
Girard C-M. 1995. Changements d’échelle et occupation du sol en télédétection. Bull. SFPT, 140:10-11.
Girard C-M, Girard M-C. 1994. Aide à la cartographie d’unités paysagères par une méthode d’analyse du voisinage
des pixels: application en Basse Normandie. Photo-interprétation, 3-4:145-154.
Girard C-M, Gilliot J-M, Girard M-C, Thorette J. 1997. Comparaison de la cartographie de l’occupation des terres
par classification de données de télédétection avec la cartographie CORINE niveau 3: application à une zone
au nord-ouest de l’Ile-de-France. Revue internationale de Géomatique, 7 ( 1 ): 57-86.
 Processing and Interpretation 207

Girard M-C. 1993. Recherche d’une modélisation en vue d’une représentation spatiale de la couverture pédologique.
Application à une région des plateaux jurassiques de Bourgogne. Thèse Doc. ès sciences, Université Paris 7.
SOLS, 12:414 pp.
Girard M-C, King D. 1988. Un algorithme interactif pour la classification des horizons de la couverture pédologique:
DIMITRI, Science du sol, 26 (4): 101.
Girard M-C, Girard C-M, Rogala J-P. 1980. Automatisation de l’interprétation de l’humidité des sols et interprétation
des paysages ruraux. OPIT, 85 pp.
Girard M-C, Mougenot B, Ranaivoson A. 1990. Présentation d’un modèle d’Organisation et Analyse de la Structure
des Informations Spatialisées (OASIS), Deuxièmes journées de télédétection: Caractérisation et suivi des
milieux terrestres en régions arides et tropicales. ORSTOM, Bondy, pp. 341-350.
Girard M-C, Yongchalermchai C, Girard C-M, 1991. Analyse d’un espace par la prise en compte du voisinage.
Gestion de l’espace rural et système d’information géographique. INRA. Florae, 22-24 octobre, pp. 349-359.
Girard M-C, Girard C-M, Bertrand P, Orth D, Gllliot J-M. 1996. Analyse de la structure des paysages ruraux par
télédétection. C.R. Acad. Agri. Fr., 82 (4): 11-25.
Ranaivoson A. 1990. Organisation et Analyse de la Structure des Informations Spatiales. DESS, Université Paris
5, INA-PG, Grignon, 36 pp. et annexes.
Robbez-Masson J-M. 1994. Reconnaissance et délimitation de motifs d’organisation spatiale. Application à la
cartographie des pédopaysages. Thèse de doctorat de l’École nationale supérieure agronomique de Montpellier,
161 pp.
Yongchalermchai C. 1993. Étude d’objets complexes, sol/plante, à différents niveaux d’organisation: de la parcelle
au paysage. Thèse de l’INA-PG, Sols, Grignon, 19:232 pp.
 12
Digital Filtering of Images
Filtering is the action of a filter. From an electronic point of view, a filter Is a tool designed to pass or
block certain frequency components of an electric signal (Larousse, 1998). Digital image filtering has
emerged from the theory of signal processing. It consists of a group of methods that analyse the
signal in spatial domain or frequency domain. The field of application of these methods in Image
analysis Is quite vast, involving all stages of processing, viz., preprocessing (noise reduction), detection
and extraction of image elements (edge detection), analysis (morphological analysis) and post­
processing. Conventionally, two major categories of filters are distinguished depending on whether
the operators employed fulfil or not the linearity criteria. Accordingly, they are known as linear and
non-linear filters.

12.1 LINEAR FILTERING

12.1.1 Image as a two-dimensional signal
■ Source of digital image: discretisation
A signal (s) constitutes the physical basis of information (Kunt, 1981). Mathematically, the signal is
represented as a function of one or several variables (xand y for an image). Depending on the nature
of the variables, two types of signals are distinguished: analog signal for continuous variables and
discrete or sampled signal for discrete variables. Further, the amplitude of the signal (grey level for an
image) can also be continuous or discrete.
An analog signal whose amplitude is discrete is a quantised signal.
A discrete signal whose amplitude is discrete is a digital signal. Let s be a digital signal of a
variable k, designated as s {k). t(k) is an offset or delayed version of s(k) with a lag /Cq such that t(k)
= s (/c-/Cq). a digital signal scan be expressed as a balanced sum of offset unit impulses d(Fig. 12.1):

/= + o 0

s(/c)= ^ { l)d {k -l)

(1)
/= - o o

A digital image / is an example of a two-dimensional digital signal generated from two stages:
— sampling, which is spatial discretisation In xand yco-ordinates;
— quantisation, which is discretisation of signal amplitude that represents the grey level.
A digital Image is hence represented by a discrete and finite zone of discrete amplitude (Fig.
12.3). It constitutes spatial representation of an object of a two-dimensional or three-dimensional
scene or of another image (Haralick and Shapiro, 1991).
 Processing and Interpretation 209

d{k)

Fig.12.1: A one-dimensional digital signal s(k) and a unit impulse d(k).

■ Gridding (tessellation) of a digital image

Division of a continuous space leading to discretisation of an image is done along a grid. A grid is a
division of a plane using the same elementary geometric figure known as a cell. Several types of grids
or meshes (graphs associated with grids) exist. The most common grids in image processing are
square and hexagonal (Fig. 12.2).

Fig. . : Square and hexagonal grids.

1 2 2

Most systems operate in square meshes. It Is possible to simulate a hexagonal grid on a square
grid, however, by means of processing. The hexagonal grid offers the advantage of isotropy since,
unlike the square grid, the distance between the central pixel and its neighbours Is identical in all
directions.

■ Digital image format

Let / be a two-dimensional image function:

/: (x,y) ^ /(x,y)

where /(x,y), an image element with co-ordinates (x,y), is termed a pixel (picture element).The value
of l(x,y) may be a grey level or a colour value for colour images which generally comprises three
components (red, green and blue in the RGB system). The value of a pixel gives information
representative of the image in the cell under consideration. Digital representation of an image in a grid
format is a matrix in which each cell contains the value of a pixel. This type of representation is known
as the raster format (Fig. 12.3). The origin is generally fixed at the top left.
210 Processing of Remote Sensing Data

Continuous domain

Subject

pixel Digital image

Fig. 12.3: Discretisation of information during image digitisation.

M Coding digital values and quantisation of pixel level

As mentioned earlier, the most common digital representation used for images is the matrix
representation. Each image pixel is associated with a cel! of the matrix. Like any computer representation
of digital values, coding of pixels is based on binary numbers. A bit (binary digit) is the elementary unit
of information which can only take the value of lo r 0. For reasons of efficiency, computers process
many bits simultaneously. The bits are thus combined into ‘machine words’ of 8 bits (octet or byte) of
16,32 or 64 bits. Coding used in these machine words facilitates representation of various mathematical
types of numbers. The discrete mode of representation allows description of only a finite number of
values coded with a finite precision (Table 12.1).
An image for which the values have been quantised over various level intervals, coded in an
appropriate number of bits, is shown in Fig. 12.4. Coding over an octet Is most common for grey level
images and provides adequate precision in many cases. Most image-processing systems can manage

Table 1 2 .1 : Coded value domains for various bit numbers for different types of numbers

Number of bits
Number type
1 8 16 32

Integer with no sign [0 ; i ] [0; 255] [0; 65,535] [0;4x10®]

Signed integer X [-128; 127] [-32,768; 32,767] [-2 x 1 0 ®; 2 x 1 0 ®]
 Processing and Interpretation 211

^ Analog

Quantisation

Discrete

2 5 5 d :1 1 1 1 1 1 1 1 b 1 5 d : 1 1 1 1b 7 d : 1 1 1b Id : 1b

Od : 00000000b Od : 0000b Od : 000b Od :0b

Coding in 8 bits Coding in 4 bits Coding in 3 bits Coding in 1 bit

8 4 3
2 = 256 levels 2 = 1 6 levels 2 = 8 levels 2^ = 2 levels

Fig. 12.4: Image quantisation and pixel coding (d: decimal, b: binary).

images in different codes (binary, octet, integer, real) and with different precision levels (1 to 64 bits).
It should also be noted that representation of numbers over machine words differs according to the
computer architecture (microprocessor), particularly in the order of coding used for the self-same
number (high-weight and low-weight octets). Various conventions are employed (Intel/Motorola).
The integer 258, for example. Is too large to be coded in a single octet (Table 12.1): 258/255 =
1*255 + 2 (the remainder of integer division is 2). Two octets are hence necessary to code this integer;
one is known as high-weight octet (1*255) and the other low-weight octet (2).

2 1

00000010b 00000001b in Intel code: high weight octet as primary

1 2

00000001b 00000010b in Motorola code: low weight octet as primary

212 Processing of Remote Sensing Data

M Neighbourhood and connectivity of a digital image

As will be seen later, filtration processing is based on computations for each point of the image. These
computations use not only the value of the pixel perse, but also the value of adjacent pixels.These are
referred to as neighbouring pixels or neighbourhood (Fig. 12.5).

Neighbourhood V4 Neighbourhood Vg

Fig. 12.5: Neighbourhoods V4 with connectivity 4 and Vg with connectivity 8 in a neighbourhood of size 3 x 3
(black: the pixel under analysis; grey: neighbourhood pixels).

The definition of neighbourhood depends on the type of grid used (Fig. 12.2) and the metric
defined for computations of distance In image space. Two types of metrics are commonly employed in
a square grid: distance and distance dg, respectively defining the neighbourhoods V4 with connectivity
4 and Vg with connectivity 8 (Fig.12.5):

0^4 (A. e) ^ Ixg-x^l + lyg -y^l

cJq (A. B) = max (\Xg- x^\, ly g - y^l)

Spatial extent of the neighbourhood is generally given by the side of the window as a number of
pixels. Most often, it consists of an odd number such as 3 x 3, 5 x 5, etc., since centred on the pixel
under consideration. The processing time increases significantly with window size. If the window is
changed from 3 x 3 to 9 x 9, for example, the side of the window is changed three times, but the total
number of pixels Is changed from 9 to 81. Increase in the side of neighbourhood by a factor n increases
the number of pixels and hence the number of computations by

12.1.2 Signal-processing systems: convolution product

H One-dimensional convolution product
Information is extracted from signals by means of signal-processing systems (Fig. 12.6). These systems
are classified, in the same way as signals, into analog, sampled and digital systems.
Let s be an input signal of system Sand fthe output signal (Fig. 12.6).

t(k) = S[s(k)]

Linearity is an important property for signal processing systems. A linear system responds to
linearity criteria with two constants, a and b:

S [a s (k) + b t(k )] = a S [s (k)] + b S [t(k)] (2)

Impulse response g {k, I) of a linear system is defined as follows.

Let s (k) be a digital signal and da unit impulse (eqn 1).
 Processing and Interpretation 213

it4..4t ttiTI
i iff

Signal s
1 System S

Fig. 12.6: Signal-processing system.

t(k) = S [s(k)] = S['Z s(l) d ( /r - /) ] = I s(l) S [d ( k - f) ] (3)

Substituting g (/c, /) = S [d ( k - I) ],

/=+00

(4)
/= _ o o

This equation shows that response t of a linear system to a signal s can be expressed as a
function of impulse response of system g (k, /), where g (k, I) is the response of the system to an
elementary impulse d.
Linear systems that are inyariant by transfer are especially important for image processing (see
sec. 12.1.3).
If t {k) is the response of a linear system S to the signal s (k) and S is invariant by transfer, then
t (/c- /Cq) is the response of S to s (/c - /Cq), where k^ is an integer.
For systems that are invariant by transfer, eqn (4) is expressed as a convolution product as
follows:

l =+CQ
t{k )= ^ s { l) g ( k - l) = s (k )® g (k )^ g (k )® s {k ) (5)
l=-co

— The impulse response is tilted over the co-ordinate axes

— It is shifted by k: g (-1 ) -» g (k -1 ).
— Product s (/) g ( k - l ) is obtained for each sample for a ll/.
— ^These products are added to obtain t(k).

■ Two-dimensional convolution product

Image processing is the analysis of two-dimensional digital signals. The definitions given in the preceding
section are extendable to two-dimensional signals such as images. The discrete two-dimensional
convolution product of finite extent is written as:
214 Processing of Remote Sensing Data

nk^.k2)= ^ 2] Silvl2)g{k^-l^,k2-l2)
/l=0 ¡2=0

f(/Ci,/C2) = s(/fi,/c’2 )® (8)g(/Ci,/C2) (6)

12.1.3 Spatial filtering

ti Definition of digital filter
Digital filtering applies a linear digital system, invariant by transfer, for modifying the frequency distribution
of components of a signal using arithmetic operations of limited precision.
Let / and J be two images and h a convolution mask such that

J {x, y) = h (x, y) o / (x, y) (7)

Equation (7) represents (with an approximate symmetry relative to the origin) the discrete
convolution product introduced by eqn (6).
The operator h is implemented in the form of a matrix. It is referred to as mask, processing
window or filter.

M Normalisation of filter computations

The result of application of the convolution product is not necessarily contained within the interval of
values permissible by image coding. Hence It is necessary in some cases to normalise the result. The
coefficient of normalisation Is generally defined as a function of the weightages of the convolution filter.

H Neighbourhood analysis by ‘moving window’

As mentioned earlier, the discrete convolution product applied to an image during filtering amounts to
making for each point a linear combination of grey levels of its neighbourhood. The coefficients of this
addition are filter values or weightages. They are stored in the matrix of the processing mask. At each
point, a linear combination of the pixel and its neighbours is computed using filter values as coefficients
(Fig. 12.7). The mask is ‘placed’ on the neighbourhood centred on the pixel under analysis and each
coefficient is combined to the pixel located below (Fig. 12.7).
Each pixel is analysed sequentially, covering the image line by line starting from the origin at the
top left up to the last point at the bottom right (Fig. 12.7). The entire operation takes place as though
the mask was ‘moving’ from top to bottom and from left to right passing through each image pixel and
hence it is called a moving window for neighbourhood analysis. It should be noted that algorithms
dedicated to computers with parallel architecture have been developed for this type of processing in
order to reduce computational time. Some linear filters known as separable can be divided into two
successive filters, one along rows and the other along columns. This reduces the number of
computations to be made.
Unlike point analysis, the result of neighbourhood analysis cannot be progressively stored in the
image per se. In fact, the result depends not only on the value of the pixel per se, but also on the values
of its neighbours. As the processing of various points in the Image is sequential, generally from top to
bottom and from left to right, a part of the neighbourhood of a point (position B, Fig. 12.8) will be
modified by computations in preceding positions (position A, Fig. 12.8).
To overcome this problem, the result of computation is generally stored in another Image. The
initial image remains unchanged at any point. These images are often called the source and result
images (Fig. 12.7).
 Processing and Interpretation 215

Convolution m a sk : h

m
Pixel and its 1x20 + 1x5 + 1x7 + 1x10 + 1x2 + 1x20 + 1x50 + 1x32 + 1x28
neighbourhood

Origin

, -
y ▼ Source image : / Resultant image : J
Line during computation

Fig. 12.7: Analysis by ‘moving window’: example of ‘mean’ filter.

3 x 3 processing window

Fig. 12.8: Coverage of neighbourhood zones of pixels.

H Problem of image borders in neighbourhood analysis

As mentioned earlier, processing by filtering is based on a specific computation of a pixel in its
neighbourhood. Digital images are geometrically finite. Definition of neighbourhood for pixels situated
at the borders of the image hence poses problems. In fact, no points exist in a part of the neighbourhood
(position C, Fig. 12.8). Several classical methods exist to tackle this problem (Fig. 12.9):
— B: analyse only those points for which the neighbourhood Is completely defined and replace
the unprocessed border by zeros;
— C: consider that the Image Is surrounded by a boundary V/2 (where V is the size of the
neighbourhood analysed) containing zeros;
216 Processing of Remote Sensing Data

Fig. 12.9: Management of image boundary problem in neighbourhood analysis with a 5 x 5 window. A: Original
Image; B: analysis frame reduced to 2 pixels (neighbourhood/2); C; Enlargement of image by a 2-pixel frame
with 0; D: Enlargement of image by a 2-pixel frame symmetrical to image boundary.

— -D: consider that the Image is surrounded by a border of size \//2, for which the values are the
same as In an image symmetrical to its border. A variant consists of taking values of opposite sign to
those of the image.

H Low-pass and high-pass filters

Filters are referred to as low-pass or high-pass according to the range of frequencies they allow to
pass through. In the case of filters for image analysis, low-pass filters attenuate level variations in the
image and hence smoothen the image, as in an averaging filter used to reduce noise. Contrarlly, high-
pass filters enhance the variations, as in the case of gradient filters used as edge detectors.

12.2 NON-LINEAR FILTERING

Methods of neighbourhood analysis, which cannot be replaced with a convolution product, are also
used in filtering. Such a filtering is called non-linear filtering in contrast to linear filtering. This category
includes various kinds of processing.

12.2.1 Order filtering: median

This filtering is based on analysis of pixels of a neighbourhood not with respect to their spatial disposition,
but as a function of a statistical measure. This measure is established for a classification in increasing
order of grey levels of all pixels of the neighbourhood. The median Is an example of an order filter, for
which the result Is the median value of an ordered distribution of points in the neighbourhood (CD
12.2). For a square window of side A/, M = A/x N pixels, the latter are classified in increasing order of
 Processing and Interpretation 217

grey levels and the position of a point in the distribution forms the order index / [1; M].The median
is the value with index (M + 1)/2.

12.2.2 M orphological filtering

H Mathematical morphology
The concept of mathematical morphology was developed by G. Matheron and J. Serra in 1967 at the
Paris School of Mines (Serra, 1982). The concept of form constitutes the base of analysis in mathematical
morphology. It consists of comparing the objects understudy with an object of known form, referred to
as the structuralelement.The basic operations in mathematical morphology are group transformations
in all-or-none (Coster and Chermant, 1989). In other words. It is based on union, intersection and
inclusion and gives 1 or 0 as the answer.

■ All-or-none group transformations

Mathematical morphology makes use of group relationships (Fig. 12.10), unlike most image-processing
techniques based on linear algebraic relationships (Preteux, 1989). These are the conventional group
operations that use structural elements.
If A and B are two groups, the conventional group operations would be:
— union: A ^ B ]
— intersection: A r\B ;
— complement A ^ :x x
— symmetricdifference A fB = A kjB - A i^B .

Fig. 12.10: Group transformations.

For an object under study, let us consider A, the group of points in the space containing this
object. Let Ebe the structural element of a known geometry. All-or-none group transformation with a
structural element involves ‘moving’ E onto all the positions In the space under consideration and
testing the group relationship between E and A each time. The relationship is either confirmed or not
confirmed. It thus constitutes a test of Boolean nature and hence the name ‘all-or-none’ transformation.

H Basic operations
Let us assume an image A and a structural element E for these basic operations.
218 Processing of Remote Sensing Data

□ Erosion
Erosion is the oldest transformation defined in mathematical morphology. Let the group >4 be defined
In space and E a structural element of known geometric figure, say, a circle. In transformation by
erosion, the origin of the structural element (here, centre of the circle) is placed on each point in the
space and whether E is entirely included in A or not then checked. The group of points satisfying this
criterion forms a new group B, called the erosion of >4 by E (Fig. 12.11) (CD 12.1).
Erosion of >4 by E: >4 © E is group B of points x e A /[x : E^ c A].

Original Eroded Expanded Open Closed

Fig. 12.11; Basic operations of mathematical morphology.

□ Expansion
By analogy, a transformation by expansion is defined In a dual manner. Let a group A be defined in a
space and E be a structural element. Transformation by expansion consists of positioning the
origin of the structural element at the point under consideration and checking whether E has at least
one point in common with A. The group of points satisfying this criterion forms a new group B and is
known as expansion of >4 by E (Fig. 12.11; CD 12.1).
The expansion of >4 by E (>4 © E) is the group B of points x e A /[x : {A^ n £) ^ 0 ]-
Transformations by erosion and by expansion are two dual transform ations relative to
complementation. In fact. If a complementary group of A is eroded, A^ by E, a result identical to
expansion oi A by E is obtained.

A^ © E = /4 © E

□ Opening
Opening of >4by E:>4 0 E= (/4 © E) e Eis an erosion followed by an expansion (Fig. 12.11) (CD 12.1).

□ Closure
Closure oi Aby E: A 9 E = (A® E) © E is an expansion followed by erosion (Fig. 12.11; CD 12.1).

□ Thinning and thickening

Thinning consists of removing a group of points corresponding to a given configuration of neighbourhood.
If {E} is a group of structural elements that gives all configurations of the neighbourhood to be
taken into consideration, thinning of the image A by {E} will be equal to:

/ \ 0 { E } = (...(((^ © E i )© E 2 )0 E 3 )© ... 0 E ^ )

The application of (E) is iterated until there is no change in the Image. At each Iteration, n masks
are applied for each point if the configuration In the image corresponds to that of a mask the point Is
removed from the image, while thickening consists of adding points corresponding to a given
configuration in the same manner.
 Processing and Interpretation 219

□ Skeletisation and trimming

Let be a group and Fits boundary; a point s belongs to skeleton S of /A if the Euclidean distance of
sto Fis reached at least in two distinct points of >4. Trimming consists of thinning using masks that aid
in removing small irregularities of boundaries, i.e., free short lines at an edge.

■ Mathematical morphology in grey levels

As mentioned earlier, SPOT images are grey level (panchromatic) or colour (multispectral) images
which can also be considered as multiple grey level images corresponding to various spectral bands.
The methods of morphological analysis described so far are applicable only to binary Images and
hence only their generalisation to grey level images will be discussed.
The concept of subgraph of a function is Introduced to relate to the group definitions of the preceding
section.
Let / (x) be a one-dimensional grey level function such that

/ = x - > /( x )e R -"

G, is the curve of function /, I.e., the group of points [x, f] e where t= I (x) (Fig. 12.12).The area
under the curve G, determines a group SG, called subgraph of function l(x).

I(x)

G is defined from SG as: G = sup (t: x, t e SG).

As SGj defines a group, the group definitions described earlier can be used for it. For example,
erosion of group SG by a structural element E to give a new group S G j would be:
SG j= S G ,e E
A structural element E ls associated with each point P (x, t) e R^. |n a neighbourhood v defined
by E around P, the function / (x) has a minimum and a maximum. For carrying out an erosion by a
planar structural element E, each point P is attributed the minimum value of / (x) in E: / (x) © E =
minimum {/ (v): v e E). For expansion, the maximum is taken: / (x) © E = maximum {/ (v): v e £}.
Expansion produces a lighter image than the original (CD 12.1) wherein dark details are reduced
or disappear. The result of erosion is a darker image in which lighter details disappear (CD 12.1),
emphasising the dual nature of erosion and expansion. Opening and closure (CD 12.1) are determined
in the same manner as for binary images, combining erosion and expansion (Fig. 12.13).

12.2.3 Homomorphic filtering

Unlike linear systems, non-linear systems are relatively more difficult to analyse and characterise
mathematically. It is difficult to analyse by non-linear methods a signal resulting from multiplication or
220 Processing of Remote Sensing Data

Original Eroded Expanded Open Closed

Fig. 12.13: Basic operations in mathematical morphology In grey levels.

convolution of several signals. Moreover, utilisation of linear filters is based on the assumption, crudely
sometimes, that the signal is effectively linear. Homomorphic filtering results from generalisation of
the theory of linear systems and Is based on the principle of generalised superposition (Kunt, 1981).
It represents generalisation of linear equations (2), for which homomorphic filtering combines linear
and non-linear processing techniques.

12.2.4 Adaptive filtering

Linear filtering has facilitated numerous applications due to simple and powerful mathematical
representation. Nonetheless, use of a linear filter is always a compromise, in terms of efficacy, between
the action of the filter in the zones to correct the image and its undesirable effect on other zones of the
image that need to be preserved. In fact, application of a linear filter is Indifferent for all objects of the
image. Thus, application of a noise reduction filter, such as an averaging filter, will not only have the
desired effect of reducing the impulse noise of the image, but also make the transition zones
corresponding to Image boundaries hazier (Fig. 12.14).
Adaptive filtering Is based on adjusting the structure or weight of the filter according to the situation
encountered at each position In the image. The adaptive mean is an example of adaptive filter designed
to reduce impulse noise without disturbing the boundaries.

Transect Transect

Image affected by Image corrected by a noise

an impulse noise reduction filter (mean 3x3)

Fig. 12.14: Undesirable effect of noise reduction linear filter at the boundaries.
 Processing and Interpretation 221

v/2 v/2

I
2 z
^
m =-v/2 h -v /2
k{x+ l,y+ m )*l{x+ l,y+ m )
J{x,y) = ( 8)
k(x+l,y+m )

1_ if_ I/^ + /, y + rr? - lx, y\< preset_ threshold

k (x + l,y + m) =

where / is the source image, Jth e resultant Image, v the size of neighbourhood and /c a coefficient.
The only difference in eqn (8) compared to a computation of the classical mean is the presence
of the coefficient k. The value of coefficient /ccan be either 1 or 0 and it does not take some pixels of
the neighbourhood into consideration for computation of mean. The coefficient Is calculated for each
position of the neighbourhood and its value directly depends on the gradient between the neighbourhood
pixel and the central pixel. If this gradient Is greater than a preset threshold, k is fixed at zero to
eliminate this neighbouring pixel from computation of the mean. In the boundary zone, only the
pixels that are on the same side of the boundary as the central pixel are used for computation of the
mean.

12.3 NOISE REDUCTION IN IMAGE PREPROCESSING

12.3.1 Noise in images
Grey level variations in Images represent not only the information contained in the scene acquired by
the sensor, but may also be due to disturbances in acquisition, for example. These fluctuations, which
carry no information, are artefacts associated with acquisition process and constitute noise. Generally,
the noise Is by nature spatially uncorrelated with the rest of the image and represented by a higher
frequency signal than the image (Pratt, 1991). Noise is most often considered a random distribution. A
preliminary stage of processing is usually necessary before the analytical phase to remove or at least
reduce the noise.

12.3.2 Linear and non-linear filters for noise reduction

Linear and non-linear low-pass filters are used for preprocessing (Fig. 12.15). Various models more or
less assume the impulse nature of noises, represented by isolated pixels that differ totally from their
neighbours.
Linear and non-linear filtering methods are used for noise reduction In images. As mentioned
earlier, use of a linear filter such as the mean filter (CD 12.2) for noise reduction makes boundaries of
objects hazier (Fig. 12.14). Non-linear filters enable reduction of this effect. Among non-linear filters,
the median filter is least sensitive to Impulse noise, while preserving regular transitions between image
zones. It has the disadvantage of modifying the geometry of image regions (Fig. 12.15). The median
filter (CD 12.2) is very expensive In terms of computation time and the number of computations
exponentially increases with the size of the processing window. Fast algorithms such as the pseudo­
median filter have been proposed (Pratt et al., 1984). The median filter and its derivatives are very
useful for noise reduction. Many other non-linear filters are also employed, among them homomorphic,
adaptive and morphological filters (Fig.12.16).
222 Processing of Remote Sensing Data

Original noise Mean Median Mode

Fig. 12.15: Reduction of salt-and-pepper type impulse noise in a portion of SPOT image
by some low-pass filters.

Fig. 12.16: Noise elimination by a morphological filter using a closure operation.

Some linear methods used for noise reduction are based on analysis in the Fourier domain rather
than the spatial domain. They are more particularly employed for reduction of periodic noise in images
(Pratt, 1991). Such periodic noises may often appear in remote-sensing images acquired with a digital
CCD sensor, for example when a defect in the sensor can produce a line effect.

12.4 EDGE DETECTION

An important field of application of filtering in image processing is edge (contour) detection. Image
analysis systems proposed by various authors are conventionally structured in three levels; low level
for segmentation, intermediate level for constructing approximate geometries and high level for
interpretation. In the segmentation stage the image is divided into several parts corresponding to the
objects present In the Image. Although no universal method of segmentation exists, two approaches
can be identified, viz., region and edge methods, depending on whether data inside regions or at the
 Processing and Interpretation 223

boundaries of regions are used for analysis. An optimal detection of edges is hence necessary for a
correct segmentation of the image under study.

12.4.1 Model of an edge

Let l{x, y) be a function describing an image, which gives amplitude variations in space. Discontinuities
in amplitude (grey levels or colour components) correspond to edges in an image.
These edges are often important indicators of boundaries of objects in images. However, they
may also arise from regions such as shadow zones and noise zones, which contain no objects. An
edge can be characterised by its height H (Fig. 12.17), its slope a relative to the horizontal, its width W
and its orientation O. If H is greater than a certain limit, it can be said that an edge exists. Various types
of edges conventionally identified in digital images are shown in Fig. 12.18. If a = 90°, the edge is
stepped (Fig. 12.18a).

z
À
I

y
Fig. 12.17: General model of an edge.

B I
Fig. 12.18: Step-type edges.
224 Processing of Remote Sensing Data

Generally, this type of edge is observed only in artificial images. In natural images, discretisation
reduces slope and introduces noise; hence stepped edges resemble those in Figs. 12.18 b and c.
When a < 90°, the edge is said to be ramp-like (Fig. 12.19). The grey level profile of a line is given in
Fig. 12.19 a and b. If the width of a line tends to zero, the edge is roof-like (Fig. 12.19 c).

1 1

Fig. 12.19: Ramp- and roof-like edges.

Two methods are commonly employed for edge filtering, viz., differentiation operators and
comparison with an edge model. In the differential approach, a filter is applied for enhancing the grey
level contrast of the Image. A second operation, which may be simple thresholding, is used to select
the pixels for which the differential is considered sufficient to make part of the edge. Two classes of
differentiation operators are identified based on first- and second-order derivatives respectively.

12.4.2 Edge detection by differential filters

As seen earlier, boundaries of objects correspond to zones of steep variations In grey levels of the
image. These zones can be detected by studying the derivatives of the function of an image (Fig.
12.20). The gradient technique for first-order differentiation is based on computation of the gradient V
along two orthogonal directions and addition of these two gradients. It is given by the equation:

. d l(x ,y ) _ dl(x,y) . _
V (x,y) = — T ^ >cos0 + — sinO
dx dy

The amplitude of v is given by: V= + Vy)

where and Vy are the horizontal and vertical gradients respectively. For reasons of efficiency, the
amplitude (norm) of the gradient Is often obtained from the sum of their absolute values:

V = IV;,l + IVyI

The direction of the gradient is given by: 9 = tan~\V;^ / V ^).

Using these definitions in the continuous domain, approximations for discrete data pertaining to
digital images are chosen (Rosenfield and Kak, 1982; Pratt, 1991):

V ^= l{x, y)~ ¡(x+1, y)

 Processing and Interpretation 225

Second derivative

Fig. 12.20: Edges and derivatives.

Vw= / ( > f , y ) - / ( x , y - i )

Several differential operators corresponding to this definition have been proposed:

'0 0 O' '0 -1 O'

0 1 -1
Gradient X = Gradient Y = 0 1 0 Difference (CD 12.3)
0 0 0 _0 0 0_

"0 0 O' '0 -1 O'

1 0 -1 0
Gradient X = Gradient Y= 0 0 Difference separated by a pixel (CD 12.3)
_0 0 0 0 1 0

"0 0 -1 ' '-1 0 O'

0 1 0 Gradient Y = 0 1 0
Gradient X = Roberts (CD 12.3)
0 0 0 _ 0 0 0_

’1 0 -1 ' -1 -1 -1

1 0 -1 0 0 0 Prewitt (CD 12.3)

Gradient X = Gradient Y =
1 0 -1 1 1 1
226 Processing of Remote Sensing Data

"1 0 -T "-1 -2 -f

Gradient X = 2 0 -2 Gradient Y = 0 0 0
Sobel (CD 12.3)
1 0 -1 1 2 1_

The second derivative is given by the Laplacian operator defined as:

J(x,y) = -V ^[l{x,y)]

di^ dr
v r =
dx^ ^ ay2

The Laplacian J is zero if / is constant or changes linearly with amplitude. J changes sign at the
inflection point of / (Fig. 12.20). Hence change of sign is searched. In the discrete domain, depending
on the neighbourhood considered, the Laplacian filter Is given by (CD 12.3):

'0 - 1 O' '-1 -1 -1"

-1 4 -1 -1 8 -1
Lapiacian_V4 = Laplacian_V8 =
0 - 1 0 -1 -1 -1

12.4.3 Edge detection by other filters

a Optimal filter
The Canny filter, for example, uses an analytical method to obtain differentiation operators (Canny, 1986;
Pratt, 1991). It consists of obtaining an optimal differential filter for a given type of edge. The Canny filter
is based on a stepped one-dimensional edge model, of height h^, with a Gaussian-type additive white
noise n, having a standard deviation I (x) = A * h (x) n (x). This technique can be used for two-
dimensional images by applying it In two orthogonal directions. If h is the impulse response of the filter
and / the image of the edge, the response of the filter on the edge is given by the convolution integral:

t+L
H =J I(x )h (x 0 -x )6 x

The image is convoluted with h and the edges are detected at the maximum of l(x)<8>h. h is chosen
so as to satisfy certain criteria for correct detection, correct location and unique response.
For a correct detection, the signal/noise ratio is maximised so as to minimise the probability of
detection of false edges and the probability of non-detection of true edges.

J h{x)dx
Signal where
= ^ S (h )
noise ][h ( x ) fd x
 Processing and Interpretation 227

For a correct location, the distance between the points detected and the centre of the true edge
ought to be minimal.

m
Location = —rr L(h) ’ where L(h) = -r*-
j- h'{x)h being the derivative of/?(x).
dx

To obtain a unique response, there should not be multiple responses for a single edge. If is the
distance between the peaks of gradient in the presence of only noise, it is fixed at a certain proportion
of the width L, i.e., x ^ = k*L. It hence consists of maximising the product S(h)L(h) under the condition
x^ = k*L. No analytical solution exists for this. For large values of x^, the Canny function can be
approximated by the derivative of the Gaussian operator.

■ Conditional filtering
The mean conditional filter Is applied only if the mean of the potential grey levels of a line differs at
least by 5% from the grey level mean of the pixels outside the line. This technique Is derived from
visual perception systems that differentiate the lines compared relative to the base. The median
conditional filter Is derived from the preceding filter. The mean values are replaced by the median
measures. The results are better but the computation time longer.

■ Morphological filters
□ Morphological gradient
Morphological gradient is defined as the difference between the expanded and eroded versions of an
image: G = [ / 0 £ ] - [ / © £ ] . Like the gradient defined in linear filtering, the morphological gradient can
be used for detection of edges In images.

□ Top-hat form: Detection of highways in SPOT images

The IGN is in the process of developing geographic databases (BDCarto, BDTopo, BDRoutiere) for
the territories of France. The amount of information required for preparing such digital maps is high,
viz., several hundred billion octets. Under these conditions, problems of initial collection of data and
updating these databases are of great significance. In this context, remote-sensing images constitute
an important source of information. One of the important elements of these geographic databases is
the network of communication lines, viz., roads, highways, river channels and railways. The images
furnished by SPOT and especially SPOT 5 have a sufficient resolution enabling detection of such
networks. Many methods based on edge detection or, more specifically, detection of lines can be used
for this purpose (Gilliot, 1994). in particular, a mathematical morphological approach based on the
top-hat shaped operator is useful.
The top-hat form is defined as the difference between the Initial Image and Its opening by E: J =
I - [(/© E) © E], followed by thresholding (Fig. 12.21). During opening by structural element E, grey
level peaks whose width is less than that of Eare eliminated. Valleys are detected in a similar manner
using the difference with closed image: J = / - [ ( / © E ) ® E].Top-hat transformation enables selection
of grey level peaks not only according to their width, but also according to their height, because of
thresholding:

J (x) = {x :l(x ) - [{/ (x) © E ) © E ] > Threshold}

The term top-hat shape Is used since the transformation preserves only the zones that enter into
a cylinder whose diameter corresponds to the size of the structural element and the height corresponds
to the threshold.
228 Processing of Remote Sensing Data

Fig. 12.21: Portion of highway network in SPOT image (a) and 3D view of difference between (a) and its opening
in (b), and result after thresholding in (c).

A morphological skeletisation is applied on the binary image resulting from the top-hat form
filtering and aids in obtaining edges of unitary size. A trimming operation eliminates small segments
with a free edge.

12.4.4 Closure of edges

Edge detection by filtering most often results in discontinuous boundaries due to noise and imperfection
of the filtration operators. In fact, application of a differentia! operator provides in the result the intensity
of a potential edge. The choice of a proper threshold value will transform the amplitude of a gradient
into a Boolean result, representing the presence or absence of an edge. This stage of thresholding
(Fig. 12.22), depending on the threshold chosen, furnishes a more or less continuous or more or less
noisy edge.
A stage of so-called edge closure is hence essential for reconstructing the initial boundary
(Gonzalez and Woods, 1992). Various methods based on graphic exploration techniques are used in
this stage.

12.4.5 Other edge detection methods

H Parametric methods based on an edge model
These methods make use of a priori knowledge about boundaries of simple geometric form that can
be represented by a simple mathematical function. Hough proposed a method for detection of lines,
which is known as the Hough transformation (Ballard and Brown, 1982; Maitre, 1985; Wang, 1990).
The problem of detecting collinear points in the image (Cartesian [y, x] space) is replaced by a problem
of detecting ‘clusters’ of points in a parametric [p, 0] space (Flg.12.23).The relationship between the
two spaces is given by the equation:

p = X cos 0 + y sin 0
 Processing and Interpretation 229

i
255 255 255
r 255
Different threshold values

Images of binary edges

Fig. 12.22: Thresholding the amplitude of a gradient with various values of threshold for obtaining an
image of boundaries.

An infinite number of straight lines can be drawn through each point (Xj, yj), all of which satisfy the
equation p = x^ cos 0 + sin e . The space defined by p and 0 is often known as parametric space (Fig.
12.23). A straight line in the Cartesian space corresponds to a point in the parametric space. A point
in the Cartesian space corresponds to a curve in the parametric space (Fig. 12.24).This curve represents
all possible straight lines that can pass through this point. In this way all possible straight lines or
candidates at each point are determined and this stage can be compared to a vote. Search for local
maxima in the battery of lines is used to determine those straight lines which have had the highest
votes and which hence are present in the Image. Thus roads present in the SPOT image are
reconstructed. This reconstruction of a road network calls for a knowledge-based system (Gilliot et al.,
1993).

H Markovian methods
Stochastic methods have mainly been used for textural classification of Images. The Markovian fields
constitute a family of stochastic models whose definition is essentially based on the concept of
230 Processing of Remote Sensing Data

Fig. 12.23: Hough transformation.

Fig. 12.24: Parametric space or battery of lines.

neighbourhood which, as seen earlier, is fundamental for detection of boundaries and, more generally,
in analysis of images. Methods based on Markovian fields have thus been developed for edge detection
(Cocquerez and Phillip, 1995).

a Methods of active edges or snakes

This method of image segmentation was proposed by Kass in 1988 (Kass et al., 1988). The boundary
under investigation is created from a curve evolving under constraints for gradually reconstructing its
exact form. The initial curve is started in the proximity of the boundary and then deformed according to
an iterative process, which is stopped depending on a convergence criterion generally related to the
stability of the curve during iterations. The active edge is modelled by a curve C:

C = {p {s ,ty ,s e [a ,b ls e [O J ]}
 Processing and Interpretation 231

where p is a point, s its curvilinear abscissa, a and b the extremities of the curve and f the iteration
time.
An energy function E{C) is computed at each iteration for the curve C. We search for the curve C
which would minimise this energy so as to provide a good approximation to the boundary in the image.
For this, E is expressed as a function of the curve, the image and interactions between the two in the
vicinity of the curve (Cocquerez and Phillip, 1995):

E(C) ^ in te r n a l ( C ) + ^ e x te rn a l ( C ) + ^image (C)

Various categories of snake are identified depending on the constraints applied on the curve
(fixed or free extremities, closed curve). In addition to continuous formulations, discrete formulations
of snakes also exist which can be readily implemented. The boundary is described by a suite of points
that represent apexes of a polygon approaching it. Conventional methods of edge extraction consist of
two phases: detection, which provides often an incomplete edge, followed by closure of edges. The
method of active edges, making use of general information, aids in directly obtaining the closed and
regular edges with a good robustness vis-à-vis anomalies in the Image (Cloppet-Ollva, 1996).

■ Neural network methods

Methods known as neuromimetic, based on mechanisms of biological vision, attempt to Implement
systems of artificial vision in the form of neural networks capable of edge detection.
Drawing from the well-known visual illusions which make edges that are not drawn appear (Fig.
12.25), Grossberg (Grossberg and MIngolla, 1985) suggested the existence of virtual edges in the
mechanism of biological vision. This consists of a low-level, rapid and attentive mechanism that allows
strengthening of edges of objects in the vision process. Artificial neural networks operating on similar
mechanisms have been proposed. Various possible boundaries depending on various orientations
are first detected using directional gradients and constitute the state of entry of neurons. From this
initial state, continuation of boundaries is obtained through a mechanism of activation and inhibition
between neighbouring neurons. Rules used for continuation of edges are simple geometric laws detected
In biological vision systems (in the direction of a boundary, perpendicular to an edge). Several
assumptions are made and a comparison mechanism is used to identify the most probable boundary.

r ^

V^
Fig. 12.25: Kamizsa square: a visual illusion showing virtual boundaries of a square that is not drawn.

12.4.6 Evaluation of edge detection operators

As can be seen throughout section 12.4, several edge detection operators exist. It may be necessary
to evaluate their relative performance for a given problem. For this purpose, the results of various
operators are compared with reference data. The reference may be taken from a map or another
processing or may be the result of a manual segmentation of image. An example pertaining to detection
of linear elements In SPOT images for automatic identification of communication lines (Gilllot, 1994) is
presented below. A manual segmentation of the main communication lines was done such that it can
232 Processing of Remote Sensing Data

be used as a reference. Let Cbe the group of points of the edges comprising the communication lines
and C its complementary group of points where no edge exists (Fig. 12.26).
Let D be the detection of an edge by an operator.
p {D/C} is the conditional probability of detection of an edge, knowing that this edge effectively
exists. _
p (D /C ) is the conditional probability of detection of an edge knowing that in fact no edge exists.
It represents the probability of false detection. The various probabilities can hence be defined as:
Pp the probability that an operator makes an exact detection;
P p ’the probability that an operator makes a false detection;

PE = f p ( D / C ) 8 s ; Pp = r p(D /C )ds
Js

where s Is the threshold value.

Fig. 12.26: Cand C , groups of points of edges and non-edges respectively.

p (D/C) and p (D/C) are determined for <each of these images and for different threshold values.
Then p (D/C) Is traced as a function of p (D/C) for various operators (Fig. 12.27).
The topmost curve in the graph, which corresponds to top-hat shape, is the one that possesses
the highest probability in the left zone of the graph where Pp is small, and confirms the visual
analysis of the significance of top-hat shape as a detector of lines. In fact, the right-hand side of the
graph corresponds to high probability of false detection and hence is not interesting.

P(0/C)

-Deriche
-Prewitt

-Laplaces
-Laplace4
-Sobel
-Gradient
-Top hat

Fig. 12.27: Relationship between p(D/C) and p(D/C ) for various edge detection operators.
 Processing and Interpretation 233

12.5 CONCLUSION
Digital filtering, which is very useful for preprocessing of images, also finds important application for
segmentation of remote-sensing data using boundaries. The methods based on linear or non-linear
mathematical tools call for global or local approaches (adaptive methods). The most efficient methods
simultaneously combine general as well as local criteria. As mentioned earlier, in ‘natural’ images
such as remote sensing, the result of filtering rarely corresponds to all boundaries of image objects.
The large variability of the signal of these data, the noise inherent in the acquisition process and
regulation of operators (thresholds) can explain the imperfect results. Most often, the segmentation
method employed uses an operator in combination with others In sequence of processing methods.
Less sequential approaches such as dynamic programming, neural networks, snakes, etc. have also
come into vogue recently.

References
Ballard DH, Brown CM. 1982. Computer Vision. Prentice-Hall, 515 pp.
Canny J. 1986. A computational approach to edge detection, IEEE Pattern Analysis and Machine Intelligence.
PAMI,8(6):679-698.
Cloppet-Oliva F. 1996. Analyse d’images de cultures cellulaires obtenues par microscopie optique: Application à
des images de neuroblastomes de souris. Thèse de l’Université René Descartes, Paris, 5, 264 pp.
Cocquerez J-P, Philipp S. 1995. Analyse d’images: filtrage et segmentation. Masson, 457 pp.
Coster M, Chermant J-L, 1989. Précis d’analyse d’images. Presses du CNRS, 560 pp.
Gilliot J-M. 1994. Traitement et interprétation d’images satellitaires SPOT: Application à l’extraction des voies de
communication. Thèse de l’Université René Descartes, Paris, 5, 207 pp.
Gilliot J-M, Stamon G, Le Men H. 1993. A knowledge-based system in image processing for communication networks
study in aerial images, a tool for cartography automation. IEEE Systems: Man and Cybernetic. SMC-93, Le
Touquet, pp. 77-82.
Gonzalez R, Woods R. 1992. Digital Image Processing. Addison Wesley Publ., 716 pp.
Grossberg S, Mingolla E. 1985. Neural dynamics of form perception: boundary completion, illusory figures and
neon color spreading. Psychological Review, 92 (2): 173-211.
Haralick RM, Shapiro LG. 1991. Glossary of computer vision terms. Pattern Recognition, 24 (1 ): 69-93.
Kass M, Within D, Terzopoulos D. 1988. Snakes: active contour model. Int. J. Computer Vision, 1:312-331.
Kunt M. 1981. Traitement numérique des signaux. Dunod, 402 pp.
Larousse. 1998. Le petit Larousse illustré.
Maître H. 1985. Un panaroma de la transformée de Hough. Traitement du signal, 2 (4): 305-317.
Pratt W. 1991. Digital Image Processing. Wiley-Interscience, 2nd ed., 698 pp.
Pratt W, Cooper T, Kabir 1.1984. Pseudomedian filter. Proc. SPIE Conf., Los Angeles, CA January.
Préteux F. 1989. Éléments de morphologie mathématique: Approche fonctionnelle pour images numériques. Cours
de DESS.
Rosenfeld, Kak. 1982. Digital Picture Processing. Academie Press, London-NY.
Serra J. 1982. Image Analysis and Mathematical Morphology. Academie Press, London-NY.
Wang F. 1990. Improving remote sensing image analysis through fuzzy information representation. Photogrammetric
Engineering and Remote Sensing, 56 (8): 1163-1169.
 13
Geometric Transformation of
Remote-sensing Images
Geometric deformations develop between the scene observed by the satellite sensors and the raw
image received at the receiving stations. These deformations need to be corrected in order to identify
the geographic reality on the ground. Remote-sensing data are often used for analysing the natural
media. They are hence interpreted and processed with the objective of producing maps that provide
spatial Information regarding the problem at hand. It is useful to make this interpretation amenable for
geometric superposition on other maps in order to compare or utilise them conjunctively. For this
purpose, a geometric correction is applied either before interpretation so that data from various bands
can be used, such as in the case of diachronic analysis, or after interpretation for making statistical
comparisons or to integrate them in a geographic information system.

13.1 METHODS OF GEOMETRIC CORRECTION

13.1.1 Causes of geometric deformation
One type of deformation arises from errors generated during Image acquisition. These errors can be
due to platform Instability (rolling, pitch) leading to a drift relative to the theoretical orbit of the satellite
(see CD: SPOT system) or due to aberrations of the sensor resulting in non-linear distortions of the
image (somersault effect). Knowledge of sensor operation and use of precise flight path parameters of
the satellite facilitate corrections for these errors in part. These corrections are generally carried out by
image preprocessing centres.
Another type of geometric deformation is a direct consequence of the curvature and topography
of the Earth. The image, which by nature is two-dimensional, is only a planar projection of a three-
dimensional view of the zone of observation. In practice, these complex deformations cannot be replaced
by simple geometric transformations.

13.1.2 Parametric and interpolation methods

Two main categories of methods used for correction of geometric deformations of remote-sensing
Images can be distinguished (Teissier and Stamon, 1994). These are: parametric methods and
Interpolation methods. The former are based on modelling the phenomena that produce deformations.
The latter are not concerned with mechanisms of origin of distortions but only assume that these
deformations can be determined by means of a polynomial approximation. Parametric methods
 Processing and Interpretation 235

necessitate knowledge of the operative mechanisms and their parameters. They do not directly use
the image and are specific to each sensor-platform pair. These methods are complex to employ and
hence much simpler interpolation methods are mostly used. They are more general than parametric
methods, requiring no information about the sensor-platform system, and essentially based on the
geometry of the image. Only these methods are described below. Intermediate methods, essentially
polynomial but which also partially take the scene parameters into consideration, are also used.

13.1.3 Global or local modelling of interpolation methods

Depending on the mathematical model used, interpolation methods are of two types: methods that
correct for global deformations and those that correct for local deformations of the image. In the first
type of methods, a unique mathematical model is determined and uniformly applied to the entire
image. In the second, several sub-models are applied locally to various regions of the image. Polynomial
approximation methods are most commonly employed but provide corrections only for global distortions.
Transformations that can be replaced by these simple geometric operations or a linear combination of
these operations are called linear transformations of the first degree or order 1 (formulated by a first-
degree polynomial). More complex transformations, non-linear, are modelled by second-degree or
higher-order polynomials.
The method based on spatial modelling using finite elements is the simplest of the local methods.
It makes use of a regular division of the image space into cells (tessellations) and generally involves
a linear Interpolation Inside triangles. More recent methods, such as the multiquadric (MQ), have
proven more efficient for processing stronger local distortions. They are based, in fact, on a global
model that integrates local distortions.
Satellite Images are generally characterised by geometric distortions that progressively vary in
the image. Such distortions can be analysed by a global model. That is why the polynomial models
most often give satisfactory results. Aerial photographs are subjected to additional distortions such as
vibrations, speed variations and topographic effects. In such cases local methods have to be employed.

13.1.4 Direct and inverse transformations

Digital images are defined in a discrete space. The geometric transformations generally employed are
based on Euclidean geometry and are defined in a Cartesian space using real co-ordinates. This
involves an application oi The co-ordinates of the transformed image can hence be real in
some cases.They ought to be converted into integers for assignment to a digital image. This conversion,
generally to the nearest integer, may lead through rounding off to a situation wherein some addresses
in the final image are not covered.The final image will hence comprise gaps (Pratt, 1991). To overcome
this problem, inverse transformation is often employed. For each address in the final image, the nearest
address in the initial Image is computed by inversion of the transformation equation. Coverage of all
pixels in the final image is thus ensured.

13.1.5 Interpolation in geometric transformation

Creation of a geometrically corrected image by an interpolation method implies two types of
interpolations, viz., geometric and radiometric. Geometric Interpolation is used to model geometric
transformations and enables computation of co-ordinates for each pixel in the rectified image. For
each position In the rectified image, the value to be assigned to the pixel has to be determined and this
Is done by the second interpolation, viz., radiometric.
236 Processing of Remote Sensing Data

13.2 INTERPOLATION METHODS

13.2.1 Radiometric interpolation
utilisation of inverse transformation enables analysis of ail the pixels in the transformed image, but the
address obtained in the source image is no longer necessarily integer. It is hence necessary to determine
the method by which the value of the pixel In the transformed image Is interpolated from the nearest
pixels in the source image.

■ Interpolation to the nearest neighbour

From the address computed by the inverse transform, the value of the pixel can be determined by
taking the value of the nearest address in the image source. This is the simplest interpolation method,
known as nearest neighbour interpolation (Fig. 13.1). Although this method does not provide a smooth
image, it is fast and offers the advantage of preserving the initial value of the pixel (Fig. 13.4). This
property is often appreciated by thematic specialists who attempt to model the characteristics of
objects from their radiometric properties.

Fig. 13.1: Interpolation by nearest neighbour method. For the computed address (20.3, 51.4), the nearest
address in the source image (20, 51) is determined. The value of this pixel (200) is used in the resultant image.

■ Bilinear interpolation
In this method, the value of the pixel Is computed as the weighted average of the values of the four
nearest pixels. The nearest pixel has the highest weightage for computation of the value. It thus consists
of a bilinear interpolation which can be divided into two successive linear interpolations in two orthogonal
directions (X, V) in the image. Let us consider the example of Fig. 13.1 and shift the top left pixel of co­
ordinates (20, 51) to (0, 0) (Fig. 13.2). The analysis is carried out in two stages. Four values are
Interpolated, two at a time, along X-directlon in the plane VOX and then the two intermediate values
between them along /-direction in the \/0 /plane.T husinthe l/OXplane, two straight lines represented
by the equation V= a x X + b can be defined as follows:
200 = a1 X 0 + b^ b^= 200
187 = a1 X 1 + b1 => a1 = 1 8 7 -2 0 0 = - 1 3
Vi^ = - 1 3 x 0 . 3 + 200 = 196.1
190 = a 2 x 0 + b2 52 = 190
170 = a2 X 1 + b2 => a2 = 1 7 0 -1 9 0 = - 2 0
\//2 = - 2 0 x 0 . 3 + 190 \//2=184
 Processing and Interpretation 237
y (Value of pixel)

200

-►X

Fig. 13.2: Bilinear interpolation.

The linear equation relating Vi\ and V/2 in the plane VOY\s then considered using the equation V
= a X Y+ b.
Vi\ = 196.1 = a 3 x O + bS b3 = 196.1
V/2 = 184 = a 3 x 1 +b3 a3 = 184-196.1 = -1 2 .1
W = -1 2 .1 X 0.4 + 196.1 Vi= 191.26

The value 191 is assigned to the resultant pixel. This method is slower than the nearest neighbour
method and the radiometric values are not preserved. However, the ‘staircase’ effect of boundaries in
the preceding method is reduced {anthaliasing effect) and a better visual display with more continuous
transitions is obtained (Fig. 13.4).

H Bicubic interpolation and convolution method

The preceding method, defined with a 2 x 2 neighbourhood, can be extended to larger neighbourhood
sizes. Nonetheless, due to complexity of algorithms and computational speed, generally the
neighbourhood size is limited to less than or equal to 4 x 4. For larger sizes, the assumption of linearity
is not valid and in such cases cubic Interpolation functions, such as cubic B-spline, are applied.
Convolution methods are useful for some geometric analyses (see Chapter 12). Especially In the case
of enlargement operations, the convolution method is an efficient means of making interpolations (Fig.
13.4).

13.2.2 G eom etric interpolation

■ Global models: polynomial methods
□ First-order geometric transformations

Translation
Translation of an Image relative to origin is done by the equations:
238 Processing of Remote Sensing Data

yj = y + A y ( 1)

where {Xp is the point shifted from the source point (x, y) by A^in xa nd in y.

Symmetry
If symmetry with respect to any axis is defined, operations relative to a vertical axis or a horizontal axis
are most often employed. In these specific cases, as the axis coincides with a row or column of pixels,
the symmetry simply becomes an exchange of values between two addresses of the image (Fig.
13.3). For a symmetry of horizontal axis, defined by y = S, the relation is:

Xg = X

ys = 2 S - y (2)

Axis of symmetry

Original image Symmetry/ horizontal Symmetry / vertical

Fig. 13.3: Symmetry relative to a horizontal or vertical axis.

Homothety
Scale changes of the Image can be obtained by means of multiplication factors:

h-
(3)
Yh = ^yY

If the multiplication factor is greater than one, an enlargement of the image is obtained (Fig. 13.4).
If It is less than one, the image is reduced. As emphasised earlier, interpolation will be necessary for
determining the values of the initial Image.
In the case of enlargement by an integral factor, the convolution method can be used.
For enlargement by a factor of two, for example, the image is at first transferred into an Image
twofold larger in Xand yco-ordinates, wherein each pixel Is separated from its neighbours in row and
in column by zeros (Fig. 13.5). The image is then filtered with a convolution mask to make interpolation
for replacing the zeros. For an interpolation to the nearest neighbour, the following convolution mask
can be used:

1 1

1 1
 Processing and Interpretation 239

Source image 2 X enlarged image; 2 X enlarged image;

interpolation to the bilinear interpolation
nearest neighbour

Fig. 13.4: Image enlargement over an agricultural plot extracted from SPOT image of Brienne.

E a

Original image Enlarged image Image after convolution

Fig. 13.5: Enlargement by convolution with interpolation to the nearest neighbour.

Based on the same principle, other convolution masks are used for higher-order interpolations.

Rotation
Equations for rotation with respect to origin are defined as:

= x c o s 0 - y s in 0
Yf. = xsin0 + ycos0 (4)

A rotation relative to any arbitrary point is readily resolved by combining the rotation with a
translation. The size of the destination image ought to be larger so as to contain the entire zone after
rotation (Fig. 13.6).
For example, for a rotation of 45°, the size of the Image should be increased by a factor of the
square root of two. For some particular angles (90°, 180°), rotation is simply replaced by a symmetry
and corresponds to a simple exchange between two addresses.
240 Processing of Remote Sensing Data

Original image Rotation by 20° Rotation by 20°

in same image in enlarged image

Fig. 13.6: Rotation of images.

Composition of first-order transformations

The preceding geometric operations, viz., translation, rotation and homothetic conversion can be
combined for defining more complex transformations. Complex transformations can be defined in the
form of a linear algebraic combination of elementary operators. It should be noted that these operations
are not commutative and hence the order in which various operators are applied is important. For
example, a transformation T 1, comprising a translation followed by a rotation (Fig. 13.7), will not always
give the same result (depending on the parameters of operations) as a transformation T2 consisting of
the same rotation followed by the same translation.

First-degree polynomial modelling of any linear geometric transformation

While the operations are not commutative, the parameters can however be modified to obtain the
same result. In the preceding example (Fig. 13.7), use of another translation (TR2) allows defining a

Rotation 0i

Transformation T1

Translation TRI

Transformation T2

Translation TR2

Transformation T3

Fig. 13.7: Non-commutative nature of compositions of geometric transformations: example of composition of a

translation and 20° rotation with respect to the centre of the Image.
 Processing and Interpretation 241

transformation T3 equivalent to T1. Any linear transformation can be replaced by a succession of

three basic geometric operations such as translation homothety rotation. Thus by linear
combination, a transform ation matrix can be established, using which any linear geometric
transformation can be determined. Substituting x - Xu in eqn (4) and by successive substitutions in
eqns (3) and (1), we get:

iX = X/7 COS0 - Yi^ sin0 '= cos 0 - (ay y^) sin 0

[Y = X/7sin0 + cos0 = (a;fX^) sin 0 + (ayy^) cos 0

X = a;f(x + A ;f)c o s 0 -a y (y +A y )sln0

Y = a^j.(x +A;^)sin0 + ay(y +A y)cos0

X = (Bx cos 0)x - (By sin0)y + (Bx cos 0A;^ - By sin0Ay)

Y = (Bx cos 0)x + (ay cos0)y + (a;f sin0A;f + ay cos0Ay)

X = Cx-|X+ Cy2y ^3
Y = C4X + C^y + Cq (5)

Any linear transformation can thus be established by means of this single equation defined by six
constants (3 in X and 3 in V).

3D-visualisation of remote-sensing images

A remote-sensing image Is a view of a three-dimensional scene. If altitude of the scene at every point
is known, an image representing perspective view of the ground can be created. This view can be
especially useful to replace the image in its topographic context.
Altimeter data is provided In digital format by a digital elevation model (DEM). The image needs to
be geographically calibrated over the DEM. Geometric transformation that provides this pseudo-3D
view consists of computation of projection on a plane. The DEM defines a surface given by the equation
Z = f{X , Y) (Fig. 13.8a), where Z is the altitude at every point (X, V). A rectangular network is thus
defined. By connecting the points in space with one another by straight-line segments, facets are
determined (Fig. 13.8b). An approximation of the real surface is obtained by a combination of facets
(Fig. 13.8c). For two-dimensional representation of this surface, it is projected onto a plane. Various

Fig. 13.8: Example of modelling a surface in space by facets.

242 Processing of Remote Sensing Data

projections can be used for this representation and cavaliers perspective is one of the simplest
(Rousselet, 1985). Projection of a point (x, y, z) onto a point {u, v) on the plane of projection (Fig. 13.9)
is given by the equation:

u = x + ky cosa
V = z + kyslna (6)

A projection plane has to be chosen for this purpose. Any segment contained in this plane is
represented as the true magnitude. Any segment perpendicular to the plane of projection is represented
by a segment making an angle a with the horizontal, a is the angle of flight. The length of segments is
balanced by a multiplication coefficient lying between 0 and 1. This is known as the coefficient of
reduction of perspective. An additional multiplication coefficient may be used for varying the scale and
this is known as the scale factor. Intensity at each point of projection may represent the altitude or a
combination of altitude with an image (Fig. 13.10). The image is then said to be draped or ‘mapped’
over the DEM.
This intensity can be regulated at will according to the desired output effects, such as shadows as
a function of the sun’s position. Managing the problem of concealed parts is an important point for
obtaining a realistic output. As a matter of fact, some points of the scene are hidden by other points
and hence the method of image creation should be able to manage masking of concealed points.

Digital Elevation Model in 3D view SPOT image draped over 3D DEM

Fig. 13.10: Pseudo-3D visualisation of a SPOT image.

 Processing and Interpretation 243

A simple software known as ‘painter’s algorithm’ is sometimes used for this purpose. Masking of
concealed points is entirely managed by the order of display of points. It consists of displaying the
points according to their distance from the observer, with the farthest points displayed first. When a
nearer point is drawn, if it is higher in zthan the points behind, it masks them by display over them.
Three-dimensional animations can be created using this model. A trajectory in space is defined
and an image corresponding to the projection computed at each point. The system thus carries out a
sequence of animations, image by image. Computational time for each image may be relatively long
depending on the image size and desired output effects. However, such animations can be achieved
almost in real time on the latest computers equipped with 3D graphics and large live memory. This
enables interactive exploration of the zone under study in 3D space.

□ Second-order geometric transformations

Complex non-linear geometric deformations
Most often, distortions in remote-sensing images cannot be restored by simple, i.e., linear, geometric
transformations. Complex variations In the platform’s trajectory and sensor response lead to non­
linear distortions of the image (Fig. 13.11).

Fig. 13.11 ; Distortion (somersault effect) due to a sensor aberration.

Conditions of image acquisition and nature of terrain give rise to complex image distortions.
Depending on the platform altitude and angle of view (see Chapter 2), deformations can be large or
small. Images obtained by satellites, orbiting at high altitude with a small field of view, provide a better
geometric quality than aerial photographs (Jensen, 1986). When one Image Is compared with another
image or a map, it must necessarily be geometrically deformed for superposing on the latter. An
acquisition of a scene is in effect a projection on a plane. The large number of parameters involved in
data acquisition and the geometry of the terrain observed (Earth’s curvature and topography) make
the projection complex (non-linear deformation) and highly variable depending on the image. It is
difficult to mathematically reconstruct it from acquisition parameters. This geometric transformation
can be approximately estimated by a polynomial modelling by means of a second- or higher-order
polynomial.

Second-degree polynomial modelling

A second-order polynomial is used for modelling such complex transformations:

— C-jX^ + ^2^^ ^ C^xy + C4X + C^y + Cq

(7)
I = CjX^ + Cgy^ + CgXy + C-^q X + C-j -jj/ + C12

As mentioned earlier, in modelling first-order deformations, the coefficients of the polynomial can
be directly related to the elementary geometric distortions Involved (rotation, translation, homothety).
This is not so in the case of second-order modelling and no physical meaning can be given to each
244 Processing of Remote Sensing Data

coefficient. Transformation is not continuous and the polynomial equation provides an approximate
estimate. The coefficients are so computed as to minimise the mean quadratic error between a set of
image points and a set of geographically similar points used as references. These reference points are
generally known as Ground Control Points or GCP.They are obtained from a spatial reference dataset,
such as a map or another Image, which represents the undistorted image. An equation of an order
higher than two can also be used but inversion of the relation becomes more complex.

□ Conclusion on polynomial methods

Depending on the number of points used for solving the polynomial equation, the method is referred to
as either interpolation or approximation. If the number of control points Is strictly equal to the number
of unknowns in the polynomial expression, the system of equations Is exactly determined and termed
interpolation. In this case, the position of the control points computed by the model is the same as that
in the reference image. There Is no residual error since the model is exactly determined on the control
points. If the number of control points is greater than the number of unknowns, the polynomial coefficients
are approximated using various points. The model no longer corresponds exactly to the control points
and a residual error appears at every point relative to the reference position. This deviation is an
Indicator of the quality of the model.
A polynomial approximation of an order higher than one may give rise to undesirable image
distortions (such as stretching, compression). To overcome these problems associated with
extrapolation, the control points should be chosen all around the zone to be corrected. For a n-degree
polynomial, the number of control points required to solve the system of equations Is (/7+1)(n+2)/2.
The minimum number of control points hence is 3,6 and 10 for first, second and third degree polynomials.
For example, in the case of a first-degree polynomial for which 6 constants are to be determined, three
control points allow formulation of 6 equations (3 in X and 3 in V), sufficient for solving the system.

Ü Local models
Polynomial approximation provides a global model and hence corrects deformations applicable only
to the entire image. In the case of Images with large local distortions, a rectification technique based
on use of an image grid ensures better results.

□ Finite element modelling

In image processing the finite element method is mostly used as linear interpolation in a triangular
grid. The grid divides the image into elementary regions (finite elements) which correlate the source
image with the reference image. A local distortion model is applied region by region. Division of image
(gridding) can be done automatically, using Delaunay triangulation (Fig. 13.12) for example (Worboys,
1995). A grid covers the surface under analysis by an arrangement of non-overlapping polygons.
Grids can be distinguished as regular (repetition of the same elementary polygon) and irregular.
Delaunay triangulation and the Voronoi diagram (Thiessen polygons) are the most commonly
used tessellations. A triangular grid can be created using control points. Triangles are so constructed

Fig. 13.12; Delaunay triangulation.

 Processing and Interpretation 245

that the circle circumscribes a triangle and contains nodes of only this triangle. The triangles should
be as regular as possible. Polynomial coefficients are determined for each triangle and rectification is
applied to the image points, triangle by triangle. Triangulation can also be fixed to correspond to the
known zones of deformation. Higher-order polynomial approximations are also used. This method is
well suited for locally linear distortions.

□ Multiquadric functions
Multiquadric functions were introduced by Hardy (1971) and are used for image rectification. The
hypothesis assumes that a mathematical surface and more generally any surface can be approximated
to a given precision by a sum of mathematically defined elementary surfaces, in particular those of
quadratic form (Hardy, 1977). These methods are intermediate between global and local methods. As
a matter of fact, they use a global model that takes local variations into consideration, by integrating a
term which is computed between the point to be relocated and the control points (Fogel, 1996):

X = ^ a j^ {x - X if + { y - y jf + R ^ + P^{x,y)
/=1

y=2 N
b j^ ix - x , f +(y- y if +fl 2 +
-----------------------------------
y) (8 )

/= 1

where (x, y) is the point under consideration. A/the number of control points, [Xj, y) control points, a,,
and byConstants specific for each control point and and polynomial terms which are not always used.
It may be noted that { x - x f - + { y - y f - in the main term of eqn (8) represents the distance between
the point under study and the control points. The optimal value of the term depends on the type of
distortion. It may also depend on the distance between the control points or their scatter, for example.

= 0.6 min ^(Xf - X jf + (y/ - y j f V /, y e [1...A/] for example, depends on the minimum distance
between the control points. can also be replaced by a set of Py^, one for each control point.
These methods are found to be very efficient and Hardy later (1990) cited nearly a hundred
references regarding their use for rectification of remote-sensing images.

13.2.3 C onclusion
Polynomial methods based on global modelling of deformations are well suited for satellite Images
exhibiting mainly global distortions. Contrarily, in the case of aerial photographs, local distortions are
often more severe and the orthophotographic approach becomes necessary. Methods using division
of the image by gridding (tessellation) are good for local linear distortions. This is the type of deformation
generally observed in scanned images, comprising distortions due to folding.

13.3 IMAGE RECTIFICATION

13.3.1 General problem
Application of the methods described in the preceding section for remote-sensing images corresponds
to the same general problem, viz., placing an image in spatial conformity with other spatial data.
246 Processing of Remote Sensing Data

These methods include relocating, rectification, placing in conformity, georeferencing (relative to a

map), warping and stretching.

13«3.2 Systems and data for rectification

M Systems of geometric correction
Various processing systems are used for applying geometric corrections.

□ Image enhancement systems

Most image enhancement systems (such as Photoshop®, Image’in®) include simple geometric
transformation functions. Two modes of making these transformations are currently in vogue: one
assigns digital values to the transformation parameters (ex: angle of rotation) and the other interactively
carries them out using tools that enable moving, turning and enlarging the image on the computer
monitor. A visualisation mode called the ‘transparency’ mode is also quite common. It allows viewing
another Image ‘below’ the image to be deformed. The image can also be visually adjusted over a
reference image (by linear deformation).

□ Image processing systems

Image processing systems such as Visilog® (Noesis, 1997), Khoros® (Khoros, 1997) contain an
additional module of deformation (warping) by control points which can be used for rectification of
remote-sensing data.

□ Remote-sensing systems and geographic information systems

Special processing programs for remote-sensing data, such as Erdas® (Erdas, 1997), ER/Mapper®
(ER/Mapper, 1997), TeraVue® (Teravue, 1997), generally have a versatile module that allows
manipulation of raw images, interactive choice of control points, working with cartographic co-ordinate
systems, displaying cartographic data, assembling images (mosaic), viewing the image in three
dimensions using DEM, etc. Some geographic information systems such as Arcinfo®, ArcView® (Esri,
1997), Mapinfo® (Mapinfo, 1997), Geoconcept® (Alsoft, 1997), Idrisi® (Idrisi, 1997) also employ more
or less versatile sophisticated modules. While they all enable use of polynomial models, some also
propose models based on local triangulation. However, programs comprising rectification software
based on multiquadric functions are rare, except for some specific algorithms (MQReg, 1997). Most
versatile systems are equipped with an orthophotographic module.

M Data used for rectification

□ Level of correction
Images supplied by organisations that distribute remote-sensing data are already often subjected to
preprocessing. SPOT images, for example, may have been preprocessed to correct for acquisition
errors or to directly provide a georeferenced image, conforming to a conventional cartographic
projection system; such an image is known as a spatial map. We can also obtain geographically
rectified and grouped images that form a mosaic covering a large area, an entire district for example.
In such cases, radiometric corrections are often necessary to reduce response variations between
images comprising the mosaic. According to the level of processing, various products can be identified
(see Chapter 2).
Aerial photographs, because of their wide field of view, are much more deformed than satellite
images. Moreover, given the lower altitude of their acquisition deformations related to topographic
 Processing and Interpretation 247

variations cannot be ignored. Combined use of cartographic reference and altimétrie reference to
produce a geographically conformable image is known as orthophotography.
Data available for processing may be of different qualities depending on the remote-sensing
product. The cost of these various products, of course, is not equivalent. According to the number of
images and frequency of processing, one can acquire rectified images that can be used directly or set
up one’s own sequence of rectification. In the latter case, reference data for image rectification are
mandatory.

□ Reference data used for corrections

The same principles of rectification are used irrespective of the nature of reference data, its projection,
co-ordinate system or physical support. Most often, the reference system Is another remote-sensing
image or a map.

• Another image
In order to compare two images or to combine them in a single processing, the two images can be
rectified relative to a reference map or simply one Image rectified with reference to the other. In the
latter case, no reference map is needed. The images in such cases cannot be superposed on a
cartographic base and are not corrected for geometric distortions. The rectified image will inherit the
distortions of the reference image. These comparisons or combinations of images can be used, for
example, in diachronic analyses or for combining data from various sources.

• Reference map
Whenever possible, it Is more useful to rectify images with respect to a reference map. This effectively
enables superimposition of the result over other maps, more accurate spatial control on the ground
and its direct integration in a geographic database in the form of GIS, where it can be compared with
other data. Co-ordinates of the control points are determined in a cartographic projection system,
usually the Lambert system in France. Depending on the method of determination of these co-ordinates
from a map, two cases are identified, viz., using a paper map or a digital one.

Cartographic co-ordinates from a paper map

When digital maps are not available, the co-ordinates can be determined directly from a paper map.
For this, the map should have reference points in Its projection system. Maps often contain a square
grid representing the co-ordinate system. The grid may be drawn completely or more often, to avoid
overloading the map, only crosses at grid Intersections or marks for boundaries are indicated. While
co-ordinates defined in a projection system are almost invariably represented in large-scale maps,
such Is generally not the case in small-scale maps in which geographic co-ordinates (longitude and
latitude) are marked instead. The distance of a control point from the nearest grid points is measured
with a scale as accurately as possible. It should be noted that the grid lines are not parallel to the
borders of the map and hence the measurements should be imperatively made parallel to the grid
lines. Knowing the scale of the map, a simple three-point rule Is used to determine the co-ordinates in
the projection system of the control point. The precision of the co-ordinates obtained is limited In this
type of measurement. At best, a precision of 0.5 mm in reading the scale can be expected, which is
equivalent to 25 m on a 1:50,000 IGN map for example. Several errors accumulate in this measurement:
errors in reading the scale, in determining the position of a point (such as centre of a road) and in
referring to the grid.
These measurement errors can be reduced using a digitising table. At first the table is calibrated,
in general, on the grid and then the co-ordinates of the control points are directly determined using the
mouse. Advantages of the digitising table, compared to the preceding method, are: minimisation of
248 Processing of Remote Sensing Data

reading errors given the very precise pointing of the mouse, elimination of errors in referring to the grid
and availability of an adjustment model for correcting deformations due to the paper.

Map co-ordinates derived from digital data

Most publishers nowadays supply digital versions of their maps in raster (image) or vector form. Raster
data are scanned maps. Because of their nature, they can be used directly In all Image correction
systems. The National Geographic Institute (Institut géographique national, IGN) supplies a range of
scanned maps. Scan 50 (IGN, 1997) or 1:50,000 maps, for example, are scanned at 254 dpi (dots per
inch) giving a resolution of 5 m reportedly for a geographic precision of 10 m. When only a paper map
is available, it can be scanned but several problems need to be taken into consideration. For large
format maps, an AO scanner may be necessary. Deformations of the paper map, in particular its folds,
may lead to a deviation of several mm in width of a map with a theoretical grid. These deformations
can be corrected with a rectification model using the theoretical grid as reference. A finite element
model is well suited In such cases. It should also be noted that digitisation of a map, IGN for example,
is not freely permitted and is subject to copyright. Vector data are most often used in cartographic
systems and in GIS. Some rectification programs are also capable of processing these data. The
‘BDcarto’ and ‘BDtopo’ of IGN (IGN, 1997) are the most widely used vector products in France for
these corrections. Digital Chart of the World (DCW) is the world vectorial base map and can be used
for small scales.
Utilisation of digital reference data offers some advantages. A high precision of map reading is
obtained thanks to software zoom. Images can be effectively visualised nearly to a pixel level and
pointing error almost corresponds to data resolution. As in the digitising table, reference errors are
eliminated since the software ‘measures’ the position of the pixels. The possible distortions of reference
data can be corrected when a paper map is digitised. The advantage of a cartographic base lies in
verification of rectification by a posteriori superposing of the base on the rectified image.

®Altimétrie reference (DEM) for correction of parallax errors due to topography

Image distortions due to topography are proportional to the relief and angle of view (Fig. 13.13).

Fig. 13.13: Parallax error due to relief and angle of vision.

 Processing and Interpretation 249

For a point viewed vertically, the position in the image coincides perfectly with the vertical projection
and no deformation due to relief occurs. For a point viewed at an inclined field of view, the parallax
effect (see Chapter 14) gives rise to an offset between the position in the image and its vertical
projection. Let us consider a datum plane (at minimal altitude, for example) on which the relief of the
zone Is projected. The size of the pixel and its position on the projection vary with altitude relative to
the datum plane and with slope. For an altitude difference from the datum plane AA and an angle of
Incidence of the satellite a , the offset between the observed position of the point and its vertical
projection is given by D= AA tan a. For an angle of 20° and altitude difference of 150 m, D is equal to
55 m or nearly 3 pixels in the SPOT image. For zones of high relief variations, these parallax errors
should be corrected if a georeferenced image Is required. For this, altitude, angle of incidence and
local slope should be known. A digital elevation model is used to make these corrections.

13.3.3 Formulation of rectification model

■ Choice of control points
The control points should belong to objects visible both in the image to be rectified and the reference
map. Some parts of a map represent elements that are not present in the physical medium
(administrative boundaries for example) and not observed in remote-sensing data. A large difference
in resolution may also give rise to the fact that features present in the reference do not appear in the
image (or vice versa) or appear with a variable form, making a precise location difficult. A difference in
date of acquisition of the reference and the image may also have the same effects, since image
features may appear, disappear or be modified by human activity (roads, deforestation) or natural
phenomena (such as floods). Cloud cover may mask some features. Generally, objects that can be
identified with no ambiguity and located precisely are chosen as tie points. Objects whose form and
spatial organisation aid in locating a reference point at the level of an image pixel are used. Temporal
stability of morphological characteristics of the objects is essential for comparing data not necessarily
acquired on the same date. Constructions and management elements are often preferred over natural
objects, since they are more regular in geometry and stable. Buildings and especially road networks
satisfy these criteria. Borders of agricultural parcels or forests are avoided since they are not stable.
Objects used for calibration are at the limit of image resolution and may exhibit contrast variations,
making them difficult to precisely locate pixel by pixel. Hence, alignments and intersections are used
as reference for locating a precise point In the image. A set of aligned points, rather than a single point,
is relied upon. Intersections of road networks are therefore most commonly used reference points. If
the reference document is another image, a map is useful to mark the reference points in the Images.
In some zones, only a few or no control points may be precisely located in the image.This is particularly
the case in areas of least human activity and in methods using reference forms (group of pixels) rather
than existing reference points (Hubault, 1994).

H Number of control points

In the case of first-order rectification, the polynomial equation has three constants to be determined (3
in Xand 3 in V). Three control points are therefore needed to formulate 3 equations to determine these
3 unknowns. For a second-order rectification, six points are also necessary. This is the minimum
number of points mathematically required for solving the polynomial approximation. More number of
points than the minimum are recommended so as to reduce possible errors related to quality of reference
data and location of points. A dozen points are generally used for a second-order correction of a
slightly deformed satellite image (Fig. 13.14). On the other hand, taking too many control points for a
polynomial model Is not useful. In fact, since it is a global model, while the number of points is sufficient
250 Processing of Remote Sensing Data

Reference : georeferenced SPOT image (CD-ROM) LANDSATTM image to be rectified (CD-ROM)

Fig. 13.14: Second-degree rectification of aTM image over a georeferenced SPOT image: location of
ten control points.

for regular distribution, rectification cannot be improved locally by adding points In the poorly rectified
zones. In such a case, the entire model will be altered and other parts of the image not calibrated
correctly. In the case of a locally too irregular image, a local deformation model can be used with more
number of points, covering various distortion zones. If topographic variations of a zone are not negligible
in the scale of the Image, use of an orthophotographic method with DEM provides high-quality
rectification with no need for too many points.

M Distribution of control points in the image

In higher-order polynomial approximations the extrapolation problem may give rise to undesirable
image distortions depending on the spatial distribution of control points. The image may then be
globally stretched or compressed in several directions. A regular distribution of points on the boundaries
of the zone to be corrected will minimise such distortions (Fig. 13.15a). Concentration of points in one
or several zones of the Image should be avoided (Fig. 13.15b).
High-accuracy restitution cannot be obtained for the entire region when the zone under analysis
significantly differs in altitude and orthophotographic and DEM softwares are not available. Mean
altitude of zones of importance is determined and control points in the image are distributed according
to this altitude. Accuracy of restitution is optimal at this altitude and gradually decreases with distance
from the mean altitude.

H Identification of control points

After defining the type, number and general position of control points in the zone under study, their co­
ordinates in the image and the reference have to be determined. Rectification programs most often
comprise an Interface that aids in interactively identifying the points. If such facility is not available, it is
necessary to determine these co-ordinates by means of a simple program of visualisation of Images
and transfer them to the rectification system, generally through an ASCII file. In integral systems, the
 Processing and Interpretation 251

(a) TM image 197-26 (b) TM image 197-26

Fig. 13.15: Distribution of control points in the zone to be corrected: The sameTM 197-26 image after rectification
over the same reference image (georeferenced SPOT) with the same number of control points, distributed around
the image in case (a) and at the centre of the image in case (b).

image and the reference are displayed simultaneously. Any part of the zone can be zoomed or moved.
Some software programs enable simultaneous view of several reference Images or several views with
different level of zoom. All the control points are displayed In these views as and when they are
identified. In some cases, residual error in X and / at each tie point is displayed as a small vector.
These data facilitate correction of the model during identification. It is necessary to determine the
correct level of zoom for identification of points to an adequate precision but not necessarily to the
pixel level, because of local contrast variations (see the section on ‘Choice of control points’). Several
types of error may arise during identification of control points: errors of pointing, location and
interpretation.
Pointing errors are less common if a sufficient zoom level is used. Errors of pointing with the
mouse depend on the level of zoom.

plot plot

\ plot plot plot

A , A
plot ■

plot
plot \
■

(a) SPOT reference image (b) TM image to be corrected

Road used for locating point Number 10
shown in the two images

Fig. 13.16: Error of location of tie point No. 10.

252 Processing of Remote Sensing Data

Errors of location of a tie point arise due to incorrect interpretation of the pixel configuration of the
object used as reference. In the example of Fig. 13.16, an intersection of roads following an agricultural
parcel was chosen as the control point. The reference is a georeferenced SPOT image with a resolution
of 20 m and the image to be calibrated is a LANDSAT TM image with a resolution of 30 m.
The first rectification model (Table 13.1a) has shown a larger error in X for point No. 10. It can be
noted, using a map, that point No. 10 has been incorrectly located in theTM image. In fact, the resolution
of the TM image does not allow a clear view of this small road. It Is located based on the boundaries
of the parcel. However, a dark band can be noticed in this image, to the left of the road in the area of
plot A. This band has caused confusion with plot boundary where the road really exists. A correction of
point No. 10 has minimised the residual error of this point (Table 13.1b).

Table 13.1: Variation of rectification model after correction of an erroneous point

(a) Model with erroneous point 10 (b) Model with corrected point 10

Points Residual in X Residual in Y Points Residual in X Residual in Y

1 0.25 -0.046 1 0.39 -0.059

2 -0.65 -0.391 2 -0.29 -0.424
3 -0.16 0.323 3 -0.29 0.336
4 0.66 0.132 4 0.39 0.157
5 0.67 0.083 5 0.48 -0.065
6 -0.82 0.207 6 -0.39 -0.248
7 0.20 0.289 7 -0.09 0.320
8 0.58 0.230 8 0.20 0.263
9 0.75 0.593 9 -0.14 0.677
10 -1.50 -0.840 10 -0.24 -0.957

It may be noted that the correction of error in X of point 10 reduced the same error for most other
points of the model. In fact, in the case of polynomial methods, only one global model exists, computed
for all the points, and hence an error for one point leads to an error in the entire model. Errors of
location may also be generated by an Incorrect interpretation associated with the scale of objects in
the data. Difference In resolution of the images and effects of cartographic generalisation of maps
(thickness of representation of some objects not to scale) make it necessary to use, in the case of
linear objects, the axis of the objects, but not their boundaries for fixing control points. Only the image
points present on the axis of the object can be considered geographically similar to the reference
points.
Identification errors arise when the object identified in the Image does not correspond to that
chosen In the reference data. For example, in a zone of high urbanisation road intersections are
‘mistaken’.

H Final resolution of rectification

Resolution of the rectified Image produced generally corresponds to that of the reference image.
Some systems enable determination of this resolution.This facility is particularly useful when multisource
data, which most often are of various resolutions, are integrated. While most processing programs
require that images to be combined should be of the same resolution, some are capable of dynamically
combining images of different resolutions.
 Processing and Interpretation 253

B Verification and improvement of rectification accuracy

One way of verifying the accuracy of rectification by an interpolation method is an a priori verification
of the accuracy of the approximation model. This is estimated from the residual errors at each point. A
posteriori verification can be made by superimposing the rectified image over its reference image. If
residual errors are too large, the model should be corrected. Various approaches can be followed for
this: verifying the points commencing with those having the highest residual errors; adding points if
they are too small In number or if their distribution is not regular in the zone (see section on distribution
of control points); removing or replacing the most ambiguous points; checking the distribution of points
In the zone according to altitude and verifying that altitude differences are acceptable for a polynomial
model. If this is not so, DEM should be used along with an orthophotographic program.

13.3.4 Conclusion
Image rectification is an important stage for using remote-sensing data. Choice of reference data and
methods used depend on the scale of study. The nature of remote-sensing data directly determines
the mode of rectification to be employed. Resolution of the image, acquisition parameters (altitude,
angle) and relief of the zone of study are also equally important In choosing a model. Numerous
studies have been carried out on development of more efficient models. Local or intermediate models,
such as the multiquadric, have enabled improvement of spatial accuracy of rectification. Rectification
can be applied for images of different resolutions derived from the same sensor or from various
sensors. Multiscale models based on multiresolution analysis of images using wavelet transformation
have been developed (Djamdji, 1993). The mosaic consists of rectifying several images together so as
to cover an extensive field. It facilitates studies at district or regional levels. Integration of rectified
images in a GIS often constitutes the final stage of rectification. It opens many possibilities of analysis
by facilitating combination of remote-sensing data with other geographic databases.

References
Alsoft. 1997. http://www.alsoft.fr.
Djamdji JP, Bijaoui A, Maniere R. 1993. Geometrical registration of images— the multiresolution approach.
Photogrammetric Engineering and Remote Sensing, 59:645-653.
Erdas. 1997. http://www.erdas.com.
ER/Mapper. 1997. http://www.ermapper.com.
Esri. 1997. http://www.esri.com.
Fogel DN. 1996. Image registration using multiquadric functions, the finite element method, bivariate mapping
polynomials and the thin plate spline. National Center for Geographic Information and Analysis, University of
California USA. internal report n° 96-01, pp. 1-44.
Hardy RL. 1971. Multiquadric equations of topography and other irregular surfaces. J. Geophysical Research, 76
(8): 1905-1915.
Hardy RL. 1977. Least square prediction. Photogrammetric Engineering and Remote Sensing, 43 (4): 475-492.
Hardy RL. 1990. Theory and applications of the multiquadric-biharmonic method. Computers and Mathematical
Applications, 19:163-208.
Hubauit J. 1994. Corrections géométriques utilisant des formes d’appui d’image à image ou d’image à cartes; un
critère d’ automatisation. Séminaire SFPT-RSS, Qualité de l’interprétation des images de télédétection pour
la cartographie. Grignon, pp. 9-11.
Idhsi. 1997. http://www.idrisi. corn.
IGN. 1997. La gamme de produits geomarketing IGN.
Jensen JR. 1986. Introductory Digital Image Processing, a Remote-sensing Perspective. Prentice-Hall, 379 pp.
Khoros. 1997. http://www.khoros.unm.edu.
254 Processing of Remote Sensing Data

Mapinfo. 1997. http :www.//mapi nfo.com.

MQReg. 1997. Multiquadric rectification, http://pollux.geog.ucsb.edu/~fogel/mqreg.html.
Noésis. 1997. http://www.noesisvision.com.
Pratt WK. 1991. Digital Image Processing, Wiley-lnterscience, 698 pp.
Rousselet M. 1985. Graphisme 3D, Éditions techniques et scientifiques françaises, 223 pp.
Teissier L, Stamon G. 1994. Geometric ortho-rectification of flash radar images. First IEEE International Conference
on Image Processing, Austin, Texas, USA
Teravue. 1997. http://lacan.grignon.inra.fr
Worboys MF. 1995. GIS Computing Perspective. Taylor & Francis, 376 pp.
 14
Fundamentals of Aerial
Photography
It is necessary to review the knowledge base of aerial photography for two reasons. Firstly, visual
interpretation has regained its place in the gamut of methods of interpretation of remote sensing data
such as satellite images, satellite photographs, aerial photographs, etc. Secondly, satellite remote
sensing including development of satellite missions, image interpretation, relief measurements, etc.,
has emerged on the basis of data, concepts and methods concerning aerial photography: acquisition
of aerial photos, their quality, organisation of flight paths, photogrammetric measurements and
interpretation of aerial photos. Initially, in fact, satellite remote sensing combined photo-interpretation
and photogrammetry by associating physicists and aeronautical engineers.
Fundamentals of photo-acquisition, stereoscopy and photogrammetry are presented below.

14.1 PHOTO-ACQUISITION
Acquisition of aerial photos during a denominated mission flight is dependent on three factors:
1) Factors related to the data to be generated: focal length of the objective of the receiver, format,
coverage and scale.
2) Factors concerned with the region under study: area to be studied, maximum altitude,
3) Factors pertaining to platform: speed of aircraft, stability in flight, etc.
A good aerial photograph should satisfy three demands. Firstly, it should furnish maximum
information, which necessitates correct choice of optics of the camera used, combination of film and
filters and conditions of acquisition. Secondly, the photographs should cover the entire region under
study. Thirdly, it should facilitate accurate measurements of length and area, which necessitate metric
cameras and precise geometric analysis of the entire sensor-platform system.

14.1.1 Photographic sensors

Aerial photographs can be taken with an amateur photo-camera in an aircraft but the photos will not
be of excellent quality for several reasons.
For a colour emulsion, the photos appear with dominant blue colours and a yellow filter has to be
used to avoid this mist which is due to atmospheric disturbances (Chapter 1). Further, there can be
reflection of the porthole glass and a better angle has to be chosen to avoid this reflection or the
camera has to be directly inserted in the wall of the aircraft cabin. In this case, the photograph will not
be vertical. Hence, the aircraft has to be turned so that the camera is vertical when clicked manually,
which, of course, would be an acrobatic feat. Alternatively, the camera should be kept in a vertical
256 Processing of Remote Sensing Data

position under the aircraft wing (but provided with a long clicking mechanism) or directly in the floor of
the cabin.
Basics of emulsions and films are mentioned in Chapter 3.

B Metric cameras
Special cameras, known as metric cameras, were developed for acquiring aerial photos. They represent
cameras for which any geometric and chromatic distortion can be detected and measured. These are
employed for obtaining systematic photographic views, which can be subsequently processed for
relief restitution.
A navigational telescope was attached to these metric cameras so that the navigator of the aircraft
could verify the flight path. Today GPS (Global Positioning Systems) are used for this purpose. All data
about the flight, viz., its speed, position, etc., are recorded In an electronic control unit.
Aerial photographic cameras are characterised by film formats ranging from 35 mm (24 x 36 mm
photo) to 70 mm (55 x 55 mm photo) and for metric cameras (Fig. 14.1) 240 mm (230 x 230 mm
photo).Thickness of the film on the base should be uniform to avoid distortions in measurements to be
made subsequently on the photographs. This Is ensured by a depression in the focal plane punched
with numerous small holes. These cameras are equipped with an automatic trigger which should be
regulated according to flight plan: altitude, coverage, flight height and speed. Film magazines should
contain very long film, 120 m for example (or 500 snaps), to avoid frequent changing. An automatic
system of film movement should be regulated as a function of speed of aircraft and the ‘base’,
i.e., distance covered between two successive photographs. Exposure times are very short, 1/10 to
1/1000 s. Roll due to aircraft motion during photo acquisition should thus be avoided to preclude photo
blurring. For an exposure of 1/100 s, displacement is of the order of 20 cm for an aerial photograph
and 80 m for a satellite Image. Compensating mechanisms are presently used. Correction for roll is
obtained by a displacement of focal plane in the direction of flight during the exposure time with a
velocity v\

\/= V x fIH

where V\s the flight speed (for example, V= 720 km h""* or 200 m s“ "*), /the focal length of the metric
camera (f= 152.4 mm) and H the flight height above the ground (H = 6096 m).
For the example given, v= 5 mm s“ ^ and the focal plane should be displaced by 5/100 mm during
the exposure time of 1/100 s.

Cabinet

Objective

Fig. 14.1: Schematic diagram of a metric camera.

 Processing and Interpretation 257

Aerial cameras are equipped with an objective (Fig. 14.1) comprising a shutter, a diaphragm,
filters and an objective which most often consists of an assembly of several lenses. The objective is
defined by its focal length which varies from 90 to 300 mm. Focal lengths of 125 and 152.4 mm are
common In cameras used by the National Geographic Institute (IGN). A camera is also defined by Its
field of view a which varies from less than 75° to more than 120°.
The optical axis of the objective is perpendicular to the plane of the film at the centre. Consequently,
aerial photos are conical projections whose field of view is dependent on the angle of aperture a and
flight altitude H and whose scale E is dependent on the focal length /and flight altitude. Focusing is
evidently at infinity and hence the distance between the film and the optical centre of the objective is
equal to the focal distance.

■ IVpes of cameras
Various types of cameras are used for several specific cases.
Large format cameras were developed for satellites such as the Gemini and Apollo missions.
They had a focal length of 305 mm and took 230 x 460 mm photos. The European Space Agency had
a program with a metric camera mounted on SPACELAB during a flight of space shuttle. The camera
constructed by Germany had a focal length of 305 mm and 230 x 230 mm size photos (Schuhr et al.,
1984). The aim of this program was to study the possibilities of Interpretation of spatial data for
preparation of maps. Preparation of current maps based on satellite data shows that the result is
positive (SFPT special number 99, 1985).
Multispectral cameras have also been developed. They consist of four objectives equipped with
various filters and take photos on black and white films. The four photos taken at the same time are
viewed simultaneously with four filters to obtain colour composites. Thus composites with ‘real colour’
and ‘false colour’ or infrared colour are prepared. Later, several missions were launched in which four
Hasselblad cameras were combined, each with a different filter, since these cameras were less
expensive.
Panoramic cameras are also available which, equipped with a rotating objective, photograph a
long strip perpendicular to the flight path. The disadvantage of these photos is that the scale changes
continuously. However, the Apollo missions photographed the surface of the Moon using these cameras.
A camera of this type used by NASA has a focal length of 610 mm, aperture angle of 120° and film
length of 2000 m.
Several other cameras are marketed but not described here. Only digital cameras, which are
undergoing continuous development for usage by amateurs as well as professionals, need to be
mentioned. Digitisation of data is evidently an important advantage since nowadays data is processed
using computers. It may be noted that aerial photos can also be digitised by scanning. Modern scanners
provide high-resolution data, comparable to the geometric resolution of aerial photos in black and
white or colour.
Criteria for comparison between aerial photos and digital images (Table 14.1) have been developed
in recent years since digital Images ought to provide significant improvements.
Table 14.1: Comparison between photographs and digital images

Operation Photographs Digital Images

Data acquisition Film CCD digital sensors
Storage Film or paper Magnetic tapes, CDs, hard disks
Interpretation Visual Computer processing
Transmission Post, Fax Radio, networks
Visualisation Projection, paper maps Video projection, television,
computer monitors
Reproduction Photo printing Electronic printing, ink jet, laser
258 Processing of Remote Sensing Data

14.1.2 Types of aerial photos

Several types of aerial photos can be distinguished depending on the position of optical axis with
respect to the vertical and motion of the aircraft. The main aim in aerial photography has been to take
photos as fast as possible in the best meteorological and geometric conditions and with as wide a field
of view as possible. This objective has remained almost the same for satellite images, aside from
increasing revisiting capability.

H Vertical photos
The optical axis of the camera is vertical (Fig. 14.2). When the ground is horizontal, the relationships
existing between the ground (point Y) and the photo (point y) are hom othetic. In vertical
photographs, the central point of the photo (p) coincides with the nadir (F), point at the vertical of the
optical centre of the camera. The optical axis passes through the nadir for these photos. This position
is used to acquire satellite images. The homothetic ratio H/f, where H is camera height above
ground and f the focal length of the camera, is equal to the inverse of scale. The photo is a conical
projection.
Knowing the aperture angle of the camera and the flight height, the field of view of the photograph
and area from which the information is gathered by the sensor are derived.

9’ y d’

Fig. 14.2: Schematic diagram of taking a vertical photograph.

H Oblique photos
When the optical axis of the camera forms an angle of more than 5° with the vertical, oblique photographs
are obtained. The nadir and central point of the photograph are not coincident.
A square outline on the ground appears as a trapezium In the photo. These photographs can be
distinguished as panoramic oblique and low oblique (Fig. 14.3). In panoramic photos the nadir is no
longer photographed and the horizon Is visible In the photo.
 Processing and Interpretation 259

Lens

(a)

Lens

(b)

Fig. 14.3: Panoramic (a) and low oblique (b) photographs.

Two oblique (at 60°) panoramic photographs together with a vertical photograph constitute a
trimetrogon system (Fig. 14.4). They enable rapid reconnaissance mapping. They were employed
prior to satellite images to obtain aerial photographs of Nordic regions where one can fly only for short
periods. For example, such a mission was launched to photograph Spitzberg. Thus strips separated
by 25 to 30 km could be photographed at a flight height of 8000 m.
The horizon does not appear in low oblique photographs. Two photographs can be taken to the
left and right of the camera or one in front and the other in the rear (Fig. 14.5). In the latter case a
convergent view Is obtained. Depending on the angle of Inclination of the camera, 100% coverage can
be achieved. With the advent of super-wide angle cameras interest in low oblique cameras has
decreased. However, this concept has been applied in one of the SPOT-5 projects with the objective of
obtaining a stereosicopic pair from a single pass of the satellite in nearly 10 s.

■ Emulsions and formats

Emulsions were described in Chapter 3. Conventionally emulsions are classified as panchromatic and
infrared (black and white), colour and colour infrared (CIR). Aerial photographs are acquired in all
these four emulsions in IGN.
The most common format used at present In France is 23 cm x 23 cm, with 1:30,000 scale. Earlier
formats of 18 cm x 18 cm with 1:25,000 scale and still older ones of 13 cm x 18 cm are also available.
260 Processing of Remote Sensing Data

Fig. 14.4:Trimetrogon photography.

 Processing and Interpretation 261

It may be noted that aerial photographs sold by IGN can be enlarged by about 25 to 40 times.
Enlargements are carried out by IGN up to scales of about 1:1000.

14.1.3 Regular missions

The Hurel-Dubois aeroplane, with a flight height of 6000 m, was specially designed for aerial photography
and proved very stable. In France, regular missions of aerial photography were undertaken for a long
time on propeller bombers B 17 (‘flying fortress’) with a flight height of 8000 m. The Mystere 20 aircraft
with an operational altitude of 12,500 m and range of 4000 km was also used. Today various aeroplanes
are used for aerial photography. A pilot, navigator and photographer man an aircraft for a photographic
mission. They are equipped with a navigational telescope for viewing the ground below the aircraft,
two cameras, an intervalometer for fixing correct spacing between photographs, a GPS, as well as an
apparatus for determining the aircraft attitude and height. Three hatch doors are therefore required in
the cabin floor, which sometimes pose problems in designing the aircraft profile. This results in higher
costs, especially when the mission employs other sensors such as metric cameras, some of which do
not accept glass or quartz structures as in the case of middle or thermal infrared scanners for example.
Aeroplanes used for remote sensing cannot be employed for other missions.
When roll, pitch or tilt occur in flight, the track of the aeroplane and hence of the camera Is
modified, which changes the area covered on the ground by photography (Fig. 14.6). Thus a spherical
surface gives the deviation of the axis of the metric camera with respect to the vertical for each
photograph. Sometimes, stabilisation is achieved using gyroscopes and servomotors, which ensure
verticality of the metric camera in spite of aircraft oscillations.
Aerial photographic surveys have several disadvantages:
— choice of date of photography is not certain since flight plans depend on meteorological
conditions;
— area covered is relatively small for a given mission;
— cost is relatively high.
The advantage of aerial photography still evident to date is Its very high geometric resolution,
which probably won’t be enjoyed much longer.

PITCH

ROLL

Fig. 14.6: Roll, pitch and tilt of aeroplane: modifications in ground coverage by aerial photography.
262 Processing of Remote Sensing Data

One foresees that with miniaturisation of electronic components of sensors, the possibility of
operating sensors outside the cabin and use of GPS that give very precise position of the aircraft,
application of aerial remote sensing will witness new developments.

14.1.4 Organisation and overlap of aerial photos

The National Geographic Institute (IGN) conducts regular aerial photographic surveys in areas
corresponding to 1:50,000 topographic maps of France or districts (to the requirements of National
Forest inventory, IFN, In particular). The survey consists of a series of photographs taken successively,
if not the same day, at the shortest possible time. The successive flight lines are parallel to one another
(Fig. 14.7), thus providing a strip of photographs.

R = 55 to 60 %

I N

I N

_____ Boundary of zone to be photographed

___ Aircraft flight oath
__ Aircraft flight path when
No photograph is taken

Fig. 14.7: Positions of aerial photos in a survey: overlap.

The photographs should have an overlap for obtaining a stereoscopic view. This successive overlap
(endlap) (R) is 55 to 60% and may go up to 80 to 90%. Photographs of each strip overlap a few grid
cells of the adjacent strip. The lateral overlap (sidelap) (r) varies from 10 to 20%. In most cases the
strips are taken from west to east. In this direction north is to the left of the aircraft whereas, when it
takes a half-turn to photograph the adjacent strip farther south, south is to its left. As some acquisition
parameters are indicated on the borders of a photograph, they represent north and south alternatively.
Thus, when aerial photographs of IGN are viewed, numbers of photographs are alternatively observed
in the direction north then south.
In the case of high mountains, as in the Alps, surveys are carried out from north to south. For
surveys launched for a specific objective, flight directions can be arbitrary.
An overlap of 50% ensures stereoscopic view; below this value, no ground object is seen at two
different angles and hence a perspective view cannot be obtained. Contrarily, if the overlap is 60% or
more, ground marker points, which are situated in three photographs, can be Identified. Consequently,
the resolution is improved. If a constant flight height Is taken relative to the sea level and if the
photographed zone shows high relief, it is possible to observe portions of the ground in three or two
photographs. If they exist and are observed in only one photograph, a perspective view cannot be
obtained and some parts may not be photographed at all (Fig. 14.8). For this reason, another set of
photographs of the strip is taken at a different altitude.
The minimum clearance between the flight line and the maximum altitude required for a minimum
overlap of 50% can be deduced. Zq corresponds to the altitude below which the overlap at any point
would be at least 60% (Fig. 14.9).
 Processing and Interpretation 263

Fig. 14.8: Various overlaps of a zone of aerial photography,

a — aperture angle of camera; numbers 3, 2,1 and 0— number of photos covering a zone
designated by a horizontal line.

The distance separating two successive photographs (Fig. 14.10) on the same flight line, known
as air base B, depends on the aperture angle a of the camera, the overlap R and the flight height
above ground H.
The format F o f the photograph is computed as:

F = 2 H ta n a /2

Since B = F(1 - R), we get:

e = 2 H (1 » F )ta n a /2
264 Processing of Remote Sensing Data

Fig. 14.10: Computation of air base B.

The base-height ratio B/H is defined as the ‘characteristic’ that depends only on the aperture
angle, which is constant for a given sensor. This ratio is used to determine the overlap. For aerial
coverage of France, the base-height ratio is close to 0.58. In the case of SPOT satellite, when images
are taken at a day’s interval, this ratio is 0.5 and for images taken with an inclination of 27°, it is 1.19
(see Chapter 2 and CD: SPOT system).
The distance between two flight lines is computed in the same way as for B, but by taking the
sidelap r, the values of H and a remaining the same.

14.1.5 Flight plan (example)

A flight plan Is determined from various parameters mentioned earlier regarding the types of photographs
and onboard cameras, and specific characteristics of the platform, here aircraft. Among others, its
height and speed should also be taken Into consideration. All these parameters can be computed. As
an example, information obtained from IGN for a photographic mission available for public sale is
given below:
1988 Bourg-Saint-Andéol-Nyons.
Date: 2 June 1988 from 7 h 21 to 7 h 59.
Photographs 1 to 99.
Focal camera: 15 UAg. 1041. Focal: 151.83.
Altitude above ground: 4590 m, above sea level: 5090 m.
Scale 1:30,000
Aircraft: Mystere 20.
For this survey, with one photograph out of two, a 60% overlap is possible.

14.2 STEREOSCOPY
14.2.1 General
Sharpness of visual discrimination e is defined as the minimum angular distance between two points
to be seen as separate. It depends on the illumination, shape, colour and contrast of the object and
 Processing and Interpretation 265

acuity of the observer’s vision. On average, e = T , which corresponds to a distance of 5 pm on the

retina.
In monocular vision, the eye estimates distances from modifications in convergence of the lens.
Distances are also interpreted with reference to normal objects to which the memory assigns a value.
Estimation of perspective is also based on the position of shadows. An illusion of 3D relief is
obtained when the shadow is situated at the bottom and on the right, which indicates that the light
source is on top and to the left. If the observed photograph Is rotated by 180°, the object and shadow
are interchanged and we get the Impression that the perspective Is reversed. This is observed in
images when no other reference objects exist, which our brain identifies as reverse to the impression
of relief. Consequently, cloud masses or images of planets give this type of impression. Conventionally
a geographic map Is drawn with the north showing upwards and the south downwards. In such a
reference frame, the sun is never at the top left In the northern hemisphere. Hence, shadows never
occur at the bottom right, i.e., south-east. This is one of the reasons for the difficulty in observation of
aerial photographs and more so In the case of satellite images acquired over relief, when they are
compared with shadowed topographic maps. In fact, shadows in topographic maps are situated in the
south-east, whereas no shadow exists in this exposure.
In binocular vision, distance is estimated by convergence of optical axes, i.e., parallax, of the
eyes, and the relief is estimated by parallax. This perception of relief is a complex phenomenon but
can be represented by a simplified model.
If an observer looks at a point A (Fig. 14.11), his left eye G and his right eye D accommodate a
distance cfg. The lines of vision of the two eyes converge at A at an angle a. If the observer now looks
at point C, the distance of accommodation would be d^and the angle of convergence y .The segment
AC will be viewed at an angle p . A relationship exists between and a on the one hand and between
dç and y on the other.

Fig. 14.11 : Perception of relief.

When aerial photos separated by a distance B (base) are taken, the image of point A at height
is viewed at and the image of point A/at height Hg Is viewed at in photograph P.|. Segment NA Is
seen under angle p and its image in photograph is segment a^. Similarly, in photograph P2,
segment NA is seen under angle p and its image Is segment a2 Distances and 32^2 are
parallaxes which can be measured and from which distance NA can be derived. This provides an
266 Processing of Remote Sensing Data

estimate of topography. Thus two photographs (or images) in which the same point is seen under two
different angles of view, suffice for obtaining the perspective view.

14.2.2 Stereoscopic vision

When two photographs overlap the same region, In which objects A, Band Care situated at the same
altitude and object D at a different altitude, the four objects will be observed in a different sequence in
the two photographs: a, b, d, c In the left photograph and a, d, b, cin the right (Fig. 14.12). In the same
photograph, segments ab and be are equal since they are at the same altitude, but segments ad and
dc are not.
When a stereo pair Is viewed, one notices that the vertical scale is greater than the horizontal
scale, resulting In exaggeration of relief. This results in a false estimation of slopes, heights, etc. On
the other hand, a stereoscopic pair enables detection of small differences In slope or breaks in slope.
Hence It is useful for visual Interpretation.
A nomogram (Fig. 14.13) connecting three parameters, viz., true slope (a ), apparent slope (p)
and vertical exaggeration of relief (e^), Is used to determine one parameter from the other two.
The apparent slope (p) is measured from the aerial photograph. We can fix a pin at the base of
the slope and, under the stereoscope, tilt it until an Impression is obtained that it is in contact with the
relief model, and finally measure this slope with a protractor. The true slope can be measured in a
topographic map or In the field with a clinometer. On the nomogram, the point of intersection of the
lines of apparent slope (moving along the perpendicular to the abscissa) and true slope (moving along
the inclined line) is located, and its position on the ordinate gives the vertical exaggeration of relief. For
common stereoscopic plates, the relief exaggeration is about 2.5. With such a value, an apparent
slope of 55° corresponds to a true slope of 30° and an apparent slope of 20° to a true slope of 8°.

Fig. 14.12: Perception of relief from two aerial photographs.

 Processing and Interpretation 267

■ a : True slope

n r-r-r-;rT ",
! |j I I /I I /
IiM I /1 II /'
I 11,1
,,, 11 I 1,'i 1/
n, , I - I ( f i- - ..... v'l
1/ I |/ /
II I Ij /
1/ ! /
4 /
III II -/•- - I 1r-
III I
III 51 'I 1/ I /
III
III illill
nTjit-

'/!1/!1 ■
III III ■'1....
III III I
111(11 I ^
'iilii 1
3UJJIIJ
H'i! itr
lll!ll !I 1/ 1/ I
mill ji 1/ !/' i
P III
mill/I II-I I' < I, f I1 / / 1 /
r---
^ 11111Ij II11j IU;i M /1 1I // '-^-7'^....
/
iiii|ii|/i/i !/
II< I'i 1/1
II I 1/ I /i l-J. / I / /
'7' F--_ri— --.7^ ----- ----
!!!!U ( \ /\/
I IIÍÍ /1I//II /I1 /!I
rrjr;íl^-íV--i/ r-|
iilt/iin 1/ / 1/1
ii |i / i r / i / /I 1 ■

f
llilill/l/i/!/f 1/
258101520 25 30 35 40 45

ß : Apparent slope

Fig. 14.13: Nomogram of true slope (a), apparent slope (P) and vertical exaggeration of relief (e^).

In practice, this value of relief exaggeration corresponds fairly well to the impression in the eye and
can be recommended for drawing sections or three-dimensional diagrams to represent geomorphology
and landscapes.
Relief exaggeration is due to six factors, three dependent on the sensor-platform system and the
other three on the stereoscope-observer system.
— Air base B: If this increases, relief exaggeration increases. Thus, in the case of satellite images,
largest base and hence highest vertical exaggeration is obtained by taking SPOT images with -2 7 °
and +27°. This is necessary for morphological interpretation of these Images.
— Focal distance f: If this Increases, vertical relief exaggeration increases. This is an important
factor at present, since photographs and images of Increasingly longer focal distances are being
used: the latter have changed from 70 mm to 125 mm, then to 152 mm and even 300 mm.
— Flight height H: If this increases, vertical exaggeration of relief decreases. This is the case of
satellite Images and photographs.
— Interocular distance y is connected with the observer. If y decreases (eyes closer), vertical
exaggeration of relief increases.
— Eye-to-photograph distance d: If this increases, as in some stereoscopes, vertical exaggeration
of relief increases.
— Overlap of photographs R: Overlap R is inversely proportional to the distance between the
principal points (centre of each photograph from the stereo pair): it is the separation s of the photographs
over the support on which they are observed (Fig. 14.14). If R decreases, or s increases, vertical
exaggeration of relief Increases. Some stereoscopes have magnification which enhances the virtual
overlap. The latter decreases the vertical exaggeration of relief.
The relationship between these factors can be summarised as:
268 Processing of Remote Sensing Data

BHd/yRH

which can be represented in various forms:

— Depending on the two systems, stereoscope and flight, e^= (Bf/H) [d!yR)\
— or e^= [f! H) (Bd/yR)] this is the product of scale and a hyperstereoscopy factor;
— or using the base-height ratio (5 / H), focal length of the camera f and d of the stereoscope,
= (B/H) {fd/yR).

14.2.3 Stereoscopes
A stereoscope aids the observer to fix one eye on each photograph of the pair and see them as only
one photograph. Further, the distance of between the eye and the photograph is maintained constant.
A stereoscope (Fig. 14.14) consists of a double optical system (lenses, mirrors, prisms, etc.)
mounted on a rigid frame supported on legs. In this way, distance dis fixed. The optical system is such
that the virtual Image is cast at infinity and consequently stereoscopic vision is obtained without eye-

Fig. 14.14: Types of stereoscopes.

 Processing and Interpretation 269

strain. Moreover, in most stereoscopes the interocular distance y is variable and can be adjusted to
suit the observer.
A simple lens stereoscope is made up of two achromatic convex lenses. The focal length is equal
to dcorresponding to the height of the stereoscope above the plane on which the stereo pair is placed.
The only inconvenience of this stereoscope is that it does not facilitate vision of the entire area of the
photographic pair of format 23 cm x 23 cm.
A mirror stereoscope comprises two metallised mirrors, two prisms, two lenses and two eyepieces.
It enables viewing the entire area of the photographic pair of format 23 cm x 23 cm.
In some stereoscopes the optical part is fixed on an arm and the photographic pairs are arranged
on two different planes. They facilitate analyses of several stereo pairs consecutively without changing
the arrangement.
Many models exist in which the scale of the photograph can be changed and hence stereoscopy
of photos of different scales can be obtained and/or Interpretation can be made on a topographic map,
thus Integrating a camera lucida.
These systems, more and more bulky, have to face stiff competition from digital systems, which
are highly developed for satellite images and furthermore because photographs can be digitised.

14.2.4 Anaglyphs and vertographs

Some other methods are also used for obtaining stereoscopic view.
Most common is an anaglyph, a figure comprising two traces of the same object, one in blue-
green (seen under a particular angle) and the other In red (seen under a different angle). Green and
red glasses are used to get a 3D view. The left eye sees through a red transparent film and hence
perceives, as black, only those lines that are drawn in green-blue and not those in red. The right eye
sees through a green-blue transparent film and perceives, as black, only the red lines. Thus the same
object Is seen as black under different angles. This suffices to produce a sensation of three dimensions.
If the glasses are reversed, a pseudo-stereoscopic view Is obtained which gives the impression of a
reversed relief: valleys correspond to high points and peaks to basins.
A vertograph (polarised platen viewer) is based on polarisation of light. The observer wears
polarising eyeglasses. One glass allows only vertically polarised light and the other only
horizontally polarised light. Thus, each eye perceives only one image. The two images are projected
onto a screen, one with vertical polarisation and the other with horizontal polarisation. Once again,
each eye perceives a different image, since viewed at different angles and the observer obtains
perspective view by synthesising them. The principle of the vertograph is very useful at present for
stereo restitution.

14.3 PHOTOGRAMMETRY
Photogrammetry Includes all aspects concerned with obtaining quantitative Information from
photographs, viz., measurement of distances, altitudes, etc. It provides geometric descriptions of
locations and extensions of phenomena interpreted from photographs. Photogrammetry is an accurate
and complex science described in many books; hence we present below only the basics needed for
understanding how results of Interpretation of aérospatial photographs and images can be linked to a
precise geometric space.
Photogrammetry is employed for numerous applications such as restoration of historic
monuments and ancient musical Instruments, analysis and surveillance of some structures (dams,
ship propellers, etc.) as well as in medicine. Its most common use, however, remains production of
topographic maps.
270 Processing of Remote Sensing Data

14.3.1 Orientation of photographs

A vertical photograph is usually oriented in the same way as a map, i.e., when it is on a horizontal
plane— which is most common— north is in the farthest position from the observer and when it is in a
vertical plane, north upwards. For quick appraisal, it is considered that in the Northern Hemisphere
shadows are more or less directed northwards. Otherwise, points identifiable in the photo and a
topographic map should be taken for identifying directions. Segments are drawn to connect various
points and the photograph arranged In such a way that the directions of the corresponding segments
are parallel. It will be noticed that the directions of the segments do not correspond exactly when high
relief is present.

14.3.2 Use of a stereo pair

A stereo pair consists of two photographs having a certain percentage of overlap. The two photographs
1 and 2 should be placed in the order in which they were taken during the mission; otherwise a
pseudo-stereoscopy is obtained. The centres A and B of each photograph or their principal points are
marked (Fig. 14.15). On each photograph, the Images a and b corresponding to the principal points
are marked and the straight lines Ab and Ba are drawn. The two photographs are adjusted in such a
manner that these two straight lines are coincident. To achieve good stereoscopic vision the distance
Aa and Bb should be equal to the interocular distance y.

I ^ 1 -.. B

1 ---------------- ! » ______ 2

1
.b . ■
a
• B

1
■ • 2

Interocular distance y
 Processing and Interpretation 271

14.3.3 Scale of a photograph

■ Measurement of scale
The scale of a map Is defined as the ratio of a distance measured on the map to the same distance on
the ground. The scale of a photograph can be defined in the same way. Let N bea reference plane of
altitude n above sea level at a distance Hfrom the objective of the camera (Fig. 14.16). For example,
point G and points U, V, Pand D, projected respectively from the points (on the ground) O (photograph
centre), R (river), A (plateau) and C (hilltop) are located on the reference plane. On the photograph,
these points are located at g, o, v, p and d Scale of the photograph E is hence given by:

E =og/U G or E^f/H

Hence, at a given altitude, such as for example above a contour line, the scale of a photograph is
constant if the axis of the camera is strictly vertical.
The focal length of the camera and height of the platform hence determine the scale. In the
example of the Bourg-Saint-Andéol-Nyons regular mission (sec. 14.1.5), the scale of photographs on
the reference plane was:

£=151.83/4590 or £=1:30,231

This precision of computation assumes there was no relief (not at all the case for this mission)
and that the aircraft flew at exactly the same height.

■ Scale variations
In practice, variations occur in the scale of aerial photographs due to defects in the objective of the
camera, deviations in the optical axis relative to the vertical and topographic fluctuations.

Fig. 14.16: Geometry of a vertical photograph.

272 Processing of Remote Sensing Data

Scale variations due to defects in the objective are relatively very small. Evidently, camera objectives
are never perfect since small distortions always exist. However, these are totally negligible for specialised
aerial photographs. Nevertheless, wide-angle optics of commercial cameras should be carefully
checked.
Scale variations due to fluctuations In Inclination are more common. In fact, these variations are
essentially due to modifications in the attitude of the aircraft (or satellite): common causes are roll,
pitch and tilt (see Fig. 14.6). For example, an inclination variation / in the angle of an 18 cm x 18 cm
photograph, taken with a focal length of 125 mm, causes a displacement of that results in a relative
error erin the distance measured on a half-diagonal (Table 14.2).

Table 14.2: Scale variations due to inclination variations (in degrees)

Inclination / Displacement d Relative error e r

1G 1 .2 mm 1%
3G 3.6 mm 3%
5G 5.8 mm 4.8%

Large variations in scale are produced by relief variations.

For the plane passing through Oand R, situated at the same altitude (Fig. 14.16), the scale of the
photograph is given by:

El = or/ OR Hence, o r- El - OR and El = f/H^.

For points U and V, which are equivalent to O and R but at a greater altitude and hence closer to
the aircraft (H < H^), the scale would be:

E= ov/UV, and as OR = UV, we get E= ov/OR, i.e., ov= E • OF?and E= f/H.

Consequently,

o r/o v - E1/E, i.e., or/ov= H/H..

Since oris smaller than ov(Fig. 14.16), scale El for OR is smaller than scale E for L?\/.Thus the
points A, C and E on the photograph will be at a larger scale than those of the reference plane N.
As a result, the scale of a photograph can be determined only when the photographed zone is
entirely at the same altitude. In fact, scale can be measured only on a reference plane assumed to
represent, at best, the mean altitude of the region under study. Contrarily, scale can be measured at a
point as well as on a contour line.
The concept of scale can be generalised as follows (Fig. 14.17). If the height between the objective
of the camera and the ground is measured starting from the objective, an inverse relationship is
obtained between scale and this height.

If sensor-object height decreases, scale E increases.

In a single photograph, height decreases when passing from a plain to a hill and to a mountain; hence,
scale Increases (Fig. 14.17, right).
For a given terrain, with the same camera (same aperture angle), height decreases according to
whether the platform is a satellite (800 km), a space shuttle or an aircraft (8 km). Hence for this terrain,
scale increases when passing from satellite photograph to aerial photograph (Fig. 14.17, left).
Similarly, the effective field of view is proportional to this height for an aircraft, the effective field of
view is smaller than for a shuttle or a satellite.
 Processing and Interpretation 273

Three fields of view for the three platforms

Fig. 14.17: Relationship between scale, height of camera and field of view (in this diagram,
focal point of camera is fixed and altitudes varied).

14.3.4 Aerial photo mosaics

If the photographs are assumed to be really vertical and after identifying the principal points of all the
photographs, they can be positioned with respect to one another on a topographic map. Thus, an
assembly oi aerial photos is obtained. This facilitates a quick recovery of photographs. Such assembled
charts can be consulted for regular missions of IGN.
A mosaic corresponds to an assembly of aerial photographs prepared for obtaining a view of the
entire area of an aerial survey.

■ Non-controlled mosaic
A non-controlled mosaic is prepared by matching the edges of aerial photos relative to one another.
Obviously, only one-half of each photograph is required. The other half is preserved for stereoscopic
observation. A satisfactory result Is obtained if we attempt to determine only a small field of view that
necessitates only a small number of photos. In fact, due to scale distortions, we cannot have a map
from these mosaics. Correct matching Is obtained over a single strip of coverage since overlap is
sufficiently high, but this Is rarely so between strips. If the relief Is too high, the result is rarely acceptable;
however, an impression of continuity is obtained and this facilitates marking reference points.
Stereoscopic study of such a mosaic can be carried out using the other part of the photographs.
274 Processing of Remote Sensing Data

H Semi-controlled mosaics
To prepare semi-controlled mosaics, first a mean scale of all the photographs is determined. Secondly,
a kilometric square grid is drawn on a large sheet, corresponding to the area studied, with the mean
scale of the mosaic. Using a topographic map, co-ordinates of the photo-centres are determined and
marked on the grid. Lines of centres are drawn on the grid. The central points on the grid are made to
coincide with the centres of the photographs and, by rotation, the lines of centres on the sheet and the
photographs are matched. Quite often some parts are not in contact between two successive photos.
This is secondary since the main aim Is to maintain angles and distances correct as far as possible.
Photographs can be divided into groups In such a way that the best possible matching is achieved
between them; such a division can be made, for example, at changes in altitude. Stereoscopy can be
obtained with these mosaics if two sets of photographs are available.

■ Controlled mosaic and rectification

Controlled mosaics are prepared based on the same principle as given above but using rectified
photographs.
Rectification refers to reconstruction of a photograph for which the optical axis of the camera is
strictly vertical and the scale is uniform irrespective of altitude of the terrain. Photographic rectification
is obtained through projection by distorting the photo in such a manner that images of four marker
points of the photo coincide with the corresponding four points on a topographic map. An assembly of
these rectified aerial photographs is also known as photomap or orthophoto. These mosaics provide
more Information on land cover and land use than topographic maps.

14.3.5 Radial distortions

As a photograph Is a conical projection, distortions during projection are of the radial type. The water
tower CE on the ground in Fig. 14.16 projects as ce in the photograph. If It were a planar projection,
i.e., an orthogonal projection, points Cand E ought to coincide in the projection, but in the photograph
ce is a segment and not a point. The nearest point from the photo centre c has the lowest altitude.
Flattening of the water tower CE in the photographic plane has the effect of shifting the lowermost
points towards the centre of projection and elevated points towards the circumference.
Points P, below A, and D, below C, evidently appear coincident on a map. However, on an aerial
photograph, each corresponds to a straight-line segment, pa and dc. These segments are formed
since the highest points are pushed outwards and the lowest points towards the photo centre, conforming
to conical projection.
The higher the points, the farther they are from the centre.
For two objects of the same height, the lengths of the two corresponding segments on the
photograph depend on their position in the photograph. A segment in the centre of the photograph will
be smaller than on the border. Measurement of segment length does not directly give the height of an
object.
The farther from the centre, the larger the distortion.
Thus, for a focal length of 125 mm, a scale of 1:25,000 and at 10 cm from the photo centre, an
altitude difference dh gives rise to a radial distortion drior a true scale E (Table 14.3).
A consequence on flight plan is that In the case of high relief difference, one is forced to take
repeated photos by changing the reference altitude, in order to cover the entire zone by stereoscopy
(case of insufficient coverage) and to obtain photographs of approximately equal scales.
As mentioned earlier, scales vary and the same length on the photo does not represent the same
distance on the ground. So it Is evident that two slopes which appear similar are not necessarily equal
on the ground. Therefore, one has to be cautious in evaluating slopes in a photograph. Nonetheless,
 Processing and Interpretation 275

Table 14.3; Radial distortion of a photograph due to altitude difference

dh dr E

1 0 m 0.3 mm 1:25,000
1 0 0 m 3.2 mm 1:24,200
500 m 19 mm 1 :2 1 , 0 0 0

the best interpretation of morphology and relief of a region is achieved from stereoscopic vision and a
three-dimensional virtual model.
Stereoscopic vision is obtained by observing two successive photographs. As the aircraft moves
between the two photos, the optical centre of the camera Is also displaced. Thus, objects seen in the
first photo are not at the same position in the second photo and their radial distortions are not represented
by the same segments. Based on these various considerations, three-dimensional relief can be
reconstructed.
None of this applies to sun-synchronous satellites such as SPOT.

14.3.6 Parallax and altitude determination

In two successive photos 1 and 2 separated by a distance equal to the air base B, the centres and
N2 of the two project as and H2. Point A projects on the two photos as and a2 (Fig. 14.18). The
relative displacement of point A between the two photos is called parallax p. To represent this parallax,
a pseudo-image of point A is drawn at the focal distance of the camera. The difference in angle of view
of point A In the two photos determines the value of parallax at the focal distance f in the pseudo­
image:

p = 3 2 ri2

We get the relation f/H = p/B. As B and fare constants for a given stereo pair, i.e., B f= K,

H = K/p

Altitude and parallax are inversely proportional.

If there is no difference In altitude for a stereo pair, H is constant and hence p is constant. Any two
homologous points will be separated by the value of p. Contrarily, If an altitude difference exists between
point B and point A, their heights H will differ and hence the parallaxes also differ. Altitudes can be
computed by measuring parallaxes In the two images. For this, the distance between the Images a
and a* of point A is measured in each photo of the stereo pair. A micrometric screw gauge is used to
measure distance to a tenth of a millimetre.
Thus, parallaxes of three points A, B and C in a stereo pair were measured and the following
values obtained:

aa' = 7.73 cm bb' =7.95 cm cc' =8.17 cm

A is at an unknown altitude, B at 500 m and C at 600 m. Consequently, the difference in parallax

between B and C Is:

cc' - b b ' = 8 .1 7 -7 .9 5 = 0.22 cm

for an altitude difference of 600 - 500 = 100 m.

276 Processing of Remote Sensing Data

A/i

Fig. 14.18: Parallax.

Thus, a parallax difference of 0.22 cm corresponds to an altitude difference of 100 m.

As bb’ - aa' = 7.95 - 7.73 = 0.22 cm, it can be inferred that the altitude of A is 400 m.

14.3.7 Stereo restitution—Orthophotos

Restitution consists of conversion of a conical projection to an orthogonal projection. Aerial photographs
are thus converted into maps. Stereo-restitution Instruments facilitate this conversion of photos into
maps. At present, most topographic maps are refined using aerial photographs or, more recently,
satellite Images.
Stereo-restitution equipment comprises three components, viz., a projection system, an observation
system and a plotting system.
The underlying principle is to reconstruct an elevation model (or plastic Image) exactly similar to
the object by creating a projection system based on flight parameters and ground reference points for
 Processing and Interpretation 277

which X, y and z co-ordinates are accurately known. The procedure includes a relative orientation of
the photos, determination of scale and absolute orientation. Preparatory operations, which hitherto
took a very long time, are now done using microcomputers that provide much more accurate
computations than earlier.
A hyperstereoscopic viewing system enables the operator to fuse, over an elevation model, two
cells Into a single one for a given altitude. While modifying the x and y co-ordinates of a cell, the
operator can follow a contour line, since if he (she) moves away from it the two cells get dissociated.
The method of anaglyphs or the method of vertographs is presently used.
Lastly, a tracing system permits drawing the details to be mapped (planimetry, contour lines, etc.)
using either a pantograph or a computer by storing in memory the co-ordinates of points situated on a
contour line. Nowadays, the contour line is directly drawn on the three-dimensional image.
Digital stereo-restitution systems also exist, which can be directly used for analysis of digital
images, such as those of SPOT, by means of digital correlation of images. A digital elevation model
(DEM) is obtained in this way. A DEM consists of a series of points for which geographic position and
altitude are determined.
Orthophotos are obtained automatically using a differential rectification process that corrects for
scale variations of aerial photographs. Some specific elements of topographic maps are often drawn
on orthophotos. This product is developing rapidly since it can be used in geographic information
systems. Orthophotos generated from aerial photographs or satellite images can thus be superposed
over one another. Recent data-processing software includes systems that enable this transformation.

14.3.8 Historical background of aerial photography

Study of the Earth’s surface from an elevated view dates back to the Bronze Age.The first interpretations
of such views were discovered at Mont Bego and Val Camonica in the Alps. Shepherds of this period
carved a landscape in the rock that overhangs a valley with fields, farm enclosures (planted or empty),
cattle and habitations. It is noteworthy that they depicted rectified shapes of fields, viz., squares
representing fields that were distorted by perspective.
In the Middle East, land survey records existed, since a perspective plan of the village Telloh
(4000 B.C.), inscribed on baked clay bricks, was found.
In Chaldeens, land surveys must have been carried out minutely for fiscal reasons.
Up to the 16^*^ and 17^^ centuries, maps with imaginative drawings could be distinguished from
very accurate land survey records In which figures of various objects such as trees, bushes, fences,
etc., are depicted, giving a synoptic view similar to that of aerial photos.
In 1812, Nicéphore Niepce invented a procedure for obtaining negatives and positives. In 1839,
Louis Daguerre used them to make direct positive proofs from the black camera, the ‘daguerreotype’
could not be printed in multiple copies: photography existed. In 1840, Francois Arago, Director of the
Paris Observatory, advocated the use of photography for topographic surveying. In 1841, William H. F.
Talbot patented a system for producing a negative that enables printing of multiple positives.
The first known aerial photograph was taken in 1858 from a balloon at a height of 80 m above
Petit-Clamart by Félix Tournachon, known as Nadar. He demanded patent for invention ‘of a new
system of aerostatic photography’.
On 13 October 1860, James Wallace Black and Samuel A. King took a balloon photograph from
365 m above Boston (United States), which is the earliest existing aerial photograph to date. In 1862,
Col. Francois A. Laussedat prepared plans for Paris fort using ground and aerial photo-topographic
sketches. During the War of Secession in America between 1861 and 1865, photographs were taken
from balloons.
The term photogrammetry came into existence in 1876.
After 1878, when collodion plates were replaced by gelatine-silver bromide emulsions, many
aerial photos were taken. The earliest known aerial views of Paris were taken by Tallandier and Ducom
278 Processing of Remote Sensing Data

on 19 June 1885 during the rise of a free balloon. The first aerial photograph taken from a kite is
credited to E.D. Archibald, an English meteorologist (around 1882).
In 1886, in North America, aerial photos were used for topographic mapping. In 1890, A. Batut of
Paris had published a book on aerial photography. Between 1854 and 1898, the Vallot brothers mapped
Mount Blanc using aerial photos.
In 1897, Scheimpflug conducted aerial surveys along parallel traverses at a constant altitude with
an airship. Photos taken at a regular interval covered more than half of the zones and two adjacent
strips partially covered their boundaries.
Thiele, a Russian surveyor, used aerial photos for topographic purposes in Trans-Caucasus from
1898 to 1908.
On 28 May 1906, G.R. Lawrence photographed San Francisco immediately after the earthquake
and the great fire with a panoramic camera.
The first photomap, dated 1908, was prepared by the Italian scientist Tardivo for archaeological
investigations.
On 24 April 1909, the Wright brothers took the first motion photograph from a plane over Centocelli
(Italy).
Aerial photographic technique developed during the First World War of 1914-1918. Development
slowed down after the war and again continued with the advent of equipment and methods of
photogrammetry, which enabled preparation of aerial maps covering large areas. An aerial survey
was carried out from 1919 to 1923 for resettlement of 1500 communities. The United States Department
of Agriculture (USA) launched a programme of aerial coverage in 1937.
In the Proceedings of the French Academy of Agriculture of 1919, Captain Bouché outlined the
possible applications of aerial photography for agriculture: ‘reconstruction of land records, details of
land division, particulars of crops in a farm, inventory of trees in orchards and other plantations.. .Study
of form and use of land according to geologic formation, surface relief, investigation of slopes, etc’. All
these are visible in aerial photographs taken at altitudes between 500 and 1000 m.
In 1942, Kodak company produced black and white Infrared emulsion and later colour emulsion.
More than fifteen years later, false colour emulsion, now known as colour infrared (CIR), was developed.
The Army Geographic Service (SGA) up to 1940 and later National Geographic Institute ensured
aerial coverage of all of France and for a number of years, overseas regions. At present, the entire
territory is photographed on average every five years or less. Aerial photographs are sold at IGN.

14.4 CONCLUSION
Aerial photography has long been employed for topographic and thematic investigations for which a
large number of surface maps have been prepared using aerial photographs. They are used as guide
maps and essentially for delineating boundaries between map units.
The first Interpretations of colour emulsions were made using aerial photographs, which constituted
a step forward In the direction of development of modern remote sensing. Aerial photographs have
become complementary to satellite images. The latter evidently have the advantage of a much wider
field of view, while the small angle of vision of some sensors gives images close to an orthogonal
projection. However, photographs even today have the advantage of better resolution than images,
but perhaps not for long.
For several years satellite photographs have been acquired which can be processed and Interpreted
as aerial photographs, particularly with respect to stereoscopy.
Use of the term ‘aérospatial remote sensing’ clearly indicates the complementary nature of these
sources of information.
 Processing and Interpretation 279

References
Société de photogrammétrie et télédétection. 1985. Bull. SFPT, 99, Saint-Mandé, 69 pp.
Ducher G. 1985. Intérêt, exploitation et avenir des prises de vues photogrammétriques de Spacelab. Bull. SFPT,
99:13-14.
Girard C-M, Girard M-C. 1985. Interprétation du paysage à petite échelle à partir de clichés de la chambre métrique
spacelab. Approche botanique et Pédologique, Bull. SFPT, 99:41-51.
Girard M-C, Girard C-M. 1970. Cours de photo-interprétation. Polycopié. INA-PG, Grignon, 208 pp.
Konecny G. 1985. La mission photogramméthque Spacelab-1. Bull. SFPT. 99:5-12.
Schuhr W, Engel H, Konecny G, Lohmann P, Schuring A, Wu J. 1984. Investigations of metric camera data quality.
Inter. Archives of Photogram. Remote Sens., 25 (part A1): 64-69.
 D
QUALITY
ASSESSMENT
 15
Scale Changes

15.1 INTRODUCTION
The term ‘scale’ is often used ambiguously. When a manager envisages processing a ‘large-scale’
problem, he implies a vast spatial field. From the cartographic point of view, such vast areas are
represented on a ‘small-scale’ map, from 1:250,000 to 1:1,000,000 for example. Thus the term has
opposite meanings. We prefer to use the term scale only in its cartographic sense, i.e., as the ratio
between a distance on the map and the corresponding distance on the ground. The main categories
of scales used in analog (paper) maps are given in Table 15.1.
Moreover, the concept of scale hardly has any meaning in processing of remote-sensing data
since these data correspond to pixels semantically represented by digital numbers. For about a decade
the question of ‘scale change’ has been the subject of several Investigations and discussions
consolidated by theses and scientific articles as well as in seminars and study reports (GSTS-CNRS,
CEMAGREF, International Journal of Remote Sensing, SFPT, etc.).

Table 15.1: Major categories of map scales

Scale category Dimensions

Local 1 :1 0 , 0 0 0 or higher
Regional 1:10,000 to 1:50,000
National 1:50,000 to 1:250,000
Continental 1:250,000 to 1:1,000,000
Global 1 : 1 ,0 0 0 , 0 0 0 or lower

15.2 SCALE AM) ORGAMSATIONAL LEVEL OF MEDIA

The objects of the Earth’s surface exhibit many levels of organisation, from molecule to biosphere. A
level of organisation corresponds to the set of relationships and functions between objects that manifest
a geometric, semantic and functional coherence, defining a new entity or new pattern. A level of
perception corresponds to the tools, methods and estimates employed for detecting a given
organisational level. A visual observation using a microscope (perception level) enables study of a
vegetative tissue (organisational level). However, perceiving the organisation of a forest canopy requires
an aerial view. The levels of spatial organisation of vegetation as well as the order of magnitude of its
objects and means of perception used are shown in Table 15.2.
Organisational levels are related to the components of ecosystems, which are of two types, viz.,
abiotic and biotic. Abiotic components are climate, geologic substratum, hydrological regime, soil,
284 Processing of Remote Sensing Data

Table 15.2: Levels of spatial organisation of the plant world.

Each level includes properties of the lower level besides having new properties resulting from
the structure of association of lower-level components

Organisational level(*), Dimensions Examples Means of perception

Nature of objects (order of magnitude)

Macromolecule 1 0 "® m DNA

Organite 1 0 "® m Chromosomes, Electron microscopy
leucoplast
Cell 1 0 -6 to 10-4 m Cell Microscopy
Tissue 1 0 "^ to 1 0 "2 m Parenchyma
Organ 1 0 “ ®to 1 m Leaf Laboratory
Individual organism 10-®to50m A daisy Photography Spectrometry
Subpopulation 1 0 "® to 1 0 m A spot of daisies Field observation
in a grassland
Natural population 1 0 "® to 1 0 ®m Daisy population in
a group of grass­ Photography Radlometry
lands, not necessarily
limitrophs
Plant community 1 0 ""* to 1 0 ®m Grasslands on
or synusia rich soil Airborne and satellite
Vegetation formation 1 0 to 1 0 ^m Grasslands and observations
or biome pastures
Vegetation landscape 1 0 ®to lO^m Agricultural plains
Ecological region lO^to 10® m Euro-Siberian region Aerial photography
Biogeographic region 10® to 5 X 10® m Holarctic zone Multispectral data
Biosphere 4 X 10^ m

(*) From lower to higher

etc., and biotic components are flora, fauna and man. The latter can be used as a basis for mapping.
Generally, climate, geomorphology and substratum belong to small-scale features, while fauna,
vegetation and soil are of large scale.
This hierarchic division of objects can be continuous or discontinuous depending on the
organisational level; some changes in level correspond to thresholds or stages, as shown in Fig. 15.1.
Cartographic representation of various levels of perception is done differently depending on whether
a perception level is between two thresholds or at the level of a single threshold.
In the case of organisational levels for which changes occur continuously and for small ratios of
reduction, changes in perception levels can be obtained without changing the number of units by
moving from particular to general (note that the reverse approach of general to particular Is prohibited).
To change the scale of a map from 1;25,000 to 1:100,000 without changing the number of map units,
we can use a mechanical method such as photographic reduction, change of zoom factor, subsampling,
smoothing, etc.
Contrarily, when changes in organisational level correspond to a threshold, the number of units
needs to be decreased. For this, we can proceed either by regrouping the units according to methods
 Quality Assessment 285

Organisation

Number of units No change Decrease Decrease

Simplification of criteria Establishing new concepts

Transformation Mechanical reduction Logical generalisation Synthesised

generalisation

Fig. 15.1: Organisational and perception levels and scale changes (after Guillobez and Bertrand, 1995).

of aggregation of classes, provided organisation of the latter is hierarchic (see Chap. 8), or by decreasing
or simplifying the attributes of each map unit and regrouping. Thus a logical generalisation is obtained.
If the thresholds are high or many, a new interpretation has to be made based on new concepts, as in
the case of changing from a thematic map to a landscape interpretation. This method necessitates an
in-depth analysis, knowledge of concerned processes and description of chorological laws. Thus a
synthesised generalisation is achieved (see Chap. 11).
It is therefore necessary to be cautious in using the term ‘scale’, which is ambiguous except when
reserved for the final restitution stage of a spatial dataset. It is hence advisable to be concerned with
the quantity of information which is a function of spatial field, resolution and level of analysis and with
levels of organisation, means of perception and precision. These levels should be coherent with spatial
resolution, level of analysis (number of attributes used for each pixel: In remote sensing, spectral
bands analysed), the field (extent of study area), size and nature of objects under study. Adequacy
between levels of organisation and perception is essential: cellular structure is not studied with binoculars
but rather with a microscope!

15.3 SPECIFIC ASPECTS OF REMOTE SENSING

15.3.1 Resolution
In remote sensing, objects can be characterised by two thresholds of resolution (Puech et al., 1995),
as illustrated in Fig. 15.2.
The minimum reso/i/f/on characterises a given object by a given radiometric value. This threshold
is reached when the ratio between the resolution r (pixel size, minimum size of observation) and the
size of the object 7 is about 2/3, r = 2/3 7. Below this resolution, the constituents of the object are
differentiated but not the entire object.
286 Processing of Remote Sensing Data

Fig. 15.2: Theoretical model relating variance of digital numbers with resolution (after Puech et al., 1995).

Thus a tree can be identified by a given resolution but, if the latter increases, the branches, leaves
and shadows produced by tree tops are perceived and not the tree as an object.
The maximum resolution is reached when the pixel is larger than the object observed and does
not permit characterisation of the latter by a precise digital value. A mixel is obtained In this case.
Thus, the maximum resolution of a tree is reached when the pixel size is such that there is a
mixing between the tree and the openings or skirts.
This model explains how forest covers that appear homogeneous in LANDSAT MSS images
(resolution 4424 m^) become heterogeneous in LANDSAT TM images (resolution 900 m^) or SPOT
images (resolution 400 m^). This type of model was shown with the software VOISIN for determining
the most adequate window size (or over sampling resolution) for identification of a complex object
(Chap. 11).

15.3.2 Adequacy between level of observation and level of

organisation: Image segmentation
When a remote-sensing image is to be classified, it Is desirable to ensure the best possible adequacy
between pixel dimensions and sizes of objects to be studied. In other words, remote-sensing data
most suited for a given investigation should be chosen (see Chap. 16). As a matter of fact, the basic
elements to be analysed, viz., the pixels are situated between two levels of organisation rather than
corresponding to one of them. Sometimes the data available with the required repetition rate are at an
inadequate level of observation and it becomes necessary to go below the pixel resolution for obtaining
a correct classification. Moreover, mapping, one of the most common applications of remote sensing,
assumes certain conceptualisation of data as well as a minimum size of units facilitating their
representation graphically. Normally, the pixel dimensions in remote sensing do not correspond to the
dimensions of map units In the final maps and many mixels exist in the borders. Lastly, It may also be
desirable to formulate scientific objectives of data corresponding to other levels of organisation for
simulating the data before their acquisition by new sensors and preparation for use, to evaluate their
possible applications.
Several solutions exist for restoring the levels of satellite observation with levels of organisation:
either a pixel Is resolved Into Its constituent elements, or pixels are combined so that they correspond
 Quality Assessment 287

to a higher level of organisation, or isolated pixels are not classified but their neighbourhood is taken
into consideration.
The objective of segmentation is to divide an image into zones corresponding to the objects
present on the ground, these zones serving as a base for subsequent processing. Masking (Chap. 7)
is an example of segmentation. In aTM or SPOT scene covering a montane region it may be desired
to segment the image into various groups according to exposure of slopes. These groups represent
an observation level intermediate between the level perceived by the pixels and that of the study,
which corresponds to a regional level.
The conceptual model of landscapes implied in image segmentation assumes that they constitute
a mosaic of homogeneous objects, irregular in shape, of dimensions greater than a pixel. The
segmentation procedure consists of minimising the intragroup distance of digital numbers of the
segmented zones. It is assumed that the objects constituting a landscape have a low variance and a
single value of internal variance. However, verification of the latter is rare since various land-cover
categories may have very different degrees of internal variance.
This is one reason for using the maximum-likelihood method for segmentation. A joint hierarchic
model can be employed for taking into consideration various levels of organisation present in a
landscape. The joint hierarchic model combines on the one hand, the assumption that a mosaic of
discrete objects exists, as in a homogeneous-constituents model, and on the other, the assumption
that an explicit hierarchy exists that defines relationships between these objects. The hierarchic levels
correspond to the organisational levels of various processes existing on the ground and to the
chorological laws determined during the study.
For example. In an agricultural region, the first level may correspond to variations within a parcel
(heterogeneities in the canopy) due to effects of competition between individuals and microclimatic
and edaphic conditions on the latter. The attributes correspond, for example, to differences in vigour or
development of aerial phytomass.
The next level will be that of parcels: plots occupied by various crops. In this case, the
canopy occupying a plot is considered a unit In which intraplot heterogeneities are smaller than
interplot heterogeneities. The causative factors would be agricultural calendars, phenotype
characteristics, rotations and farming systems, etc. The attributes correspond to crops occupying
each plot.
The next higher level corresponds to organisation of plots in space according to chorological laws
based on major types of soils, geomorphology, regional climate, socioeconomic conditions, etc. The
attributes would be, for example, combinations of various land covers that define landscape units.
Some examples illustrating various approaches to achieving better correspondence between
organisational and observation levels and the requirements of accuracy are presented below.

15.4 RESOLUTION OF PIXEL INTO CONSTITUENTS OR

DESCENDANT APPROACH
The presence of mixels In remote-sensing data gives rise to errors in pixel-by-pixel classification. In
fact, mixels are characterised by digital numbers corresponding to none of the themes to be classified
(mapped) or have the same values as those of pure pixels belonging to other themes. This situation
may hence lead on the one hand, to semantic errors in assigning pixels to one or the other class using
a distance computation for example, and on the other, to geographic errors In determining boundaries
between two classes. While this boundary can be identified using filters (Laplacian or another),
mathematical morphological operators (Chap. 12) or interpolation methods, the thickness of the
boundary line may not be less than 1 pixel and it is not directly known In which of the two intermediate
zones it Is to be drawn.
288 Processing of Remote Sensing Data

Consequences of these errors in mapping can be significant, particularly when mapping is used
to estimate the area of a given land cover or to monitor the diachronic variations in areal extent. To
minimise these errors, it Is necessary to identify boundaries of negligible thickness compared to the
size of map units and the latter should be drawn as precisely as possible vis-à-vis the zones to be
delineated. This assumes that mixels can be resolved into their constituents. Two types of mixels can
be defined (Fig. 15.3):
— ^those corresponding to a boundary between two units: boundary mixels;
— those corresponding to a mix of constituents pertaining to two or more different units: mixed
mixels.

1 2 3 4
A
D
B □

C I I

D
r

o ^ , A3,B1,C1,C2, C3,
Boundary mixels a 2, B4, C1, C4, D2 Boundary mixels p2
A1, A3, A4, B1, D1, A1, A2, A4, B2, B3.
White pixels White pixels
D3, D4 B4, D3, D4

Grey pixels B2, B3, C2, C3 Grey pixels

Fig. 15.3: Types of mixels.

Figure shows two thematic units, grey and white, extending over 16 pixels marked in rows (A to D)
and columns (1 to 4). I: Boundary mixels, II: Mixed mixels.

The digital value of a pixel corresponds to a double integration of geometric and radiometric
parameters of radiance of various units present In it. It is hence considered that in the radiometric
space corresponding to the spectral bands under analysis, any digital number of a pixel lying between
the values pertaining to two pure map units represents a combination of these two units in various
proportions. If the geographic dimensions of the units under consideration are greater than the geometric
resolution (pixel dimension), as in the case of Fig. 15.3 (I), there is high probability that the mixels are
boundary mixels. Contrarily, if the dimensions of the units are less than those of a pixel, they represent
mixels comprising a mix of units (Fig. 15.3 II).
Boundary mixels can be identified by the following method proposed by Blamont and Grégoire
(1994). It is based on the concept of local heterogeneity defined by the standard deviation computed
from the digital values In a 3 pixels x 3 pixels window. This value is assigned to the central pixel of the
window. Computation is done for the band In which variations in digital numbers of units are high, for
example the infrared band (XS3).
Clusters of digital values of units are determined either from field reflectance measurements, or
from zones identified in the image as uniquely belonging to one of the units under consideration.
Convex envelops and major axes of these clusters are drawn (Fig. 15.4). The dividing line is represented
in this example as corresponding to a composition equal to 50% of each of the two pure units. In
 Quality Assessment 289

Fig. 15.4: Clusters of digital values of two pure units and the dividing line.

similarity between separation of two units in radiometric space and their geographic distribution, it is
estimated that the smaller the distance of a pixel from this dividing line in radiometric space, the higher
the probability of a mixel being a boundary mixel.
These pixels are then resolved according to the proportions of various units. Since the digital
value of a pixel is the weighted average of radiance values of the various units it comprises, the pixel
can be resolved into its constituents. For a pixel Q of radiance L, several possible In-situ combinations
(compatible mixtures) of homogeneous units exist, giving the value L Some examples of compatible
combinations are given in Fig. 15.5.
Thus, only the value of L and of digital values of pure units need be known to determine surface
areas corresponding to each. The existing pure units have first to be characterised. For this, several
methods belonging to two major groups can be employed.
The first group consists of methods In which clusters of digital numbers are represented by points
In cluster of /c units. In this case the signal measured by the sensor is considered a simple sum of
signals reflected by each component /cof radiance L (/c).This assumes that radiance values L (k) are
known; supervised classification or tables of spectral characteristics can be used for this purpose.
Various constituents in the pixel Q are regrouped at the most Into n + 1 classes (n being the number
of bands). If /c = n + 1, a unique solution is obtained; contrarily, if /c> n + 1, infinite solutions are
possible. Various statistical methods (such as analysis of textures and variations in radiance around
each pixel or multitemporal regression) need to be employed to complete the information missing in
such cases. Such methods assume that variability in radiance values is sufficient. Quantisation errors
can be as high as 75% depending on the types of media and number of bands available.
The second group consists of methods that represent clusters of digital values of pure units by
extended domains and boundaries by thresholds in the cluster of /c units. These thresholds must
first be defined, for which we can use mean and standard deviation, minimum and maximum values In
each band, threshold indices, distances between classes, etc. The percentage of overlap of a
homogeneous unit k \n a pixel Q can be computed by a measure of similarity or by calculation of
distance.
The disadvantage of these methods lies in assigning 0 or 100% of the pixel to a unit, due to which
heterogeneity of the composition of pixels is not taken into consideration. Only the methods based on
fuzzy groups or neural networks avoid the disadvantages of binary classification of linear threshold
models. These methods involve study of all possible proportions (that can be defined by a lower limit
and an upper limit) of various units contributing to the value L, based on knowledge of digital value of
pure units. Rather than assigning a single composition to a pixel, the degree of similarity (Intensity) of
its appurtenance to various units Is evaluated.
290 Processing of Remote Sensing Data

L {Q )

Fig. 15.5: Examples of compatible combinations for a pixel Q of radiance L (after Grégoire, 1995).

Methods using fuzzy groups have the disadvantage of lengthy and time-consuming computations
whereas neural network methods, once tested, are much faster. In neural networks, the difficult task is
to choose a balance between various datasets used for defining subpixel compositions. The latter Is
done by successive approximations which minimise differences between the subpixel compositions
obtained and the true subpixel compositions of a dataset.
The methods of quantisation of proportions of various units consider that these pure units are
radiometrically different, i.e., each unit corresponds to a distinct group of digital values. This
assumption may not always be valid when the units are defined based on field data. For example, let
us consider a single agronomic unit, viz., ‘permanent grasslands’, for a scene acquired on day y; it
corresponds to at least two groups with different digital values: grass plot yet to be harvested and that
just harvested.
Moreover, to extract clusters of units or their proportions inside mixels from a matrix of digital
numbers of pixels of n bands, it is necessary to have a sufficient quantity of pixels covering the entire
range of possible proportions or a large number of spectral bands in which each unit Is characterised
by a specific cluster of digital numbers. In practice, it is most unlikely to observe In a scene several
pixels with different proportions of units having constant digital values while the number of spectral
bands is generally limited. One solution is to use diachronic data of the same region with a greater
number of spectral bands; however, this does not solve the general problem of distinction of groups of
different digital values for characterising a land cover unit.
 Quality Assessment 291

15.5 AGGREGATION OF PIXELS OR ASCENDANT

APPROACH
This method consists of generating a new dataset, corresponding to a more global level of organisation,
from a given dataset (such as ground measurements, remote-sensing data of high geometric resolution,
etc.).
An objective of this approach could be to obtain digital values of pure units. Fischer (1994) used
as basic data field reflectance measurements over various crops taken at the rate of one measurement
for every 7 days during a crop season. Individual temporal profiles of vegetation index (NDVI) computed
from these data showed that they were all of the same form irrespective of the type of crops. This
enabled formulation of a model of temporal variation of NDVI for homogeneous canopies. Using the
temporal profile models of winter crops on the one hand, and spring crops on the other, he computed
regional temporal profiles of NDVI such as would be acquired by a satellite of low geometric resolution
like NOAA/AVHRR or SPOTA/EGETATION. Such data can be subsequently integrated in a model of
yield estimation for example.
Another application related to change of organisational level could be to determine the spatial
composition of spectral units corresponding to a higher level from an image corresponding to a lower
level of organisation, classified pixel by pixel, supervised or not. A higher level unit would be
heterogeneous, comprising several spectral classes derived from the preliminary classification, some
with, others without thematic significance. One or several classes of the first classification could be
dominant in the composition of the unit. Knowing this spectral composition of units, the entire image is
analysed using a moving window whose size is chosen according to the degree of heterogeneity of
the unit to be mapped (see VOISIN model. Chap. 11). Too large a window leads to excessive
generalisation, while too small does not adequately take Into consideration the spectral classes for
characterising a unit.
It may also be desirable to simulate data to a coarser observation level from data acquired at
higher resolution (MSS, TM or SPOT). These data are degraded to obtain an observation level (here,
a pixel) of dimensions similar to those of a satellite with lower geometric resolution, such as NOAA/
AVHRR. Degradation methods are varied. One may start with a subsampling followed by computation
of mean or use a spatial filter followed by a resampling by cubic convolution. Various examples were
compared by Justice et al. (1989) for 5 x 3 pixel blocks. These included choice of the value of central
pixel, mean of the first 4 values in a row, mean of 15 values of pixels, mean of 4 non-adjacent pixels
chosen according to a constant pattern, median of 15 values, etc. The method of using the mean has
the disadvantage that pixels of very low or very high luminance are eliminated and, on the other hand,
values that did not exist in the original data are generated.
A new orientation In obtaining classification results, more precise or usable in more applications,
involves making an ordinal classification (dense, moderate or scarce forest coverage) In addition to
cardinal (forests, crops, villages, etc.) or using non-dominant characteristics to be assigned to classes
or second-choice classes (IDRISI software, for example).
In the first case, pixels of an Image of low geometric resolution are compared with corresponding
pixels obtained from degradation of an image of high geometric resolution which, in addition, is used
for classification of non-degraded pixels. Thus, information on subpixel composition and, for example,
on the proportion of a specific class inside a pixel is obtained.
In the second case, non-dominant characteristics (or second-choice classes) that can be assigned
to pixels or polygons are identified. For this, we can use models of combination of radiometric values
permitting determination of proportions of the various initial classes (see above) as well as neural
networks or a linguistic type of notation during verification of land cover on the ground. In all cases
precision of non-dominant characteristics (second-choice classes) Is dependent on the precision of
the initial classification and hence efforts of improvement ought primarily to be concentrated on the
latter.
292 Processing of Remote Sensing Data

In all these examples, the question of exact location of pixels and their correlation between data
of various sources is not resolved. The same applies to test methods and validation of results. Some
authors propose ground verification not over pixels, which are difficult to locate In situ, but over polygons.
In such cases, one should be cautious in evaluating the precision of interpretation results and, in
particular, in computing the Kappa estimator (see Chap. 19).

References
Blamont D, Grégoire C. 1994. Données télédétectées et précision de la cartographie des limites intra-pixellaires.
Bulletin SFPT 137: 98-102.
Fischer A. 1994. A simple model for the temporal variations of NDVI at regional scale over agricultural countries.
Validation with ground radiometric measurements. Int. J. Remote Sensing, 15:1421-1446.
Girard C-M. 1995. Changements d’échelle et occupation du sol en télédétection. Bull. SFPT, 140:10-11.
Grégoire Himmler C. 1995. Étude de l’hétérogénéité sub-pixellaire des milieux naturels observés par radiométrie
multispectrale: application à la modélisation de l’état et de l’activité de la végétation. Thèse Dr. Université
Louis Pasteur, Strabourg 1,145 pp. + ann.
Guillobez S, Bertrand R. 1995. Cartographie et changement d’échelle, le point de vue du naturaliste, propositions
d’applications en cartographie informatique. Bull. SFPT, 140:8-9.
Justice CO, Markham BL, Townshend JRG, Kennard RL. 1989. Spatial degradation of satellite data. Int. J. Remote
Sensing, 10:1539-1561.
Puech C, Doumerc F, Lieutaud A. 1995. Identification des objets selon l’échelle: apport des outils de SIG. Bull.
SFPT, 140:18-19.
Woodcock CE, Strahler AH. 1987. The factor of scale. Remote Sensing of Environment, 21:311-322.
 16
Criteria of Choice for the User
The important questions faced by the user of remote-sensing data are summarised in this chapter.
The various chapters of this book in which these questions are discussed in detail are referenced.

16.1 WHAT DATA TO CHOOSE?

Development of operational applications of remote sensing demands results that satisfy certain quality
criteria. But before reaching this stage and even before commencing a study, it is imperative to ask the
following four questions and obtain clear answers.

16.1.1 What is the question posed and the problem to be solved?

Some proposals do not clearly indicate the nature of the problem, which has implications on the
nature and pertinence of the methods used for solving the problem. For example, is the objective of
study to map flooded zones or to detect zones likely to be flooded every year?

16.1.2 Are rem ote-sensing data apt to answer the problem?

Some proposals specifically mention use of remote-sensing data although it is not fully justified; a
contrario, this is not because these data are not foreseen that they may not be useful. It must be
clearly understood that remote-sensing data cannot replace the field work but most often enhance the
value of the latter by making It more effective (aiding in classifying a region into homogeneous zones,
marking zones in which a phenomenon may be generated, selection of control and validation points,
spatialisation and generalisation of limited data, etc.).

16.1.3 What are the data acquisition conditions most likely to

provide an answer?
The characteristics of existing data (platform, sensor, spectral resolution, geometric resolution, date
of acquisition of scenes, etc.) and technical (acquisition, processing) and thematic (knowledge of
phenomena or objects to be detected and mapped) information must be Identified. This assumes
a good understanding of the problem to be investigated. In addition to acquaintance with remote
sensing.
294 Processing of Remote Sensing Data

16«1«4 Are rem ote-sensing data responding to the conditions

defined and problem posed available?
To ensure availability of the required data, organisations concerned with acquisition of aerial and
satellite data need to be consulted. If their response is NO, a special acquisition may be requisitioned
if possible in view of time and financial constraints; if the answer Is YES, the most suitable data must
be chosen.
Answers to these questions are facilitated by the Centre of the Earth Observation (CEO) established
by the European Union to assist individuals and organisations in the use of remote-sensing data and
to provide information concerning types of data available, costs, dissemination policy, presentation of
service societies, etc. The same applies to developing contacts through the Internet (see principal
addresses in the annexe): for example, consultation of image catalogues of SPOT and DALI, now
called SIRIUS (see the text on SPOT satellite in CD).

16.2 CRITERIA FOR CHOICE OF REMOTE-SENSING DATA

16.2.1 Choice of spectral and geometric resolutions
Choice of spectral and geometric resolutions must be made according to the type and phenomena to
be detected (see Chap. 15). For this it should be remembered that a definite relationship exists between
organisational level and observational level of a phenomenon. For example, attempts to detect
phenomena associated with pigment modifications of vegetative tissues (chlorosis, effects of diseases
or parasite attacks, etc.) in spectral bands other than the visible will be ineffective unless a relationship
exists between such modifications and other characteristics of tissues, such as water content. In the
latter case, reflective middle infrared or thermal infrared data may be useful. However, it should be
noted that if the relationships between parameters detected by remote sensing and phenomena studied
are more complex, the results will be less precise.
On the other hand, it should also be remembered that organisational levels are characterised by
thresholds. Choice of data of higher resolution is not always the best solution for identifying a
phenomenon. For example, a sparse forest canopy, which appears homogeneous in LANDS AT MSS
data, will look very heterogeneous in SPOT data, making its mapping difficult. In the case of MSS, the
pixels exhibit the combined effect of radiance of treetops and lower storey vegetation, while in SPOT
images the pixels represent either treetops or lower storey vegetation.
If there is a choice between several datasets, it is advisable to take images with the desired
geometric and spectral resolution but with a large field of view In accordance with the phenomena to
be studied. This avoids or reduces mosaicking and permits replacing the phenomenon In its regional
context. This possibility of analysing a phenomenon in a vast geographic field Is often advantageous
and validates the data for a wider range of applications. For example, wetlands over a large region can
be mapped (see Chap. 21) more readily with LANDSATTM images (scene covering 185 x 185 km)
than with SPOT data (60 x 60 km scene), while SPOT images or aerial photos can be used for more
detailed studies of small territories.

16.2.2 Choice of date of acquisition

Date of acquisition must be selected so as to facilitate ready detection either because the phenomena
are directly observable or because those not directly observable can be interpreted using simple
relationships, in order to study various soil phenomena it Is preferable to choose, if possible, seasons
 Quality Assessment 295

in which soils are bare so as to take advantage of differences in their radiance values, permitting direct
interpretation of their surface states. Contrarily, a period in which differences in vegetative growth
manifest is useful for analysing physicochemical characteristics of soils such as earlier flowering in
well-drained soils rapidly warmed up in spring, difficult sprouting of young plants, aqueous stress of
crops in shallow soils, chlorosis over carbonate outcrops, etc. If variations in water regime of soils are
of interest, a dry period after rains should be selected. If crop inventory is to be prepared, dates
corresponding to various physiological states and phenological stages should be used, which
necessitates good knowledge of modes of growth and development of species and agricultural calendar.
Quality of results will be better In the following situations:
— with 2 or 3 dates rather than one, especially when studying phenomena exhibiting large variability
such as crop identification, differentiation of diverse categories of permanent grasslands, processes
of floods or drying of soils, or when evaluating temporal variation of objects on the terrestrial surface
(magnitude of deforestation or development of towns, etc.);
— when acquisition is made in key periods fora theme, such as flowering of rapeseed or sunflower
plants, ripening of cereals, facilitating their differentiation from other crops. Differences in dry state of
bare soils can be detected after a rainy season; differences in thermal exchanges between various
land covers become visible after a period of frost or snowfall when it is melting. As much as a thick
layer of snow hinders observation of soil surface, that much a thin layer provides useful Information on
variations in thermal inertia or microclimatic conditions.
When diachronic data are employed It is preferable to use, if possible, data from a single satellite
and (in the case of SPOT) acquired at not too different inclinations, so that geometric and radiometric
differences (and hence corrections) are minimal. Radiometric correction is imperative if classifications
are undertaken with bands corresponding to various dates. In this case, data can be calibrated on
targets considered invariant (same reflectance throughout the year) such as water bodies, concrete,
pure conifer forest, etc. Caution: radiometric and some geometric corrections (to ensure superposition)
lead to m odifications in radiometric values and influence thematic classifications undertaken
subsequently.
Moreover, natural radiometric changes (physiological states and phenological stages of vegetation)
between data of different dates must be taken Into consideration In classification of land cover
modifications, even when corrections related to data acquisition conditions such as sun’s height and
atmospheric transparency variations have been applied previously. In fact, seasonal variations of the
same vegetation canopy may be represented by larger radiance variations than those due to
replacement of a vegetation community by another. Use of scenes acquired on the same day cannot
completely eliminate errors due to Interannual climatic variations (drier or cooler years). Utilisation of
the mean of radiance values obtained from a large number of scenes acquired on different dates
within a year enables elimination of this seasonal effect and the Influence of date of acquisition on
detection of changes, but costs of study increase considerably. We feel that it is preferable to compare
the results of classifications undertaken for each date separately rather than obtain a classification
from bands corresponding to different acquisitions.
For some applications it may be interesting to combine satellite data from various sources for
enhancing the geometric and spectral accuracies (resampling of LANDSAT TM images with SPOT
panchromatic images, combination of ERS-1 and LANDSAT TM or SPOT data). However, the
radiometric values of pixels are modified and the resultant digital values cannot be interpreted easily.

16.2.3 Choice of bands

Choice of bands is possible only if several types of data are available for the zone under study:
— ^to study phenomena concerning the terrestrial surface the visible and infrared (near and reflective
middle) bands can be recommended;
296 Processing of Remote Sensing Data

— ^for subsurface phenomena, spectral bands characterised by penetration (microwave frequencies)

or covering a certain volume (thermal infrared) should be chosen.
When multiple spectral bands are available for a given scene, bands with high dynamic range for
the phenomena to be detected must generally be used. For this, it is necessary to analyse histograms
of values for the training zones identified. It is advisable to retain only those training zones that correlate
least with one another to obtain maximum Information.
Correlation coefficients between bands are not given once for all. They vary according to the date
of acquisition. For example, correlation matrices for a SPOT scene of the western Paris Basin (KJ 39/
250) of 11 March 1995, for a 900 x 1000-pixel segment of the same image and for a scene acquired
on 26 July 1995 over the same zone are shown in Table 16.1. A stronger correlation between band
XS3 on the one hand, and bands XS2 and XS1 on the other, is observed for March when there is
much bare soil and non-chlorophyllian vegetation, than for the image segment and for the scene of
June when relatively more chlorophyllian vegetation Is present.
Lastly, presence of a mask changes the correlation matrix of a scene or a segment relative to that
of unmasked images but the influence of masking is more or less significant depending on the number
of pixels involved (Table 16.2). In fact, for the segment of the preceding SPOT scene masking of pixels
for which the digital value in XS3 is greater than 101, corresponding to chlorophyllian vegetation (3%
pixels in the image segment), modifies the correlation between XS3 and bands XS2 and XS1. On the
other hand, masking of pixels with values in XS3 lower than 14, corresponding to water bodies (0.1%
pixels in the image segment), does not change the correlation matrix.
It Is hence advisable to compute the correlation matrix between bands before undertaking
classification, especially in the case of programs in which the order of indication of bands influences
the result of classification.
Caution: A classification will not necessarily be better when the maximum possible number of
bands Is used. In fact, adding one or two more bands. If they are well correlated with the bands already
taken, has risk of bringing more noise than signal. Classification In this case is likely to be less satisfactory
than with a limited number of uncorrelated or less-correlated bands.

Table 16.1: Correlation matrix between bands XS3, XS2 and XS1 for two SPOT scenes and an
image segment derived from one of them

Scene A (11-03-95) Scene A segment Scene B (26-06-95)

Bands XS3 XS2 XS1 XS3 XS2 XS1 XS3 XS2 XS1
XS3 1 1 1
XS2 0.49 1 0.23 1 0.20 1
XS1 0.69 0.95 1 0.30 0.97 1 0.44 0.94 1

Table 16.2: Influence of a digital mask as a function of size

Scene A image segment Masked image segment Masked Image segment

XS3> 101 X S 3<14

Bands XS3 XS2 XS1 XS3 XS2 XS1 XS3 XS2 XS1
XS3 1 1 1
XS2 0.23 1 0.31 1 0.23 1
XS1 0.30 0.97 1 0.36 0.97 1 0.30 0.97 1
 Quality Assessment 297

For example, addition ofTM7 to bands TM 5,4 ,3 and 2 for classification by the maximum-likelihood
method under the Gaussian assumption of training zones corresponding to 10 grassland units, results
in:
— an increase in number of well-classified pixels for 3 units only,
— an identical number of well-classified pixels for 3 other units,
— less number of well-classified pixels for 4 other units.

16.3 CHOICE OF TYPE OF PROCESSING

16.3.1 Visual or computer interpretation of data
The first choice to be made about processing is between visual or computer interpretation. In many
cases visual interpretation is faster and less expensive, especially when units are identifiable by shape
or textural environment. Contrarily, this type of processing ought to be carried out most rigorously so
as to be most objective and unbiased. This assumes use of standardised description keys (see Chaps.
5 and 18) and, if possible, verification of results by various persons. For example, in the collaborative
program of the European Community on environmental information, the 3-level CORINE Land Cover
mapping was carried out by visual interpretation of IRC satellite images, based on a 44-item
nomenclature for which the content of units Is described precisely (see Chap. 19). This visual
Interpretation was, of course, accompanied by additional field verification, imperative for identification
of some units involving definition of land-use concepts.
Choice of visual interpretation instead of computer Interpretation also depends on the quantum
of data to be processed, the precision expected for the final map and the time and financial resources
available. It is sometimes preferable to be equipped with a good photographic laboratory rather than
information systems necessitating air-conditioning, regular power supply and maintenance, which
may not always be available in all countries.

16.3.2 Supervision of digital classification methods

■ Choice of main methods
When units of homogeneous composition are to be classified, digital methods of classification pixel by
pixel are chosen (see Chaps. 7 ,8 and 9). If the units are complex, it is advisable to use methods based
on analysis of pixels neighbourhood, such as CLARAS, OASIS, PAPRI, VOISIN, etc. (see Chap. 11),
or a plot-wise classification.

■ Selection of useful part of image, level of precision and mask

A classification may not be concerned with the entire image. Preliminarily, the Image segment
corresponding to the zone of study must be selected. If the result of classification has to be superposed
on a map or another classification result, geometric restitution (see Chap. 13) must be carried out after
and not before the classification, since usually restitution modifies the digital values of pixels, leading
to errors in interpretation.
The desired level of precision should also be defined. In a number of cases, for example
classification of regional units, it is not strictly necessary to classify the data with full resolution; a
sampling of every 2 or 3 pixels Is better. Likewise, the number of groups for classification depends on
298 Processing of Remote Sensing Data

the subject of study. It is not necessary to identify groups outside the theme. It is hence essential to
clearly specify before undertaking classification the objectives pursued.
Before commencing a classification on a given theme, it is recommended to mask the known
areas not pertaining to the theme of interest. This avoids subsequent confusions and errors in
classification and permits full exploitation of the discriminative power of the method for the single
theme under consideration. Several masking procedures (see Chap. 7) can be employed according to
the characteristics of the objects to be masked and the remote-sensing data available. If a mask
obtained from exogenous data such as maps has to be applied, the mask must be geometrically
rectified vis-à-vis images and not vice-versa. Otherwise, the digital values of Images may be modified
and ultimately give rise to large errors in the results of classification.

■ Adequacy between statistical characteristics of data and classification method

In digital classifications, attention should be paid to the characteristics of the spectral bands and
population of the data to be processed. In fact, any statistical method applied beyond its domain of
utility leads to errors: Gaussian assumption for a non-Gaussian population, uncorrelated bands for
correlated bands, etc. The data may have statistical parameters corresponding to none of the statistical
hypotheses of the classification methods. It is recommended in such cases to undertake other possible
processing techniques (e.g., generation of new uncorrelated bands) or, in their absence, apply a
processing method whose structure is closest to that of remote-sensing data, keeping in mind that this
may be a source of errors.

■ Selection and verification of training zones

Training zones and nuclei necessary for application of the supervised classification (see Chap. 9)
ought to be chosen in sufficient sizes (at least 30 and preferably more than 100 pixels) for statistically
characterising the units they represent. This is not always feasible in view of relative dimensions of
objects vis-à-vis those of pixels. For this reason. It Is often necessary to use several polygons (or
zones) for characterising a nucleus or a training zone. An attempt is made to at least isolate the most
homogeneous pixels inside the boundary of a nucleus, eliminating boundary pixels.
Subsequently, during analysis of the confusion matrix of a supervised classification, the percentage
(or number) of well-classified pixels in the category which each is considered to represent is determined
for each training zone. It is necessary to repeat the classification by modifying the number and
characteristics of training zones if the latter are not well-classified to 70%; otherwise, during the final
classification there would be many unclassified points and large errors of affectation (see Chap. 17). in
this case, it Is necessary to:
1) verify that no error is made in identification of the zone or its boundaries (training zones ought to
have as far as possible homogeneous digital values or, for structural classification, similar
composition vectors);
2) remove small-size training zones that are not or, are poorly classified (less than 70% pixels in the
predicted class), and repeat the classification.

H Validation of classification
At the end of classification, the results need to be validated. Validation must be done on a set of data
other than that used for classification (see Chaps. 10 and 17). This can be done in two ways (see
Chap. 10, Fig. 10.1):
— Conducting only one field survey for acquiring ground data prior to processing of remote-sensing
data, preserving part of the data for validation. This assumes that objectives of study are clearly
defined and all classes identified at the start.
 Quality Assessment 299

— Carrying out a new field survey after data processing and collecting information for validating
the results of classification. This approach is more expensive than the first but offers the advantage of
verifying new classes identified during processing for which no reference data were acquired previously.

16.4 CONCLUSION
The discussion presented above shows the Importance of recognising the advantages and limitations
before using remote-sensing data. It also shows that it is essential to organise various phases before
and after proper processing. To achieve this pragmatic approach the personnel responsible for the
study must have training as a project leader (Master SILAT).
 17
Quality of Interpretation
17.1 INTRODUCTION
Remote sensing has emerged from an experimental phase to an operational phase for a number of
applications. So it is now Imperative to furnish an estimate of quality of interpretation on infographie
and cartographic documents derived from it. This aspect has been discussed in many articles, especially
in works by Congalton.
The definitions presented below are taken from the project report ‘Accuracy of databases of the
National Geographic Institute’ and a paper by David and Fasquel (1997).
— Precision (p): ‘The degree of agreement between a measured quantity or an estimate and the
expectation (mathematical) of this quantity or estimate’ (David and Fasquel, 1997). It indicates
fluctuations of a series of measurements around their mathematical expectation and is denoted by the
standard deviation of the mean of measurements. It Is represented In Fig. 17.1 by a double arrow,
which gives an estimate of the ‘diameter’ of the group of measurement points. If this arrow is long,
precision is low; If It is short, precision Is high, it can be estimated by the intragroup distance.
— Bias (b) measures the deviation between expectation of a series of measurements and the
nominal value, i.e., between the centre of gravity of the group and the expected value; viz., reference
value or nominal value. In Fig. 17.1 It Is represented by the distance between the centre (x) of the
target (nominal value) and the centre (y) of the group of measured points. Bias is high if the distance
between xand y is large and low if It is small. It can be estimated by the intergroup distance vis-à-vis
the reference value.
— Exactitude (e): ‘The degree of agreement between a measurement or an estimate of a quantity
and its nominal value’ (IGN, 1997). It combines the two concepts of bias and precision as:
p^ + =^

Bias ; distance between target centre (x) and centre (y) of group of points
Precision ; ‘diameter’ of group of points ◄ ------------ ►

Low bias, Low bias, Low bias,

low precision and high precision and high precision and
low exactitude medium exactitude medium exactitude significant exactitude

Fig. 17.1: Schematic diagram showing precision, bias and exactitude; black dots represent measurements and
centre of target (point x) represents the nominal value (after David and Fasquel, 1997).
 Quality Assessment 301

It is often denoted by the mean square error. The nominal value by definition serves as reference
value. To be exact, a value must be precise and unbiased.
Irrespective of data quality (for which the organisation involved In acquisition and dissemination
of data Is responsible) and adequacy of data with the problem to be solved (see Chap. 16) (for which
the person carrying out the study is responsible), an interpretation Is characterised by geometric
accuracy and semantic accuracy.

17.2 GEOMETRIC ACCURACY

The geometric accuracy of an interpretation is expressed by geometric precision defined as the estimate
of fluctuations of differences between the nominal positions (positions in the nominal ground) and the
positions obtained in the data (IGN, 1997). Geometric precision Is divided into two components, one
related to position and the other to shape.

17.2.1 Precision of position

Positioning precision Is resolved into precision of punctuate position and precision of linear position.
The former Is represented by the following indicators:
— mean of errors In x, y and z co-ordinates, which gives an estimate of bias;
— regular grid of regionalised bias circumscribing the data set;
— exactitude estimated by root mean square (RMS) error,
The root mean square error is given for x, y z, xy (planar) and xyz, accompanied by the
corresponding precision estimates. Computational formulae for a sampling of n points of nominal
ground (on a given date) and homologous points In the dataset are as follows:

= (Xi,NG-Xi,Dsf

rt=1

RMSxy = J— ¿ i^i,NG ~ ^i,Dsf + (^/,WG " y¡,Dsf

where X¡ Y¡ /v^and Z¡ /v^are co-ordinates of a sampling point of nominal ground, and X¡;
Y¡; Qg and Z¡; are co-ordinates of their homologues In the dataset.
The corresponding precision estimates are given by:
302 Processing of Remote Sensing Data

R M S^_ RMSy RMS^ RMSx.y

^(R M Sy)=- ^{RMS^) - • ^(RMSy.) '■
V2n ’ V2n ^ ■ ■J2n

— sample size (n) for which each of the preceding computations is made;
— rejection threshold.
As the method of punctuate verification is not applicable to linear objects, precision of linear
position needs to be designated by special Indices:
— degree of agreement between the trace of the dataset and that of reference (after rectification
of boundaries),
— estimate of planar RMS (in x, y).

17.2.2 Precision of shape

Precision of shape refers to the geometric elements (distances, areas, curvatures, volumes, etc.)
constructed from known co-ordinates, for which the precision of planar and/or altitude positioning is
also estimated. A description of estimated parameters, intensity, unit, sampling size and rejection level
should be furnished.

17.2.3 Reliability
Reliability describes the statistical possibility of detection of random or systematic errors resulting
from a geometric correction. We can evaluate the uncertainty by y pixels of the complete boundary
line of a class or of a segment of the line for x pixels in width. It depends on the compositional
heterogeneity of pixels and hence partly on the pixel size and partly on the degree of diversity of land-
use types Inside a pixel. The methods of resolution of pixels into their constituents (see Chap. 15) seek
to reduce this uncertainty.
It is advisable to give for an Image the reference points of ground control, their number, type of
transformation applied, standard deviation, mean square error and, if necessary, the method of
resampling.

17.3 SEMANTIC ACCURACY

17.3.1 Definition
Semantic accuracy refers to the attributes of classes. Four parameters can be distinguished to
characterise errors concerning a theme.

H Nature of classes
Classes can be differentiated as qualitative in nature corresponding, for example, to various land
covers (forests, crops, grasslands, towns, waterbodies, etc.) and of ranked qualitative nature (very dense
forest, dense forest, sparse forest). Errors concern affectation of a point in a classification to a category
other than that to which it belongs on the ground or in a reference map. In the first case {grasslands
instead of forest), the nature of confusion or error differs from that of the second {very dense forest\n
place of dense forest), since it concerns very different objects. It will be seen later that magnitude of
affectation errors depends not only on the nature of classes but also on the ultimate use of classification.
 Quality Assessment 303

Frequency of affectation errors

This parameter refers to the number of affectation errors in a class or group of classification. How to
estimate this frequency from a error matrix will be described later.

■ Importance of affectation errors

Some affectation errors are generated from a greater mistake and may result in more serious
consequences than others. For example, mistake between a coniferous forest and a water surface
(due to low radiance levels in the near Infrared band in both cases) will be greater than that between
a pure coniferous forest and a mixed deciduous and coniferous forest The latter example illustrates
the problem of defining class themes (establishment of typology), presented in the next section (17.3.2).

■ Source of errors
Affectation errors may be related to characteristics of data (undiscriminating spectra! bands for the
theme studied, pixel dimensions Inapt for the class themes, etc.), independent of the Intrinsic quality
of data, to acquisition conditions or to classification algorithms used (see Chaps. 7, 8 and 9).

■ Precision
Semantic precision is differentiated into user’s precision and precision of attributes.
Semantic precision is ‘the degree of conformity of values of units of data with those of their
homologues in nominal ground’ (National Committee on Geographic Information, 1993). It represents
the degree of agreement between a class assigned to a pixel in the classification and its true class
known from ground or other reference sources. For example, classifying a permanent pasture In the
theme grass/anc/s will be less accurate than its classification in permanent grasslands but more accurate
than classification as agricultural land.
User’s precision corresponds to estimation of accuracy of classification for a given usage. In the
preceding example, in an inventory of grazing areas, user’s precision will be low when It is classified
as an agricultural land.
Precision of attributes indicates the degree of agreement between various class themes derived
from a series of identifications of a single entity. This is designated as percentage affectation error in
the error matrix.

17.3.2 Establishment of typology

This phase should commence even before undertaking classification. In fact, classes to be used
should respond to the objective identified or the terms of contract of study. The main difficulty arises
when the themes to be mapped do not correspond to identifiable physical or spectral characteristics.
For example, unique spectral discrimination between permanent meadow and herbaceous and mowed
tracks of an aerodrome Is not possible, since objects here are to be differentiated not in terms of
constituents but their use.
For pixel-by-pixel classification, any typology must be exhaustive and the classes mutually exclusive:
a point should be classified in one and only one class. This poses a problem for boundary pixels which
may either be left unclassified or regrouped in classes In which they are incorrectly classified. This
condition does not apply to classifications based on fuzzy logic for which the points can pertain to two
or more classes and in which second-choice affectations can be used for delineating map boundaries.
A hierarchic typology between classes is desirable because it enables regrouping of classes
identified on the ground but corresponding to a higher precision level than that required for the study.
304 Processing of Remote Sensing Data

Thus classifications corresponding to various precision levels can be obtained and associations between
classes established.
It must be ensured that the precision required for a study corresponds to the precision level of the
remote-sensing data used or the precision achievable on the ground using existing observation methods
and/or measurements.

17.3.3 Establishment of error matrix

An error matrix is established between the data produced from classification and the reference data
(independent of those used for preparation of the document to be tested). A number of questions arise
immediately.

■ Collection of reference data

Reference data can be acquired either from ground truth measurements or aerial photographs or
other remote-sensing data different from those used for classification; even maps obtained by other
methods may be used for this purpose. Collection of reference data may be easy as in Identification of
various land-cover types or complex as In measurement of soil surface roughness, chlorophyllian
biomass, foliar indices, tree heights, density of plant covers, etc. This assumes in ail cases that the
reference data are precise. However, It Is not always easy on the ground to measure classes of density
of tall trees in forests.
Moreover, It is sometimes difficult to locate reference data for estimating accuracy. An example of
zoning of landscape units into various organisational levels in Normandy can be cited. No map document
corresponding to this problem existed earlier. Neither ground data nor aerial photographs were useful
for investigating landscape units. No significant statistical data existed in the form of maps. Only semantic
accuracy but not cartographic could be estimated using studies of rural geography on the one hand
and rural economy on the other.

H Number and size of samples

While thousands of reference points can be readily obtained in satellite data. It Is very time consuming
and expensive to verify these points on the ground. The number of points to be sampled depends on
the sampling procedure, desired precision of estimation, number of pixels present in a category and
estimated precision of classification. Hence the smallest number of ground verification points that
enable statistical analysis should be chosen. In other words, an average number that is statistically
correct and practically feasible should be selected.
Points that are geographically representative of the data analysed are chosen as reference. A
random stratified sampling can be considered but a random choice of reference pixels leads to high
costs for acquisition of information on the verification pixels (field trips, inaccessibility, etc.). Moreover,
precise location of points of verification vis-à-vis image pixels is often difficult.
To obtain a confusion matrix, a sufficient number of points should be used in each category,
totally ignoring how the confusion matrix will be distributed among various classes. This necessitates
a number much higher than that indicated by equations based on binomial distribution or normal
approximation of binomial distribution. These methods are not appropriate for estimating the number
of control points since they compute the number of samples necessary for assessment of global
precision or, at best, assessment of global precision for only one class. The number of control points
is hence to be determined based on experience and practice. Congalton (1991) recommended 50
points. Our personal experience suggests that at least30 correctly classified control points per category
(along the diagonal) are necessary, I.e., taking poorly classified points into consideration, in all about
40 points per class. These numbers, of course, correspond to independent sampies. If such is not the
 Quality Assessment 305

case, the number of control points must be increased. More than 250 reference pixels are necessary
for estimating the mean precision of a class at about 5%. If the number of reference samples is less,
the confidence Interval of the precision obtained increases considerably. This means the number of
points is sometimes limited simply because the unit is least represented in the zone of study and it is
not possible to find a sufficient number of points.
It is not compulsory to have the same number of control points in each category, provided a
minimum number of samples is available for each. It is recommended that a larger number of samples
be taken for important classes in the study and fewer for those of secondary interest. A small number
of samples can be taken for classes exhibiting a low variability or lower risk of confusion, as shown by
analysis of two- or three-dimensional histograms (see Chaps. 4, 7, 8 and 9).
On the other hand, the size of control sites ought to be compatible with the resolution (pixel) of
remote-sensing data, which varies from a few decimetres or metres to several kilometres (see Chap.
2). This Is in contrast to existing methods of measurement or assessment commonly used in field,
often point-by-point measurements.These need to be modified (or redesigned) according to the remote­
sensing data used. For example, an areal sampling (distributed randomly, systematically or along
strips) Is preferable to a point-by-poInt sampling, and measurements in extended areas rather than at
points.
While CORINE Land Cover type (see Chap. 19) categories can be readily described, measurements
of Leaf Area Index over areas comparable to those of a SPOT pixel are very difficult. In the latter case,
other methods such as optical densitometer measurements should be employed! Among the solutions
adopted are analysis of optical density of layers sensitive to red or near-infrared in aerial photographs
acquired at low altitudes or that of a traverse with a SPOT Cimel simulation radiometer using a ULM.
In many cases, small size ground sampling must be compensated by a large number of repetitions.
For example, the Cimel radiometer of SPOT simulation measures, at a height of 1.5 m, the reflectance
of a circular area of about 7 dm^. Spectral characterisation of a plot of more than a hectare requires
measurement of at least 100 different points taken randomly.
As a general rule, it is better to make an assessment of classification precision with a sufficient
number of reference pixels. In the absence of sufficient points, a less precise assessment is better
than no assessment at all!

H Sampling mode
The mode of sampling is an important part of selection of control points, since the confusion matrix
must be representative of the entire image. The choice is between random and stratified sampling.
Stratified sampling consists of preliminary zoning of the region according to a theme (geomorphological
units, separation of marine region from terrestrial, etc.), followed by random or systematic sampling
inside these zones. Both have advantages and disadvantages.
— Random sampling is amenable to use of conventional statistical methods. However, It is not
always easy to employ since it may lead to choice of inaccessible points in montane regions or zones
with little communication network. Moreover, there may be risk of straddling points in two units or of a
preferential sampling in large size units of low thematic interest and an underestimation of small size
units of high thematic importance. These units can be taken into consideration only by significantly
increasing the number of points, which involves more time and additional expenditure. Lastly, this type
of sampling can be conducted only after completion of classification. In other words, often too late for
locating control objects in the field (harvested crops, various drainage conditions, etc.).
— Stratified sampling has the advantage of distributing samples in various classes of thematic
interest, it offers the same disadvantages as the preceding method if points inside strata are sampled
randomly. This disadvantage can be avoided by systematic sampling. Contrarily, there could be risk of
samples being no longer independent or of introducing bias In some stratified tests (see below, use of
Kappa statistic in analysing a error matrix).
306 Processing of Remote Sensing Data

The most efficient solution consists of combining random, systematic and stratified samplings.
However, caution is needed in the statistical methods employed since they may be beyond the region
of validity. Lastly, common methods of assessment are based on comparison of pixels with control
points, and not image zones with zones on the ground.

17.3.4 Discussion of error matrix

■ Elements of error matrix
An error matrix comprises the following components:
— in columns, data derived from classification,
— in rows, data of reference classes (ground truth or other data).
Overall accuracy denotes the number of correctly classified individuals (or pixels) (correlation
between groups and classes) divided by the total number of individuals (or pixels).
Error of exclusion, omission errors computed along columns, corresponds to the distribution of
pixels of a group, derived from interpretation of remote-sensing data, in the various classes of reference
data.
User’s accuracy \s the percentage of individuals (or pixels) in a group derived from classification,
correctly classified vis-à-vis the reference data. It is equal (in %) to ‘100% - error of omission’.
Error of inclusion, commission errors computed along rows, is the distribution of a reference class
among various groups derived from interpretation of remote-sensing data. It differs from error of omission
of error matrices established for vector data (David and Fasquel, 1997). In fact, in the latter the omission
error denotes absence of data, whereas for an error matrix of a thematic classification, the commission
error concerns existent data whose identification is erroneous.
Producer’s accuracy corresponds to the percentage of individuals (or pixels) of a reference class
correctly classified by remote sensing; it is read along the rows. It is equal (in %) to ‘100% - error of
commission’.
An error matrix controlling a classification should always be given for each class, along with risks
of confusion with any other class.

H Precision estimator
The precision estimator, known as the Kappa coefficient, varies from 0 to 1; it accounts for errors In
rows and columns (Congalton, 1999). It enables global assessment as v»/el! as at each class level and
accounts for two types of errors, viz., omission (deficits) and commission (excesses) (Table 17.1).

X ii- t
Kappa -
iX M
hi

where ris the number of rows and A/the total number of observations.
it should be remembered that this formula applies to random sampling and comparison of pixels
but not of polygons of control points, i.e., only to independent samples in classification. In the case of
a systematic or non-aligned systematic sampling, the resultant bias for Kappa is small (it may hence
be used) whereas contrarily the variance estimator is biased. In the case of stratified sampling, another
estimator needs to be defined. Stehman (1996) proposed a KS estimator for stratified sampling, provided
the stratification pertains to regions covering several classes and not to classes.
 Quality Assessment 307
Table 17.1: Notation of an error matrix for computation of Kappa statistic

Classification
Total lines
U niti Unit / Unit n

Reference U n iti ^1/ A/i

Unit / X-n w,-
Unit n Xnl Xnn Nn
Total columns M, Mn N

The Kappa estimator expresses the proportional reduction In error obtained by a classification
compared to error obtained by a completely random classification. A value of 0.75 indicates that the
classification employed eliminates 75% of the errors obtained by a procedure working completely at
random.
The Kappa estimator applies to evaluation of accuracy for classifications pertaining to cardinal
units. For ordinal units, it is preferable to use other estimators such as X^-test or F-test, which measure
the degree of agreement between the true distribution and theoretical distribution of classes.

■ Example of error matrix

This example is taken from a Ph.D. thesis on Lower Normandy region. The classification by maximum-
likelihood method In a LANDSAT TM image pertains to mapping of various physiognomic units of
permanent grasslands identified on the ground from the height and chlorophylllan state of the canopy.
Reference points were chosen after delineation of marsh zones. It thus represents random sampling
after stratification.
Figures given in Table 17.2 correspond to the number of pixels, unless indicated otherwise.
Overall accuracy: 14,021 /16,353 = 0.857
The Kappa coefficient is computed here as follows:
N = 16,353
= 267,420,609

Table 17.2: Example of error matrix

Classification

Physiognomy Tall Tall Heteroge­ Short TOTAL Producer’s Error of

chlorophy- slightly neous chlorophy- precision omission
llian chlorophy- llian (%) (deficits)
llian (%)

Tall chlorophylllan 2 109 3 25 4 2 141 98.5 1.5

X Tall slightly 4 912 376 5296 92.7 7.3
h- 8 0
D chlorophylllan
CC
h- 41.1
Q Heterogeneous 874 955 2 709 61 4 317 58.9
Z Short chlorophylllan 5 4 291 4 599 99.4
D 1 0 11 0 .6

0
CC TOTAL 2 996 5 880 3121 4 356 16 353
0
User’s precision (%) 70.4 83.5 8 6 .8 98.5
Error of commission 29.6 16.5 13.2 1.5
(excesses) (%)
308 Processing of Remote Sensing Data

Xjj = 2109 + 4912 + 2709 + 4291

= 14,021

É/=1
= {1 X (3 + 25 + 4)} + {8 X (376 + 0)} + {(874 + 955) x 61}+ {(5 + 10 + 11) x 1}

= 114,635

Kappa is equal to:

K = [(16,353 X 14,021) - 114,635] / [267,420,609 - 114,635]
= [212,932,413-114,635] / [267,420,609 - 114,635]
= 229,170,778 / 267,305,974
= 0.857

The Kappa coefficient and overall accuracy are equal.

A good accord is generally seen between the map obtained from classification of satellite data
and ground truth observations. However, some classes contain more errors than do others; this is
particularly so for group ‘heterogeneous’, which has accumulation of errors of commission (13.2) and
errors of omission (41.1)^ .The errors of commission consist of pixels corresponding on the ground
mainly to the class ‘tall and slightly chlorophyllian’ and affected to the class ‘heterogeneous’. On the
other hand, the errors of omission consist mainly of pixels corresponding on the ground to the class
‘heterogeneous’ and affected by classification to the class ‘tall and slighty chlorophyllian’. These
confusions and the low producer’s precision for the class ‘heterogeneous’ might necessitate revision
of definition and modification of test-zones of reference or elimination of some. In fact, a better adjustment
of digital number values in thematic classes may produce better precision.
A confusion matrix controlling a classification should always be given for each class, along with
risks of confusion with any other class.
Contingency tables (error matrix) for a complete real case study will be given In an example of
CORINE Land Cover (see Chap. 19).

17.3.5 Limitations of conventional methods of accuracy

assessment
The conventional methods of classification accuracy assessment have the following defects:
— It is assumed that each point (of classification, such as a control point) can be unambiguously
assigned to only one theme which is subsequently compared with the theme obtained from another
method. However, it is not always easy to affect a pixel to a given class or to associate a point on the
ground to a specific class.
— Information regarding the significance of errors is limited to distribution of incorrect attribute of
a class but does not indicate their degree of severity.
— Users of maps derived from classifications require more detailed information on thematic aspects
than that provided by conventional methods of error evaluation. For example, they need a detailed
analysis of confusion, class by class, or estimates of severity of errors committed for a specific
application.
That is why other methods of classification quality assessment are presently being developed.
These developments concern the nature, frequency, origin and significance of semantic errors and

‘ The different values of these two errors, not symmetric, should be noted.
 Quality Assessment 309

make use of fuzzy logic and simulation studies. An example of qualitative evaluation of semantic
accuracy for a classification is given in Chap. 19 concerning CORINE Land Cover.

17.4 CONCLUSION
To carry out a study corresponding to operational applications of remote sensing, it is imperative to
know the quality of the data to be used and to provide a quality assessment of the product data
(thematic interpretations, maps, etc.).This necessitates a preliminary study of the problem posed and
a feasibility analysis of how remote-sensing data can solve it.
Ready-made solutions should not be expected from remote sensing.
No method can provide a complete analysis of a problem and only a combination of several
approaches will be most constructive. Hence, remote sensing is not a means for eliminating ground
studies but for making them more effective and validating them for spatial extension. An analysis of
quality/cost ratios, computed exhaustively, enables choice of the most appropriate methods of study.
Establishment of an organisation such as the Earth Observation Centre (EOC) situated in the research
centre of the European Community (Ispra, Italy), which has among others improvement of seller/
buyer relations and rendering advice to users as a mission, ought to facilitate this choice. In fact, EOC
does not supply remote-sensing data but provides information on where to procure them. This free
service concerns not only remote-sensing data but also any ground information necessary for
Interpretation, validation or calibration of spatial data. Internet communication with EWSE {European
Wide Service Exchange) enables location of research organisations, societies or companies capable
of answering your needs. EWSE is a temporary arrangement that will be replaced by an operational
system, INFEO {Information on Earth Observation). The EOC ought to become in the long term an
information tool for persons and organisations Interested in earth observation and some of its objectives
should become operational in the near future.
It must be remembered that unsuitable remote-sensing data, inappropriate classification methods
and inadequate ground (or reference) information cannot give good results.

References
Cnig. 1993. Qualité des données géographiques échangées. Groupe de travail “qualité des données géographiques
échangées”, September, 22 pp.
Congalton RG, Green K. 1999. Assessing the accuracy of remotely sensed data: principles and practices. Lewis
Publ., Boca Raton— London— New York— Washington DC, 137 pp.
David B, Fasquel R 1997. Qualité d’une base de données géographiques: concepts et terminologie, Bull. d’Info, de
l’IG N,67:1-51.
Gopal S, Woodcock C. 1994. Theory and methods for accuracy assessment of thematic maps using fuzzy sets.
Photogrammetric Engineering and Remote Sensing. 60:181-188.
Stehman SV. 1996. Estimating the Kappa coefficient and its variance under stratified random sampling.
Photogrammetric Engineering and Remote Sensing, 62:401-407.
 E
APPLICATIONS
 18
Agrolandscapes
It is possible to interpret and locate major geographic units in satellite images with a synoptic view.
Visual interpretation of these images is similar to structural processing, few developed and not common
in remote-sensing softwares. The role of interpreter thus assumes great importance since he can
integrate and define units whose content varies in given proportions and with a given spatial organisation,
viz., patterns.
The methods of integrated analysis of satellite images and small-scale photographs acquired
from aeroplanes, stratospheric balloons or spaceships are based on the concept of ‘landscape’ adapted
to these interpretations.

18.1 LANDSCAPE INTERPRETATION

18.1.1 Concept of landscape
Several definitions exist for the term ‘landscape’ covering various realities and related to the sensitivities
and modes of perception from space (Rougerie and Beroutchachvili, 1991; Lizet and de Ravignan,
1987). Landscape investigators, and especially those of the School of Landscape (Donadieu, 1993),
emphasise on the importance of perspective view which organises the landscape elements according
to various planes in the space seen by the observer. However, It Is possible to analyse landscapes in
a vertical view. In this case, the notion of flight point and relative position of landscape elements is lost:
trees do not conceal the landscape but the lower strata and soil; topographic variations no longer limit
the view.
In this book, we have followed a rural approach with the aim of management. In fact, need was felt
for developing spatial models for division of a region into various units to which information aiding their
management may be attached. At present, many divisions of the territory exist: small agricultural
zones, forest regions, climatic zones, administrative boundaries, etc. It Is not clearly known how these
boundaries were drawn. Consequently, It is now difficult to know whether thirty years later and after
many variations, they are still valid. Chorological laws prevailed in their constitution are not known and
hence these various segmentations of the territory cannot be spatially analysed.

18.1.2 Landscape: synthetic descriptor in remote sensing

The concept of landscape (Henin, 1994) generally refers to ‘an extent of land presented by nature to
an observer’, independent of Its meaning in painting. In the definitions of the Larousse and Petit
Robert dictionaries, emphasis is on the terms: nature, point of observation and extent of terrain.
With regard to the concept of landscape, landscape specialists emphasise aspects of sensitive
perception, giving priority to perspective view. However, it is difficult to define a landscape object in the
former case.
314 Processing of Remote Sensing Data

Deffontaines (1985) defined landscape as follows; ‘portion of region viewed by an observer wherein
a combination of facts and interactions are inscribed, which is perceived at a given moment only as
the global result’.
We ourselves consider landscape an 'extent of land observed in the vertical direction, in which
the main subject is the natural environment. Landscape analysis is objective, but perspective view is
absent since the observer is situated above the object observed.
Landscape, as an integral descriptor of environment, is 'the entirety of components of the
environment whose spatial organisation is studied (in a geographic sense): type of units, dissemination
of these units over the region studied, spatial distribution, neighbourhood associations and hierarchy
between them’.
This spatial structuring is defined according to a theme such as agronomy, hydrology, pedology
(see Chap. 23), geology or viticulture, etc. The theme enables choice of the most useful landscape
variables for determining spatial units corresponding to them.
Thus, hydrolandscapes are defined as:
'A combination of landscape elements: natural vegetation, timber zones, agricultural plots situated
in a geomorphological unit associated with natural water flow (elementary watersheds, drainage
patterns, etc.), controlled by topography (slope and exposure) and surface states of soil, whosespatial
organisation permits definition of an entire (or part of) a watershed ora drainage system, i.e., hydrological
and spatial structure ’.
Subgroups can be defined as hydrolandscape units and hydrolandscape elements.

18.1.3 Agrolandscape
Rural landscape is the resultant of a combination of factors of environment and human actions with
regard to it. In a given spatial field, environmental factors observed at a given resolution are organised
spatially and their distribution constitutes a source of information. For a given object, the main
environmental components can be defined which, once modelled, give landscape units.The latter can
also be analysed by seasonal or decennial diachronic iterations.
The concept of agrolandscape is based on the assumption that a spatial, but not random,
organisation of environmental factors exists. Spatial organisation of a rural environment is related to
its utilisation by man, and the human actions are imprinted In the geographic space. On the other
hand, the environmental factors constitute the base of spatial organisation. Consequently, human
actions, adapted over the course of time to environmental factors, can be identified at least partially in
the agrolandscape by means of spatial analysis.
In studies concerned mainly with rural environment, agrolandscapes are defined as 'a combination
of plots (agricultural or forest) and landscape elements, viz., natural vegetation, timber zones, topography
(slope and exposure), drainage system, surface states of soil, whose spatial organisation enables
definition of (or part of) a land: social and spatial structure’.
Subgroups can be defined as agrolandscape units and agrolandscape elements. In each case
we deal with patterns: ‘spatially organised groups of various landscape elements’. Landscape units
are units characterised by specific patterns (described by a composition vector, see Chap. 11).These
are not homogeneous, but heterogeneous zones that can be described only by a combination of
criteria of multiple origin based on environmental factors.

18.2 LANDSCAPE ANALYSIS

Interpretation of satellite images for preparation of landscape maps must be done keeping In view the
purpose for which it will be used. In fact, such a map should be readily readable by anyone who
 Applications 315

consults it. So it is desirable to represent components of map in such a way that reader can readily
identify major units which he (she) already knows or can locate on the ground.

18.2.1 Different points of vision

The point of vision of an observer (pedestrian, for example) is very different from that of a satellite (Fig.
18.1) for three reasons:
— difference in angle of observation,
— difference in field of view and hence in reference zones,
— difference in resolution.

Field

Fig. 18.1: Landscape seen by an observer and a satellite.

®Angle of observation
An observer views a landscape with a highly Inclined angle. On a plane, he (she) sees the landscape
around him (her) over a distance of 1.5 to 15 km with a height of 1.5 m or at a very large angle a, of
about 90° (arc cos 0.0Q1 = 89.9°) when he (she) observes the horizon. On a mountain, the angle may
decrease (arc tan 0.3 = 72°) since, for a given distance, the height may reach 500 m or more due to
topographic variations. However, in general the angle of view is large. Contrarily, vision from an aeroplane
or satellite is close to vertical. Rarely does an obstacle obstruct the field of view in this case, whereas
a screen of trees or more or less high relief suffices to conceal part of landscape from the observer on
the ground. Consequently, the point of vision differs between an observer and a satellite. On the
ground, friends are recognised from their appearance; from a satellite, they can be recognised by
‘implantation of their hair on the skull’!

H Field of view
It is estimated that in ideal cases an observer can see up to about 20 km. This is due to opacity of the
atmosphere which absorbs solar radiation. Nevertheless, there may be special situations when the
316 Processing of Remote Sensing Data

sky is particularly clear or when contrast between objects observed is very high. Such is the case, for
example, of snow-covered mountains exhibiting specular reflectance significantly differing from that of
mountains with no snow.
The field of view in satellite Images generally covers several tens or even thousands of km^.
Reference areas in such a case are more numerous and the landscape structure is more readily seen
than from an observation point on the ground.

H Resolution
Resolution varies according to the distance of an object from an observer. While the resolution at the
foot of the observer is a few centimetres, it increases when the object is removed to a distance. An
object appearing as square at the foot becomes a trapezium when it is far (Fig. 18.1). Similar effect is
also observed when wide-angle photographs are taken with a camera (see Chap. 14, Fig. 14.3 a).
In the case of high-resolution satellite images (such as SPOT HRV or LANDSATTM) the resolution
is almost the same at any point of the field. A square In the image is slightly distorted into a lozenge in
a map, due to the rotation of the Earth and the type of projection used (see Chaps. 2 and 13); however,
variations are very small. The same is not true with low-resolution satellites and microwave systems
(see Chap. 2).

18.2.2 Principal components of landscape

In satellite images, at the resolution of a pixel, various objects are acquired with their true dimensions.
This is not the case with common maps in which objects we wish to emphasise are distinctly enlarged
and, hence, are not In the same scale as other features on the map. Thus, on road maps, roads
usually have a width of 1 mm which may correspond to 200 m, for example. Considerable distortions
also occur in ground photographs due to the angle of view (see Chap. 14).
Hence, roads are not clearly detected in satellite images. But often a highway can be discerned
not only because it is wide enough but also because superimposed to other components of landscape.
The main landscape component in satellite images Is the land-cover unit and its use by man (see
Chap. 17). Strong correlation often exists between land cover and major geomorphological units
(mountains, valleys, etc.) which constitute the principal landscape components perceived by an observer
on the ground. Hence, satellite images have to be Interpreted to depict the following features on the
final maps:
— forests,
— valleys and drainage patterns,
— relief zones,
— lots,
— bare soils,
— towns.

* Forests
In temperate zones, forests are quite readily recognised in a satellite image when they are dense
enough and the date of acquisition Is carefully taken into consideration. Appearance of forests when
deciduous species lose their leaves differs significantly from that during foliation period. Nevertheless,
they are readily recognised although their spectral characteristics differ radically.
What can be defined as forest and what appears clearly in images is a group of trees or bushes
whose organisation in layers and in density contributes to 100% coverage and reveals certain
heterogeneity due mainly to shadows of tall tree canopies on adjacent ones.
 Applications 317

It is also necessary to differentiate between forest, grove, tree alignments or riverine forests
zones. This is Inferred from dimension and form of blocks and their location.
In landscape interpretation, a minimum size is used to demarcate forest blocks. For the European
CORINE Land Cover mapping, the minimum size used was 25 ha and 100 m In width for linear
objects. In most precise cases, it is considered that at least 25 pixels are needed for characterising an
object. This corresponds to a block of 1 ha with a pixel of 20-m resolution or an alignment of trees of
the order of 160 m (8 pixels In length and 3 in width).
In satellite images, one can particularly differentiate organisation of forest blocks with respect to
one another, their respective dimensions and distance to be covered to go from one block to another,
as well as obstacles encountered in the path. Hence, possibilities of movement of various animal
species can be readily determined; this leads to ecological applications of landscape analysis.

■ Valleys and drainage systems

Major rivers or streams can be detected by their reflectance if they are wide. The composition of pixels
in these zones is characteristic of water. Further, a particular geometric disposition of these bodies is
needed for the square grid imposed by the satellite to intercept the watercourses in such a way that
connected pixels are observed, enabling delineation of the latter.
If the river Is very narrow (less than 10, 20 or 30 m), for each pixel there will be a combination of
spectral characteristics of water and surrounding objects such as roads, wooded strips, field, earthen
banks, etc. Water will not be directly detected, since only boundary mixels occur along the rivers. This
Is very common for a number of watercourses. While this presents major difficulty In computer
processing, streams can be more readily traced by visual interpretation since the human brain and
eye can Interpolate between the points where there is no water and the points of mixels. The interpreter
has the capacity to take a decision of affecting a mixel to the theme ‘watercourse’ through study of its
neighbourhood and according to pre-existing and memorised models. Thus, tracing meanders is very
easy by visual interpretation but much more difficult by software.
Tracing valleys is a still more complex operation but can be readily achieved by visual interpretation
if care is taken to clearly define the ‘valley’ landscape. Delineation is based on the fact that the width of
the zone is much smaller than the length. Moreover, the relief in such zones is low and, when it exists,
Its traces can be observed In the land cover. Thus valleys often comprise grasslands, sand pits, riverine
forests, poplar plantations, water zones associated with water table, etc. These neighbouring objects
and their distribution characterise a valley. Roads are most often outside of water, similar to railway
tracks, or these communication lines are bordered by excavations and fillings, which can be Interpreted
from their shadows or from large brightness of mineral products of filling. Agricultural lots also often
show some relationship with valleys: in many major valleys, plots are perpendicular to the watercourse
so that all plots are assured access to the river (irrigation, drinking water for animals, pisciculture,
etc.). In major river-beds, plot boundaries of grasslands take the form of meanders and are circular
(see Chap. 5, Fig. 5.4). All these features enable ready visual demarcation of valleys.
When there are terraces, it is not always easy to identify them since the relief is flat and not seen
on Images. Nevertheless, they are differentiated because edge of a terrace runs across a wooded
border or because plots of various terraces are not arranged identically. Often, the dimensions of plots
are not equal and their size may be related to the width of the terrace or differences lie In diverse
orientations of plots.
Interpretation of smaller valleys is more difficult. As in an aerial photograph, adequate detail is not
available in stereoscopic views for identify whether the bottom of a valley is flat or curved or whether
alluvial fan exists or not. Asymmetric valleys cannot generally be delineated, except through size of
plots. Lots are commonly used to interpret drainage patterns. However, shadows of slopes can also be
useful. To use this effect, images acquired at low solar angle must be selected, I.e., winter Images in
Europe, and Images taken with a large Incident angle for SPOT.
318 Processing of Remote Sensing Data

Talwegs are readily interpreted under forest cover in satellite images, if the image is taken in
winter, or over bare soils. In fact, soil cover is thicker in the bottom of talwegs due to accumulation of
colluvium. Consequently, water reserve Is larger than in slopes and soil surface is more humid or more
organic, all these contributing to a darker image of the base of talweg. Talwegs with a density at least
equal to that observed on a 1:100,000 or 1:50,000 topographic map can be easily delineated. Sometimes
talwegs are detected based on differences in vegetation within a plot during anthesis, from Its level of
growth or ripening. In fact, in these phenological stages plants need to rapidly mobilise various nutrients
which they can more likely obtain in colluvium at the base of talweg than on slopes. Depending on the
case, vegetation flowers earlier or develops leaves rapidly and hence plant cover on soil Increases; or
it may die early or become dry. These Important modifications in vegetation tissues are readily identified
in near- and middle-infrared bands. Quite often talwegs comprise, at the base, plots in the form of
strips or at least rectangles, whereas slopes are occupied by forest. Contrast occurs between forest
and crops and enables demarcation of talweg boundaries.
These constitute indirect interpretations derived from well-known agronomic models, based on
spectral and chorological laws.

■ Relief zones
Relief can be primarily analysed by stereoscopy. In fact, hypsometric maps in countries in which they
did not hitherto exist are now prepared with the help of satellite images. Precision is of the order of 5
to 10 m in altitude and area. It can be Improved with civil satellites which have a resolution of 5 m for
SPOT and about 1 m for others. Precision of aerial photography seems likely to be attained shortly in
satellite images as well.
Relief can also be determined from shadows, especially In winter images. However, it should be
noted that shadows in images are northwards whereas in topographic maps they are south-east. In
fact, we have the impression of perspective view when the sun and hence the source of Illumination is
at height on the left (north-west). Consequently, in satellite images, the first view Is often interpreted by
our brain with an inverse relief. It Is hence necessary to accustom oneself to see relief in correct
perspective. In winter, shadows and zones exposed to the south are better illuminated and so relief is
perceived fairly correctly. However, as the sun is lower than In summer, it is common to find a slope
entirely in shadow. Contrarily, these disadvantages are eliminated in summer Images to a great extent.
Relief is also reinforced by the diversity of vegetation developing at various altitudes or on slopes
of different magnitudes. In fact, ecobioclimatic conditions vary according to the altitude: vegetation
does not develop in the same way and plant communities are not identical. On slopes and depending
on the degree of exposure, vegetation does not receive the same quantity of light and hence plant
communities also differ from each other. Thus it is possible to make correct estimation of altitudes and
slopes by studying vegetation.
Relief can also be Inferred from land-cover maps. In fact, relationships exist between altitudes
and spatial distribution of plant communities: mineral zones comprising rocks, boulders and pebbles
and rupestrian species, mountain pastures, coniferous forests and deciduous forests. According to
the depth of the phreatic water table, it is possible to differentiate various types of terraces and silviculture
zones. Slopes are also detected by the shapes of plots, which in some landscapes approximately
follow the elevation contours or, contrarily, are perpendicular to them. Lastly, It is quite common to
observe forest belts on steep slopes and on top of slopes just In front of plateaus and such zones are
easy to identify. They give an indication of plateau boundaries and steep slopes.

HAgricultural lots
In satellite Images, boundaries of lots can be differentiated if sufficient contrast exist between two
land-cover types. For example, two plots of cereals cannot be distinguished if they do not differ in
some characteristics such as phenological stage. Such may be the case since they are not of the
 Applications 319

same species or the same cultivar or not sown at the same time or did not develop at the same
rate due to differences in soil and hence did not have the same appearance on the day of image
acquisition.
Thus, in images as in aerial photos, only cultivated plots can be detected but not basic plots.
Breaks can be identified. Contrarily, variations in lots over years can be identified using
diachronic images. Forest divisions can also be detected in some cases with images taken on
satisfactory dates.
Lots are clearly differentiated In satellite Images If plot sizes are greater than 2 mm on the Image
Interpreted.Thus, in a 1:100,000 SPOT image, plot sizes ought to be more than 200 m by 200 m, I.e.,
4 ha. In this case, plots less than 4 ha are qualified as small. However, using images with a resolution
of 5 m or less It becomes possible to Interpret vineyard plots, often very small In size.
It is useful to select classes for plots according to existing plots In the field under study. Thus these
classes may be chosen such that the intermediate size is the most frequent (mode). Bertrand (1994)
used the following values for plot sizes for analysing the entire region of Yonne district (CD 18.1):
— small plots: less than 4 ha,
— medium plots: 4 to 60 ha,
— large plots: more than 60 ha.
Likewise, most visible forms only are used for interpretation. Strip-form, rectangular and polygonal
shapes were chosen for the study of Yonne.
Plots are not necessarily correctly identified on the ground due to the inclined view of the observer.
Thus it is almost impossible to identify In the field circular plots or other dispositions of historic plots
extending over several kilometres.

■ Towns
Towns are useful in locating the landscape relative to various maps since they are almost invariably
indicated in the latter.
They are represented In images by house roofs, often dark, and shadows produced by tall buildings.
Roads are characterised by variable luminance values. Rivers or channels and ‘green belts’ comprising
more or less chlorophyllian vegetation according to season are often observed in towns. Almost all
types of objects are observed in urban zones: hence they cannot be Interpreted using a reflectance
model. An urban agglomeration most often comprises all themes used in classification.
However, in infrared colour Images, urban zones appear in blue to cyan shades. Various urban
zones can be distinguished by the degree of heterogeneity of shades. Often the town centre Is Indicated
by dark shades since roads are narrower and hence shadows larger. Large commercial or industrial
areas are usually very bright. Lastly, parks and gardens are readily identified as red, provided they are
sufficiently large In size.
A town Is a common feature in which mixed mixels (see Chap. 15) occur in large number since
most of the pixels in it are represented by various objects. Towns can be better identified using structural
analysis, easy for the eye. However, if magnification is not adequate, roads cannot be readily detected
with naked eye unless they are large avenues. Contrarily, in some cases, the last axis of principal
component analysis or mathematical filtering enables detection of boundaries of units and hence
identification of various communication lines (see Chap. 12 and CD 12.3 and 7.13).
Villages are easy to identify in a contact printing of satellite Image. They are differentiated by their
heterogeneity, by hues close to those of bare soils and by the fact that communication lines often
converge towards them.
Hence a satellite image can be taken as a base map, especially in regions for which no maps are
present and when there are no resources to procure information layers of the cartographic databank
of IGN.
320 Processing of Remote Sensing Data

18.2.3 D escriptive form at and statistics

Once the principal components are set up, specific aspects of study (such as agricultural, pedological,
hydrological, ecological, etc.) need to be highlighted and a legend prepared for each objective. A
complete legend comprising all variables, along with their details, must be established so as to obtain
a description format of landscape units (Table 18.1).
Each map zone is systematically described by filling the format. Various statistical analyses can
then be made and corresponding output maps obtained. The GIS tool would be very useful for exploiting
all the resources of landscape analysis.
Zone T167 of agrolandscape unit 1 in Table 18.1 can be described as follows (see Table 18.2). It
comprises no rivers, no water bodies, no forests or trees, no habitations. Between 5 and 35% very
bright soils cover slopes and plateaus in this zone. More than two-thirds of the zone is occupied by

Table 18.1: Example of description format for agrolandscape units interpreted visually In satellite image
(after Bertrand, 1994)

Map zone Landscape unit No. 1 9 17 Etc.

Zone No. T167 SI 69 1442

Water Rivers 0 1 0

Lakes, ponds 0 0 0

Habitat type 0 2 3
Hedges 0 0 0

ForestTrees Thickets 0 1 2

Groves 0 2 1

Blocks 0 0 0

Soil colour Very bright 2 0 0

Medium bright and dark 0 0 2

Large valley 0 0 0

Morphology Slope, hillock 3 2 2

Talweg 0 4 0

Plateau 3 0 4

Mixed forest 0 0 2

Deciduous 0 1 1

Coniferous 0 0 0

Land cover Crops 2 3 4

Villages, built-up zones 0 1 1
Grasslands, heaths 0 2 2

Vineyards and orchards 4 0 0

Riverine forests, swamps 0 3 0

Small 5 2 2

Agricultural plots Size Medium 0 2 4

Large 0 0 0

Polygonal 2 3 3
Form Quadrilateral 4 0 3
Strip-form 0 0 0

Non-agricultural 0 3 2

plots
 Applications 321

vineyards or orchards and the rest by crops. All the plots are small with more than two-thirds quadrilateral
and the rest polygonal.

18.3 EXAMPLE: AGROLANDSCAPE OF YONNE DISTRICT

The Integrated or landscape method of analysis Is useful when a large area is to be studied with
various objectives. Such is the case of district-wise and regional databanks now being prepared for
administrative management of these regions.
The example presented below pertains to a study conducted in Yonne district (CD 18.1). It is
based on the agrolandscape concept. The following aspects are discussed:
— method of description,
— interpretation procedure: delineation and description of map zones,
— assessment of visual interpretation
— classification of map zones: agrolandscape units,
— verification.

18.3.1 Method of description

The variables selected and codified concern three principal aspects of landscape: land cover,
geomorphology and plots.
These three principal components of landscape, which Indicate the physical characteristics of
natural environment and impact of human activity, are well suited for defining agrolandscape units, it
has been possible to infer these principal components from four SPOT images.
By means of stereoscopic vision, possible in the boundaries of two adjacent Images, major
geomorphological features, viz., talwegs, plateaus, slopes, valleys, hillocks, crests and peaks of a
region can be readily identified and isolated and their spatial organisation understood. Four
morphological units were identified: large valley, slope, talweg and plateau. The concept of slope
Intensity was not taken into consideration because too difficult to correctly estimate without a digital
elevation model.
For zones not covered by stereoscopy, geomorphology was determined making use of shadows
and spatial analysis: tree alignments along watercourses, wooded borders of plateaus, etc. Interpretation
of 1:100,000 topographic maps of the IGN was used for additional information.
Size and shape of agricultural parcels and non-agricultural zones have been described by visual
interpretation. Size-wise classes of agricultural plots (small, medium and large) were set up according
to the norm for parcels in Yonne district. The mean quantisised parcel corresponds to the most common
plot size In agricultural zones (zones of major crops). Size limits were so fixed that small plots are not
rare in these zones and large ones rather marginal.
Land-cover types were Identified by textural and structural analyses of images. The following
themes were distinguished:
— deciduous and coniferous forests,
— mixed forests,
— crops,
— grassy areas (mowed meadows, grasslands, fallow lands, etc.),
— vineyards and orchards,
— built-up or excavated (quarry) zones,
— riverine forests and swamp zones.
On some dates distinction between plant covers may be difficult: vineyards and crops In summer,
crops and woods in spring, winter crops and grasslands in winter, etc.
322 Processing of Remote Sensing Data

These three series of variables were codified depending on their coverage in the map zones
according to the 0 to 5 scale indicated in Chapter 5. Other variables were determined and combined
with these principal components for refining and strengthening the description of map zones.
Among structural elements of a landscape, water occupies first place in exorheic regions. In fact,
watercourses represent the main flow in natural environments; their presence characterises a number
of landscapes.
Trees also constitute a distinct element of landscape. Structure of tree populations may take
varied forms (hedges, groves, thickets, strips, blocks) and result from Interaction between man and
natural environment.
Soil is also a determinative factor of landscape. Its effect manifests through vegetation, agricultural
practices and drainage patterns.
Description of the state of soil surface, and especially Its colour, by remote sensing (see Chap.
23) can be envisaged (Yongchalermchai, 1993). However, in this study description of soils Is not dealt
with since the four images used were acquired In various seasons and soils were not viewed under
identical conditions (abundance of bare soil, degree of coverage by crops).
As far as colour is concerned, soils are classified only in two categories, viz., medium and dark;
very bright soils which are detectable In any season.
The drawings in Table 18.2 are visual representations of definitions of objects. The distinction
between a large valley and a talweg corresponds to the real situation of the district: the criterion of
width greater than 1 km facilitates characterisation of the valleys of Yonne river and its principal tributaries
(Cure, Armangon, Serein).

18.3.2 Interpretation procedure: delineation and description of

map zones
Interpretation was carried out by two interpreters using photographic contacts of images and was
restored directly on transparent films. It was hence necessary to ensure pertinence of boundaries,
reliability of interpretation and a single level of analysis.

H Pertinence (local contrast)

Demarcation of map zones is based on local contrasts of principal variables of the description system.
Contrast (Girard, 1983) is defined as the ratio of semantic (mathematical) distance to geographic
distance. It is hence evident that these boundaries are more or less distinct.
Interpretation extended to description of areal units: tracing the boundaries is therefore imperative,
even in situations of very large variations In the variables. However, If the Interpretation is to be used
in a flow modelling or a detailed spatial analysis. It is necessary to describe these boundaries by their
distinctness, form and contrast.
Boundaries of map zones correspond mainly to variations in land-cover use and to transitions
from one geomorphological form to another (often manifested by distinct break in slope).
Some boundaries, which define outlines of ‘composite’ map zones having a more complex
organisation, are validated by differences observed In secondary variables such as change of size or
shape of parcels, peripheral occurrence of bright soils, interruption in a hedge system, etc.

H Reliability of interpretation
Reliability must be ensured in any interpretation program. The attitude of the interpreter towards satellite
Images should remain similar in time and space. This demands a priori definition of precise rules of
interpretation and elaboration of a common model to serve as a reference to interpreter throughout
the Interpretation stage. This condition necessitates use of a general and reproducible method of
interpretation.
 Applications 323
Table 18.2: Definition and coding of variables used for description of agrolandscape units

WATER
River: permanent watercourse 0 = absence
1 = presence assumed or low discharge
2 = major watercourse
Lake or pond: water body of more than 200 m^ 0 = absence
1 = punctuate presence
2 = dominant presence

TYPE OF HABITAT

Visible structure of built-up area 0 = absence

1 = scattered
2 = combined

3 = mixed (scattered and combined)

FORESTS/TREES

Hedge: tree alignment 0 = absence

Wooded grove: isolated wooded plots of compact form 1 = sparse presence
and area less than 25 ha
Forest strip: wooded plots of highly elongated form and 2 = dominant presence
area less than 1 0 0 ha
Forest block: more or less compact woods of more than 25 ha 3 = dense presence (only for hedges)

r
Hedges Ex: row of trees Wooded grove
along a watercourse 1
L Ex. of wooded grove
Ex: bocage pattern
L
of hedgerows — S = 1X L < 25 ha

Strip-wooding Forest biock

Area < 100 ha
S > 25 ha

1I
Ex: woods along and over slopes
Ex: Forest block
of a talweg
Despite fragmentation, the overall
form is considered a single block

COLOUR OF SOILS

Very bright: chalky soils 0 = absence

Medium and dark: non-chalky soils 1 = diffuse presence
2 = dominant presence
324 Processing of Remote Sensing Data

MORPHOLOGY

Large valley; flat-bottomed valley of width greater than 1 km 0 = absence

1=0-5%
Valley edge: hillock, regular slope and edge of talweg of 2 = 5 - 35%
width greater than 500 m 3 = 35 - 65%
Talweg: bottom of talweg at least 1 km in width 4 = 65 - 95%
Plateau: relatively flat and horizontal zone, bounded by 5 = > 95%
distinct breaks in slopes (in % area)

Large valley
Ex: profile across a
large valley
L > 1 km

Talweg, Slope, Plateau

V talweg slope plateau

a b b‘ a
1 If ab + ab’ > 500 m, talweg
and valley edge can be differentiated

a, a’, b, b’ represent breaks if bb’ = 0 (talweg without flat bottom, V-shaped), area of
In slopes (projection of lines talweg is estimated by ‘extending’ it onto valley edges such
of break) that a belt of 2 0 0 m in width be classified as talweg
b and b’ can be combined (V-shaped
form)
L = bb’ < 1 km
2 If ab + ab’ <50 m s r
/everything’ (ab and a ‘b’) is classified as talweg

LAND COVER

Mixed forest: mixed population of deciduous and coniferous trees 0 = absence

Deciduous: deciduous population 1 = 0 - 5%
Conifer: coniferous population 2 = 5 - 35%
Crops: cultivated agricultural area (other than vineyards, orchards and grasslands) 3 = 35 - 65%
Village, built-up zone: anthropic area, constructed, urban or excavated (quarry) zones 4 = 65 - 95%
Grassland, heath: grassy area 5 = > 95%
Vineyard, orchard: grape plantations or fruit trees (In % area)
Riverine forests: moist zone with tree stands (poplar, alder)
 Applications 325
PLOT

> According to size:

Small: area < 4 ha 0 = absence
Medium: 4 ha < area < 60 ha 1 = 0 -5 %
Large: area > 60 ha 2 = 5 - 35%
> According to shape: 3 = 35 - 65%
Strip-form: rectangular plot with L > 7 x 1 4 = 65 - 95%
Quadrilateral: plot with 4 sides parallel 2 by 2 5 = > 95%
Polygonal: plot with 3 or 5 sides (in % of total area)
> According to nature:
Agricultural: crops, grasslands, vineyards
Non-agricultural: forest or built-up zone

Size
N on-agricultural
forest or built-up zone

Size of agricultural parcel

H I small ( S <4 h a )

medium (4 < S < 60 ha)

large (S < 60 ha)

Form
N on-agricultural
forest or built-up zone

Form s of agricultural parcel

^ I j strip form

m Quadrilateral (4 ‘parallel’ sides)

Q j polygonal

The interpreters proceed with an experimental (or validation) phase of the method by parallel
interpretations over the same test zones and comparison of the results. Discussions following this
study enable them to standardise their procedure and calibrate the variables on the characteristics of
the district under investigation.
326 Processing of Remote Sensing Data

As each interpreter operates on the commonly formulated and approved model, it can be ensured
that interpretation is carried out identically as It progresses and over all the zones of the district covered
by four satellite images.

ii Resolution and level of analysis

Interpretation was made on the 1:100,000 maps but the desired restitution product was a 1:250,000 (1
pixel = 0.08 mm) agrolandscape map of Yonne district. In this example, the one-fourth ru/e was used
according to which it is assumed that at less than 2.5 mm x2.5 mm (one-fourth of a cnri^) a map zone
is not readily legible on a map. It may hence be Inferred that there is no need to acquire data at higher
resolution.
It follows that the resolution required for delineating agrolandscape units corresponds to more
than 39 ha (6.25 hm x 6.25 hm) on the ground, i.e., 31 x 31 pixels.
This value indicates only an order magnitude since the decision to draw a map zone is also
dependent on Its shape and its contrast in the landscape.Thus, for elongated forms, minimum distance
between pseudo-parallel boundaries will be 2 mm, or 25 pixels in the image.
This concept of spatial resolution is very important since it Imposes a particular level of analysis
for agrolandscape. In fact, an object, even of high contrast (such as a deciduous grove), with an area
less than 39 ha, does not by itself constitute an agrolandscape unit but is an integral part of the
agrolandscape unit that surrounds it. On the other hand, it represents a strong landscape element
which, together with others, defines a spatial organisation specific to a map unit of agrolandscape.

18.3.3 Assessment of visual interpretation

H Importance of pattern
Direct visual interpretation of images on paper maps enables observation of a large geographic field
in its entirety. Each object, identified by certain characteristics and patterned in a similar manner, is
instantaneously relocated in its spatial environment. The interpreter can infer textural form of objects,
their relative size, density, connectivity and spatial distribution. Detection of patterns in the image is
related to the model of representation of the references. It is hence imperative to be explicit in defining
and coding the description variables used for agrolandscape.

® Duration
For the entire district (about 7120 km^), the duration of complete visual interpretation, including choice
of description criteria, preparation of grid, experimentation in test zones, tracing and description of
map zones was of the order of 6 weeks for an Interpreter.

M Use of topographic maps and aerial photos

Totally differing views of images acquired on different dates may be the cause of certain difficulties in
interpretation of land cover and plot distribution data. Choice of date of image is hence very Important
(see Chap. 17). In the absence of images acquired on a key date, it may be necessary to employ other
sources of information such as aerial photos. However, cost vis-à-vis accuracy required must be
computed. They should be used only when detailed geometric Information of the area Is desired.
Photographs may be used for validating the results obtained from interpretation of satellite images.
Shapes of plots and evidently, morphology of a region are better differentiated if care is taken to
purchase a stereoscopic coverage (whose cost Is twice higher). However, land cover data is not
 Applications 327

necessarily more readily detectable. The significance of satellite Images lies in obtaining data several
times in a year.
The 1:50,000 topographic map is often adequate for estimation of morphology required for
description of agrolandscape If the output map is of 1:250,000.

■ Semantic accuracy
Each map zone was described by 28 variables, the modalities of each having been coded by a strict
and rigorous system with a simple and unambiguous procedure. While the precision of estimation of
area (coded in 6 categories from 0 to 5) may appear low, it is sufficiently strong for reducing errors and
disparity between various interpreters. It must be noted that the semantic accuracy thus obtained is
largely adequate considering the description of each agrolandscape unit by a combination of 28
variables.

18.3.4 Classification of map zones: agrolandscape units

Principal component analysis was carried out with 1111 map zones, each characterised by 28 variables.
Axis 1 of principal component analysis differentiates between:
— non-agricultural plots, deciduous and coniferous blocks on the one hand,
— and polygonal plots, medium and dark soils, crops and medium size plots on the other.
The first structural element of a cluster of points is the contrast between agricultural and non-
agricultural units (essentially forests and secondarily anthropic zones). Thus 311 non-agricultural units
and 800 agricultural units were identified.
Subsequently, the minimum sorting distance method was employed (Girard, 1983).The first iteration
was based on the choice of initial reference zones, viz., the nuclei. Detailed visual interpretation enabled
step by step representation of structure and characteristics of the landscape of the district. Information
of the region under study acquired during this stage was evaluated for choice of nuclei.
After several iterations (13 in this specific case), classification was stabilised and 28 agrolandscape
units were obtained (CD 18.1). Histogram analysis of each variable enabled description of the semantic
content of each agrolandscape unit detected by this classification.
Besides the semantic significance of these units (homogeneity and originality), their spatial
distribution Is a second criterion of pertinence. In most cases, the chorological laws manifest distinctly.
This is very important In view of analysis of small regions (a higher level of analysis) characterised by
a specific spatial organisation of landscape units (CD 18.1).

18.3.5 Verification
Verification of the results of the entire study on agrolandscapes of Yonne district was carried out on 5%
of map units. The latter were selected by random choice and observed in the field. Ground-truth
verification was done regarding description of map zones, correctness of boundaries and especially
significance of agrolandscape units. For each zone, the codified description was compared with the
findings of visual observation in the field and a photograph was taken. No major error was committed
during description from satellite images: no confusion in land cover types, no Incorrect representation
of relief, no omission of tree populations, no false estimation of shape of plots, etc. In some cases and
for certain variables, small deviations were observed between descriptions from images and descriptions
In the field using the same variables. They concerned presence or absence of a stream, presence of
isolated hedges, proportions of slopes and plateaus, overestimation of vineyards and orchards
(abandoned plots), overestimation of small plots. These errors notwithstanding, visual Interpretation
of satellite images produced a reliable representation of ground truth.
328 Processing of Remote Sensing Data

With respect to correctness of boundaries, it is more difficult to draw conclusions from ground
observations. In fact, the field of view on the ground is often small (because of the shape of the Earth
or high plant cover). This considerably hinders the observer from understanding the structure of the
landscape and, in particular, organisation of basic patterns. Moreover, boundaries are often indicated
by a change in this organisation. Location of boundaries on the ground is hence difficult.
Only the boundaries corresponding to strong contrasts in a few criteria are distinctly discernible in
the field when the point of observation offers a synoptic view to the observer (Fig. 18.1). However, a
precise global assessment of accuracy of these boundaries is Impossible.
Lastly, ground observation permits verification of landscape classification over the entire region
of the district. Of 62 map zones observed, only 6 (10%) were incorrectly classified: one or two
characteristics of landscape do not correspond to the agrolandscape profile to which it belongs. These
characteristics suffice to make it appear closer to another unit.
However, caution is needed in using the ground-truth data. The concept of landscape is strongly
related to perception from space. The points of observation are so different between a satellite and a
ground observer (Fig. 18.1) that the ground-truth data must be considered only as a partial tool for
verification of results.

18.4 OTHER EXAMPLES

18.4.1 Auge region
Mapping landscape units of Lower Normandy region using satellite data was verified by socio-economic
information. The Atlas of Normandy and a thesis in economic sciences (Laurent, 1992) were used for
this purpose. This comparison validated division of Auge region Into two subunits, north and south
(Girard et al., 1995).

18.4.2 Rhone district

In Rhone district, the objective of remote-sensing study was to prepare a vocational plan of agricultural
and forest zones. The agrolandscape units, which were georeferenced on BD CARTO of the IGN,
were verified by ground truth (Boutefoy, SIRA) and compared with climatic, agronomic and other data.
They aided In augmenting the databank by geographic delineation of the perimeter concerning the
vocational plan and contributed through description of land-cover zones, to evaluation of the agronomic
potential of various regions, identification of zones sensitive to erosion and location of zones susceptible
to effects of afforestation and deforestation.
The description format was similar to that used for Yonne district. It comprised 22 variables
systematically described for each of the 2547 polygons traced (directly on the computer monitor using
a geographic information system) for Rhone district. They were regrouped into 72 map units and
through statistical analysis into 12 simplified landscape zones (CD 18.2).

18.4.3 Aube district

The objective in the case of Aube district was to define agrolandscape zones and integrate them with
soil data acquired In various parts of the district at different scales. This study enabled preparation of
a database on soils and land cover for various agricultural applications and In particular identification
of zones vulnerable to soil erosion (CD 18.3).
 Applications 329

An image showing maximum coverage of bare soil was chosen from the DALI (now SIRIUS)
catalogue (CD: SPOT system). We interpreted it on the basis of surface state of bare soils. The study
was carried out on an image enlarged to 1:50,000 scale. Vulnerability to soil erosion was categorised
in four classes after regrouping about fifteen classes. This soil vulnerability map was compared with a
slope map using G IS. Vulnerability of the region to soil erosion was thus determined. This was verified by
farmers and consultants of the Agricultural Chamber of Aube. Possible locations for plantation of hedges
and grass strips were recommended based on this vulnerability map and boundaries of agricultural plots.

18.4.4 Champagne-Ardenne region

The Champagne-Ardenne region was also investigated by computerised Interpretation of landscape
units. A mosaic was prepared with several LANDSAT TM images and processed by supervised
classification to obtain a single Image comprising 20 agrolandscape units (CD 18.4).

18.4.5 Agrolandscapes and “small agricultural zones” or

“agricultural reference units”
In various cases (Yonne, Aube, Champagne-Ardenne, Rhone), the results of agrolandscape analysis
were compared with boundaries of agricultural reference units (Baize, 1993; Eudes, 1993). It was
necessary to combine several agrolandscape units that were more detailed than agricultural reference
units or small agricultural zones. This was done using ascending hierarchic classification and OASIS
methods (see Chap. 11 ).

18.5 CONCLUSION
This method of studying landscapes described here has been used for various applications and various
places. It assumes an adaptation of agrolandscape descriptions to the objective of the investigation,
region understudy and precision required.
If GIS software is used, it is advantageous to make zoning directly on the computer monitor rather
than on paper prints of satellite images, since Integrated maps are readily available under GIS. For
working on a district with 2500 map zones, visual zoning requires a month for an interpreter. On a
computer, visual interpretation can also be carried out using several colour composites and various
techniques of Image processing, such as filtering (see Chap. 12) which enhances boundaries of plots.
With a digital elevation model (DEM), the image and the elevation map can be overlapped (CD
2.2) and several output images prepared using the DEM and the various bands of the Image.
Visual interpretation and zoning of landscape units are largely facilitated by a preliminary analysis
by OASIS which automatically creates map zones on digital criteria (digital number values of pixels
and neighbouring pixels). Thus mosaics of multiple images can be processed ensuring homogeneity
of visual Interpretation.
This method of interpretation by image-processing techniques (statistical and structural
classifications) and visual methods results in division of the region under study into a number of
agrolandscape units. The latter are defined by a reduced number of variables that suffice for analysing
the structure of the region with high accuracy. Agrolandscapes obtained are more homogeneous and
more accurate than agricultural reference units. Hence they can be more readily included in a spatial
model and used In simulation studies. In fact, the problem of semantic analysis of inhomogeneous
geographic units is not yet satisfactorily solved.
The landscape approach as an Indicator of environmental conditions and human activity, along
with a map representation, is very important for comparing various regional data sets under a geographic
Information system.
330 Processing of Remote Sensing Data

References
Baize D. 1993. Petites régions naturelles et paysages pédologiques de l’Yonne. Institut national de la recherche
agronomique, Orléans, 191 pp.
Bertrand P. 1994. Élaboration d’une base de données localisées sur les agropaysages à partir d’images satellitaires.
Application à l’étude des organisations spatiales et à la segmentation du département de l’Yonne. Mémoire
de Mastère “Système d’informations Localisées pour l’Aménagement des Territoires”. Institut national
agronomique, Paris-Grignon, 46 pp.
Deffontaines J-P. 1985. Étude de l’activité agricole et analyse du paysage. Espace géographique, Paris, 1:37-48.
Deffontaines J-P. 1986. Un point de vue d’agronome sur le paysage. Lectures du paysage. Coll. INRAP Éd. Foucher,
191 pp.
Donadieu P 1993. Du désir de patrimoine aux territoires de projets. Paysage et gestion conservatoire des milieux
humides protégés: le cas des réserves naturelles du plateau de Versailles-Rambouillet et de quelques marais
de l’Ouest, Thèse de Doctorat, Université Jussieu-Paris, 280 pp.
Eudes S. 1993. Étude des relations spatiales entre les sols et les options financières des exploitations. Mémoire
Diplôme d’Agronomie Approfondie, institut national agronomique, Paris-Grignon. 41 pp. + ann.
Francoual T, Gilliot J-M, Girard M-C. 1997. Réalisation d’une base de données localisée sur les agropaysages du
département du Rhône. Détermination des agropaysages par interprétation visuelle de données satellitaires.
Association Sol Info Rhône-Alpes, Lyon, 34 pp.
Gilliot J-M, Girard M-C. 1997. Étude de la vulnérabilité des sois à l’érosion. Chambre d’Agriculture de l’Aube,
Troyes, 14 pp.
Girard C-M, Girard M-C. 1994. Aide à la cartographie d’unités paysagères par une méthode d’analyse du voisinage
des pixels: application en Basse-Normandie, Photo-interprétation, 3-4:145-154.
Girard C-M, Girard M-C, Gilliot J-M. 1995. Qualité des méthodes d’interprétation, application à la caractérisation et
la cartographie d’unités de paysage. Qualité et validation des résultats. Bull. Soc. Franc. Photogr.Télédétection,
137: 62-66.
Girard M-C. 1983. Recherche d’une modélisation en vue d’une représentation spatiale de la couverture pédologique.
Application à une région des plateaux jurassiques de Bourgogne. Thèse Doc ès sciences. Université Paris 7.
SOLS, 12:414 pp.
Girard M-C. 1993. L’agropaysage, concept permettant de segmenter spatialement le milieu naturel à des fins
d’analyse agronomique. Rapport ministère de l’Agriculture et de la Pêche, DERF, Bureau des sols, 17 pp.
Girard M-C. 1995. Apport de l’Interprétation visuelle des images satellitaires pour l’analyse spatiale des sols. Un
exemple dans la région de Lodève. Étude et gestion des sols, 2 (1): 7-24.
Girard M-C, Girard C-M, Bertrand P, Orth D, Gilliot J-M. 1996. Analyse de la structure des paysages ruraux par
télédétection, C.R. Acad. Agri. Fr., Paris, 82 (4): 11-25.
Hénin S. 1993. Le paysage: étude sémantique. C.R. Acad. Agri. Fr., Paris, 79 (7): 29-30.
Hénin S. 1994. Le paysage, une entité pour l’appréciation du milieu. C.R. Acad. Agri. Fr., Paris, 79 (7): 30-43.
Laurent C. 1992. L’agriculture et son territoire dans la crise. Thèse ès Sciences Économiques Université Paris 7,
454 pp. + ann.
Lizet B, De Ravignan F. 1987. Comprendre un paysage, guide pratique de recherche. INRA, 147 pp.
Mollet S. 1994. Élaboration d’une base de données des agropaysages du département de l’Yonne— application à
l’étude des dynamiques financières agricoles. Mémoire Diplôme d’Agronomie Approfondie, Institut national
agronomique Paris-Grignon., 55 pp.
Rougerie G, Beroutchachvili N. 1991. Géosystèmes et paysages: Bilan et Méthodes. Armand Colin, Paris, 302 pp.
Yongchalermchai C. 1993. Étude d’objets complexes, sol/plante à différents niveaux d’organisation: de la parcelle
au paysage. Thèse de l’INA-PG. Sols, Grignon, 19:232.
 19
CORINE Land Cover

19.1 INTRODUCTION
On 27 June 1985, the European Union adopted the CORINE (Co-ordination of Information on
Environment) program under the responsibility of the DG XI (Environment, Nuclear Safety and Civil
Protection) of the European Commission. The objectives of this program, implemented during 1985 to
1990, pertained to three aspects (Cornaert and Maes, 1992):
— Collection of information on the state of the environment and development of an information
system;
— Imposition of uniformity in existing nomenclatures and development of nomenclature and
methodologies necessary for execution of the program;
— Co-ordination of activities undertaken in the member states or at the international level, aimed
at enhancing information on environment.
This program is now managed by the European Agency on Environment. Of the various themes
included in CORINE, covering geographic, biological, agricultural and other aspects concerning the
environment, land cover has been the major component since its extension to Central and Eastern
Europe from 1991. By the end of 1996, CORINE Land-Cover mapping (level 3) had been completed
over 3,000,000 km^, covering 17 European and 2 North African countries mapped with a single legend
(Table 19.1) and 17 national databases combined as an European database. In France, the Institut
français de i’environnement (French Institute of Environment) (IFEN, 1995; Bossard, 1996) is the
chief of operations of the CORINE Land Cover program within the framework of collection and public
dissemination of data necessary for environmental policy planning and choice of economic development.

19.1.1 CORINE Land-Cover mapping

Data used as the main source of Information for the program are provided by the satellites SPOT and
LANDSATTM (Thematic Mapper). In fact, satellite data constitute objective information in digital and
readily usable form. Image interpretation was carried out visually by photo-interpreters using paper
displays of 1:100,000 colour composites, aided by topographic maps, aerial photos and ground-truth
data.
The scale of the map obtained was 1:100,000 and the dimension of the smallest map unit 25 ha.
This choice emerged from the minimum information content considered necessary and the possibilities
of identifying changes over a time interval not exceeding 5 to 10 years on the one hand, and the time
and cost constraints on the other.
The nomenclature, grouped into 44 classes in the third level (CORINE Land Cover terminology),
is adopted as standard at the European level and enables comparisons between states, regions,
districts and zones of study (Table 19.1 ).
332 Processing of Remote Sensing Data

Table 19.1: Nomenclature of 44 classes of CORINE Land Cover, level 3

(after IFEN, 1995)

Level 1 Level 2 Level 3

1. Artificial surfaces 1.1 Urban fabric 1.1.1 Continuous urban fabric

1.1.2 Discontinuous urban fabric
1.2 Industrial, commercial and 1.2.1 Industrial or commercial units
transport units 1.2.2 Road and rail networks and
associated land
1.2.3 Port areas
1.2.4 Airports
1.3 Mine, dump and construction 1.3.1 Mineral extraction sites
sites 1.3.2 Dump sites
1.3.3 Construction sites
1.4 Artificial non-agricultural 1.4.1 Green urban areas
vegetated areas 1.4.2 Sport and leisure facilities

2. Agricultural areas 2.1 Arable land 2.1.1 Non-irrigated arable land

2.1.2 Permanently irrigated land
2.1.3 Rice fields
2.2 Permanent crops 2.2.1 Vineyards
2.2.2 Fruit trees and berry plantations
2.2.3 Olive groves
2.3 Pastures 2.3.1 Pastures
2.4 Heterogeneous agricultural areas 2.4.1 Annual crops associated with
permanent crops
2.4.2 Complex cultivation patterns
2.4.3 Land principally occupied by
agriculture with significant areas of
natural vegetation
2.4.4 Agroforestry areas

3. Forests and semi­ 3.1 Forests 3.1.1 Broad-leaved forests

natural areas 3.1.2 Coniferous forests
3.1.3 Mixed forests
3.2 Scrub and/or herbaceous 3.2.1 Natural grasslands
vegetation associations 3.2.2 Moors and heathland
3.2.3 Sclerophyllous vegetation
3.2.4 Transitional wood-land scrub
3.3 Open spaces with little or 3.3.1 Beaches, dunes and sands
no vegetation 3.3.2 Bare rocks
3.3.3 Sparsely vegetated areas
3.3.4 Burnt areas
3.3.5 Glaciers and perpetual snow

4. Wetlands 4.1 Inland wetlands 4.1.1 Inland marshes

4.1.2 Peat bogs
4.2 Coastal wetlands 4.2.1 Salt marshes
4.2.2 Salines
4.2.3 Intertidal flats

5. Water bodies 5.1 Inland waters 5.1.1 Watercourses

5.1.2 Water bodies
5.2 Marine waters 5.2.1 Coastal lagoons
5.2.2 Estuaries
5.2.3 Sea and oceans
 Applications 333

Organisations involved in the mapping of the French territory were not the same in all regions.
Land-cover mapping of the Ile-de-France region was carried out by the Institut d'aménagement et
d’urbanisme (Institute of Management and Urban Affairs) of the Ile-de-France region (lAURIF) in a
1:100,000 scale, conforming to the recommendations of the CORINE Land Cover program. However,
the visual Interpretation employed corresponded to a more detailed nomenclature (levels 4 to 5) and
minimum resolution of 4 to 5 ha. Lateri only the 3'^^ level Interpretation was retained during integration
of the map of Ile-de-France with the rest of the CORINE Land Cover database.
The entire Ile-de-France region was mapped using 9 SPOT scenes acquired in equivalent periods
(May and June) during two successive years (1989 and 1990), during which urbanisation had just
commenced. Preparation of CORINE Land-Cover map involved the following operations:
1) geometric corrections using control points,
2) re-sampling by restoration (pixel of 20 m side) in Lambert Zone I projection, to ensuring maximum
accuracy, and
3) construction of a mosaic, subsequently divided Into four zones, on 1:100,000 photographic contact
of colour composites.
The work was completed In two phases:
1) mapping of agricultural areas of the rim districts (accuracy 25 ha) with ground control of sectors
of doubtful Interpretation;
2) mapping of urban zones of the Paris agglomeration and peripheral districts (accuracy 4 ha).
22 classes of level 3 were thus interpreted, in which 27 classes of levels 4 and 6 of level 5 were
integrated. Maintenance of CORINE Land Cover (level 3) norms was later entrusted to a service
society.
Interpretation of all the nine images of the Ile-de-France region required employment of two
photo-interpreters for a period of about 6 months. Despite the experience of the photo-interpreters,
the visual Interpretation showed certain subjectivity, however. For a given date, risks of error varied
according to the classes and for a given class, according to the date of acquisition. The accuracy of
Identification of classes hence depends on their nature and the date of acquisition of the satellite data.
The lAURIF compared the CORINE Land-Cover map with the existing land cover modes in the region
(lAURIF, 1995), for which only the polygons of more than 5 ha were taken into consideration. Correlation
between the two maps was observed in 84 to 85% cases.

19.1.2 Automatic mapping

A 20 by 18 km zone covering the southern part of the Val-d’Oise district and the northern part of the
Yvelines district was mapped by lAURIF using the SPOT scene (KJ 038-251) of 17 May 1989. Nineteen
classes were identified In this zone.
The land-cover map of this small zone west of Ile-de-France, obtained from CORINE Land Cover
program, was compared with a map of the same theme prepared by applying methods of computerised
classification and aggregation in a SPOT image (see Chaps. 8, 9 and 11). The objective of this
comparison was to assess the potential of computerised methods for preparing a map at level 3 and
a more detailed level (level 4). The aim was not to model the results of visual interpretation, but to
compare the results obtained by various methods.
Computerised mapping was carried out on a multispectral SPOT scene of 11 March 1995. This
date was particularly favourable since a strong radiance contrast existed between forests (mainly
broad-leaved or mixed) and agricultural plots. Moreover, within agricultural areas, annual crops
corresponding to bare soils or sparse or very sparse chlorophyllian vegetation, are readily distinguished
from short-duration fallow lands (dense and non-chlorophyllian vegetation) at the end of winter and
perennial herbaceous vegetation— grasslands and long-duration fallow lands (dense and chlorophyllian).
The results were verified using ground observations in the spring of 1995 and July 1996 and from
aerial photographs of the National Geographic Institute. Quantitative validation was applicable only to
334 Processing of Remote Sensing Data

invariant zones and this posed problems for classes of artificial surfaces, since in this region construction
openings are common. Hence, a qualitative assessment of the results was also made. Contrarily, no
risk of modifications was observed for other classes, which was confirmed by ground check-up.
Land-cover classification was conducted pixel by pixel and the zones obtained were subjected to
spatial integration. The final results were entered In Mapinfo® geographic information system, after
geographic restitution using TeraVue program. Flow chart of the method employed is shown in Fig. 19.1.

Fig. 19.1: Flow chart of the method used for computerised mapping.

19.2 DATA PROCESSING AND DISCUSSION OF RESULTS

19.2.1 Supervised classification of land cover
In the infrared colour composite, 251 reference plots identified on the ground were delineated on the
computer monitor and used to determine the parameters of 17 classes. The number of reference plots
for each theme varied from 4 to 30. In fact, there was no need to increase the number of reference
areas for themes with no ambiguity such as water (7 references). On the other hand, It was very
difficult in some cases to locate a large number of reference areas for the classes few represented in
the subscene, viz., quarries and industrial units (4 references). Contrarily, more numerous reference
areas were necessary for characterising heterogeneous areas such as fallow lands (24 references) or
spring crops (30 references).
The entire subscene was then classified by ‘maximum-likelihood method with Gaussian assumption'
using MULTISCOPE (Multiscope, 1993) program which In fact pertains to the Bayesian group, with a
priori choice of probability. In order to obtain a land-cover map as close to CORINE Land Cover as
 Applications 335

possible, summer crops (corresponding to various bare soils), winter crops and spring crops were
combined under the theme ‘arable lands’. To compare the results of a classification with a map in
which no ‘unclassified’ category exits, it was imperative to classify maximum number of pixels of the
image. A classification of 60 to 94.9% of reference plots was obtained (Table 19.2), with more than
80% correctly classified for 8 of the 13 classes. Final classification of the entire image segment resulted

Table 19.2: Results, in number of pixels and percentage (bold), of classification of reference areas after application
of maximum-likelihood method.

H Accords (number and percentage of correctly classified pixels), major diagonal of the error matrix Kappa =
0.75

Ground- 10 11 12 13 rejec­ Number

truth tions of pixels
classes (%)

1 Quarries 702 31 48 781

89.9 4 6.1 100
2 Arable 401 6661 654 152 453 143 28 171 8663
land 4.6 76.9 7.5 1.8 5.2 1.7 0.3 2.0 100
3 Broad-leaved 5 158 158 4 68 23 12 2164
forests 0 .2 7.3 7.3 0.2 3.1 1.1 0.6 100
4 Coniferous 30 6 I 289 3 18 3 1389
forests 2 .2 0.4 20.8 0.2 1.3 0.2 100
5 Mixed 112 79 8 8 0 1794
forests 6.2 4.4 0.4 0.4 0 100
6 Old gravel 16 2 28 123 3103
pits 0.5 0.1 0.9 4.0 100
7 Recent 2 1 16 17 449
gravel pits 0.4 0.2 3.6 3.8 100
8 Fallow 148 59 12 11 1 46 349 5 2434
lands 6 .1 2.4 0.5 0.5 0.0 1.9 14.3 0.2 100
9 Grasslands 41 3 96 1486
2 .8 0.2 6.5 100
1 0 River 63 5 5 1444
4.4 0.3 0.3 100
11 Continuous 6 65 70 4 4 122 384 333 56 2888
urban fabric 0 .2 2.3 2.4 0.1 0.1 4.2 13.3 11.5 1.9 100
1 2 Discontinuous 10 242 595 29 71 1393 24 1497 8 6 107 8101
urban fabric 0 . 1 3.0 7.3 0.4 0.9 17.2 0.3 18.5 I 1.1 1.3 100
13 Industrial 5 2 365 114 5087

J
8 6

units 1.7 1.1 0.0 7.2 2.2 100

Classifi­ 1205 7279 2736 1164 2120 2934 492 3981 1523 1373 4285 4975 4959 757 39783
cation 3.0 18.3 6.9 2.9 5.3 7.4 1.2 10.0 3.8 3.5 10.8 12.5 12.5 1.9 100
% commis- 41.7 8.5 30.8 10.7 25.1 16.1 54.7 11.7 0.1 57.0 18.7 10.0
Sion
errors
336 Processing of Remote Sensing Data

in a rejection of 1.86 % or 669.6 ha of unclassified area, corresponding to pixels of very heterogeneous

contents in the boundaries of discontinuous urban fabric, gravel pits under excavation or forest blocks.
Accords In Table 19.2 indicate the number (or percentage) of correctly classified reference pixels
(shaded squares along the diagonal of the error matrix). Errors correspond to pixels affected to a class
when they truly belong to another. Errors of commission (inclusions) computed along rows of the error
matrix are distinguished from errors of omission (exclusions) computed along columns. A classification
Is statistically judged as very correct when agreement corresponds to 80% or more pixels of reference
plots (Lillesand and Kiefer, 1994). The themes ‘old gravel pits, recent gravel pits, grasslands, rivers,
industrial zones’ are correctly classified (87.7 to 94.9%) with a few errors (0 to 16.1 %), whereas ‘fallow
lands’ and ‘continuous urban fabric’ are less correctly classified (63.9 and 74.1 %) and present numerous
errors (54.7 and 57%). The themes ‘quarries’, ‘broad-leaved forests’ and ‘mixed forests’ are correctly
classified (87.5 to 89.9%), but with significant number of errors (25.1 to 41.7%). Lastly, the themes
‘arable lands’, ‘coniferous forests’ and ‘discontinuous urban fabric’ are incorrectly classified but with
fewer errors (8.5 to 18.7%).
Excluding major crop areas, the classification result is presented as a mosaic of small zones of
various colours. Indicating the complexity of spatial organisation of land-cover units, in particular in
semi-urban regions. Such a situation is observed In the classification of Saint-Christophe sector of the
new town of Cergy-Pontolse, in which forest and agricultural areas are intermixed. For catering to the
needs of managers who desire integration of land-cover maps In a GIS, we should have maps that
synthesise Information in the form of compact map units, albeit less accurate. It is hence necessary to
conduct spatial integration of the classification results obtained.

19.2.2 Spatial integration

Spatial integration of the results was carried out using OASIS and VOISIN programs (Girard, 1992;
see Chap. 11). The OASIS method was applied to the results of the preceding classification. For this,
one or several polygons corresponding to various classes were selected, avoiding the class ‘quarries’
because of its small size and confusion with ‘arable land’ (Table 19.2). For tracing the polygons, the
classes ‘continuous urban fabric and ‘discontinuous urban fabric’ showing small and complex units
were combined Into a single one. The heterogeneity index, expressed as a function of window size,
varied from 5 for the most homogeneous (river) to 23 for the most heterogeneous (summer crops).
Processing was successively repeated with several window sizes close to that of the most homogeneous
class, viz., 3 X 3, 5 X 5 and 7 x 7. In fact, use of a larger window resulted in disappearance of small-
size units corresponding to the most homogeneous classes.
These precautions notwithstanding, spatial integration resulted in erosion or even complete
elimination of linear forms. Hence, we have chosen to insert, in the final map, units entitled ‘watercourses’
obtained from supervised classification.
Note the incompatibility between spatial integration, which results In delineation of fewer and
more compact map units, and semantic accuracy (see Chap. 15). Therefore, we have retained the
result corresponding to window size 5 x 5 or a precision of 100 m x 100 m on the ground.
This land-cover classification took 1 week: 3 days for supervised classification (selection and
tracing of reference plots is the longest) and 2 days for spatial integration (data analysis for choice of
moving windows was the longest operation). After geographic restitution, the results of classification
were compared with CORINE Land Cover data using GIS.

19.2.3 Assessment of results

The quality of a map can be expressed in terms of the parameters semantic accuracy and geometric
accuracy (Chap. 17). The accuracy assessment presented below corresponds mainly to semantic
 Applications 337

accuracy. The results were verified with aerial photos and ground-truth survey. 334 ground control
points were chosen by systematic sampling. To ensure adequate coverage of units of small size and/
or of sinuous forms, and hence undersampled by this method, 28 additional points were selected by
directed sampling. In total, 362 points were used for assessment of results.

B Semantic evaluation of automatic mapping vis-à-vis ground truth

The nomenclature chosen as reference was that of CORINE Land Cover. Some classes recognised in
aerial photos did not figure in the land-cover map obtained from computerised classification. The
various errors are discussed below.

□ Errors associated with automatic mapping method

Some classes such as ‘Fruit trees and berry plantations, road and railway networks’ were not identified
since it is impossible to distinguish them by spectral criteria alone without structural and shape criteria.
The boundaries of the class ‘watercourses’ were obtained from supervised classification. All the 15
sampling points were correct with no affectation errors. The use of a 5 x 5 window for spatial
generalisation can explain overestimation of ‘continuous and discontinuous urban fabric’ at the expense
of ‘non-irrigated arable land’.

□ Errors associated with use of single-date image

Confusions between ‘pastures’ and ‘non-irrigated arable land’ and between various types of forests
are eliminated to a great extent by using another image of a different date, such as June 1995,
corresponding to other physiological states of vegetation.

□ Errors due to content of certain CORINE classes

The classes ‘transitional wood-land scrub“' ’ , ‘complex cultivation patterns^ ’ constitute ‘mixed’ classes,
considering a minimum area of 25 ha. Here it corresponds only to a verification point, which explains
their non-identification in the classification and the affectation errors observed.

□ Errors related to the nature of some CORINE classes

These errors are associated with ‘industrial or commercial units, airports, construction sites, mineral
extraction sites, green urban areas, sport and leisure facilities’ and ‘water bodies’ which in reality do
not correspond to a land-cover type but a land-use category. In fact, the first class represents mineral
areas, hence classified as ‘continuous urban fabric’, whereas ‘airports’ in the region are large herbaceous
areas and hence classified as ‘pastures’ or even ‘crops’. The ‘green urban areas’ and ‘sport and leisure
facilities’ are likewise grouped under ‘forests’ or ‘pastures’. In the case of ‘mineral extraction sites’,
which Include underwater gravel, some may be confused with ‘water bodies’ or even with ‘sport and
leisure facilities’.

B Use of the concept of fuzzy subgroups

The existence of the four types of error mentioned above has led not only to a quantitative assessment
but also to a qualitative evaluation of accuracy of this classification. The latter is established according
to the ultimate use of the map obtained from the decision of an expert who identifies the severity of
errors. For example, confusion between ‘broad-leaved forests’ and ‘continuous urban fabric’ is more
serious than between ‘broad-leaved forests’ and ‘mixed forests’. This error is expressed in the form of

^Definition: shrubby or herbaceous vegetation with scattered trees. Such formations may result from degradation of
forest or recolonisation/regeneration by forest. C itedin: lAURIF, 1995.
^Definition: Juxtaposition of small plots of various annual crops, grasslands and/or permanent crops. Cited in:
lAURIF. 1995.
338 Processing of Remote Sensing Data

a notation according to the following linguistic scale, close to that described by Gopal and Woodcock,
1994:
5: title of class exactly corresponds to the land-cover type— no error;
4: title of class pertains to a similar land-cover type— small error;
3: title of class corresponds to a neighbouring category— medium error;
2: title of class pertains to a different category— large error;
1: title of class pertains to a totally different category— very large error.
In the context of land-use type, expert judgement enables construction of a table of affectation
errors between the land-cover map and the CORINE nomenclature (Table 19.3).

Table 19.3: An expert’s qualitative assessment of errors of affectation between the land-cover map and CORINE
nomenclature for verification points

CORINE Land 1 . 1.1 1.1.2 2.1.1 2.3.1 3.1.1 3.1.2 3.1.3 5.1.1 5.1.2
Cover classes

Continuous urban fabric

Discontinuous urban fabric 1 1
Industrial or commercial units 1
Roads and rail networks
Airports
Mineral extraction sites
Construction sites
Green urban areas
Sport and leisure facilities
Arable land
Fruit trees and berry
plantations
Pastures
Complex cultivation patterns
Broad-leaved forests
Coniferous forests
Mixed forests
Transitional wood-land scrub
Watercourses
Water bodies

It was thus possible to establish a fuzzy relationship between various units and accuracy levels
(Table 19.4), based on the preceding Table and the error matrix using the 362 control points.
The units ‘watercourses, old gravel pits, coniferous forests, non-irrigated arable land, industrial
units’ have more than 80% points classified in category 5 (no error), sometimes with a certain percentage
in class 4 (small error). The unit ‘mixed forests’ has 98% of its points classified in categories 5 and 4.
The units ‘continuous and discontinuous urban fabric’ and ‘broad-leaved forests’ have more than 70%
points classified In category 5, with about 10% points in category 1 (very large error). Lastly, the unit
‘pastures’ shows 58% points distributed in categories 1 ,2 (large error) and 3 (medium error). In the
last case, these errors are due partly to the use of a single date and partly to the nature of CORINE
 Applications 339
Table 19.4: Representation of the fuzzy relationship (expressed as percentage) between the land-cover classes
and qualitative accuracy classes for comparison with control points

Classes Qualitative accuracy notation

_ _ _ _ _

Continuous and discontinuous urban fabric 14 2 3 1 80

Industrial or commercial units 4 0 0 11 85
Arable land 9 0 7 2 82
Pastures 17 1 2 29 0 42
Broad-leaved forests 8 3 0 13 76
Coniferous forests 0 0 0 2 0 80
Mixed forests 0 2 0 24 74
Watercourses 0 0 0 0 1 0 0

Old gravel pits 3 0 0 0 97

For entire image 11 3 6 6 74

Land Cover classes. The entire image shows more than 74% points as very correctly classified (category
5) and 11% with very large error (category 1).
In conclusion, while the accuracy of this map vis-à-vis the control points is overall acceptable for
an user (Kappa coefficient equal to 0.72), the magnitude of errors varies according to categories,
which may have more or less serious consequences depending on the ultimate use of the map.

■ Comparison between land-cover map and CORINE Land-Cover map

Overall, 53% points are classified correctly (Kappa coefficient equal to 0.50) but very large disparities
are observed depending on the classes.

□ Very correctly classified classes

There are no classes with high percentage of accords and low values of commission and omission
errors.

□ Correctly classified classes

‘Pastures’ and ‘water bodies’ as well as ‘continuous and discontinuous urban fabric’ and ‘Industrial or
commercial units’ have a high percentage of accord and omission errors, while ‘watercourses’ and
‘non-Irrigated arable land’ have medium percentages of accord and omission errors and low levels of
commission errors. The reasons for this poor classification have already been given.

□ Incorrectly classified classes

These classes exhibit low percentage of accords, medium level of commission errors and high omission
errors. They comprise ‘broad-leaved forests’ confused, on the one hand, with ‘discontinuous urban
fabric’ (question of a spectral differentiation threshold between these two categories) and, on the
other, with ‘green urban areas’ and ‘sport and leisure facilities’ (difficulty in identifying land-use classes
and not land-cover classes). Confusions with ‘arable land’ are due to the effects of the 5 x 5 window.
As earlier, an accuracy assessment of errors was also done.The results are shown in Table 19.5.
For the entire image, 59% points are classified without error and 71 % correspond to the accuracy
classes 5 and 4. Contrarily, 18% points correspond to a very large error (category 1). Detailed analysis
of the units shows that only the class ‘watercourses’ has 74% of its points in category 5, while the
classes ‘coniferous forests’ and ‘mixed forests’ have no points in category 5 and 92% and 82% points,
respectively, in category 4. The classes ‘broad-leaved forests’, ‘old gravel pits’ and ‘industrial or
commercial units’ have, respectively, 73% and 62% points in categories 5 and 4, whereas tow-thirds of
340 Processing of Remote Sensing Data

Table 19.5: Matrix representation of fuzzy relationship (in percentage) between units of the land-cover map and
qualitative accuracy categories for comparison with the CORINE Land-Cover map

Classes Qualitative accuracy notation

1 2 3 4 5

Continuous and discontinuous urban fabric 44 6 4 4 42

Industrial or commercial units 38 0 0 31 31
Arable land 29 0 15 1 55
Pastures 1 0 17 63 8 2

Broad-leaved forests 19 5 3 45 28
Coniferous forests 0 8 0 92 0

Mixed forests 18 0 0 82 0

Watercourses 26 0 0 0 74
Old gravel pits 27 0 0 40 33
Entire image 18 3 8 1 2 59

points for the last two units lie in category 1. ‘Arable land’ and ‘continuous and discontinuous urban
fabric’ respectively show 55% and 42% points in category 5, with 29% and 44% points in category 1.
Lastly, ‘pastures’ have 63% points in category 3, the remaining being distributed in other classes.
Low correlation was observed between the land-cover map and the CORINE map and hence
explanatory hypotheses have been simply indicated. However, for a better understanding of the sources
of these errors, the CORINE Land-Cover map has also been compared with aerial photographs and
ground data.

8 Verification of CORINE Land-Cover map with aerial photos and ground data
A complete analysis of various maps also necessitated verification of the 362 reference points of the
CORINE Land-Cover map using aerial photography and ground investigations (Kappa coefficient
equal to 0.56 and overall accuracy 0.58).
Various cases can be distinguished but the unit ‘transitional wood-land scrub’, which had no
control points, will not be discussed here.

□ Unmapped classes
A study of the error matrix (Table 19.6) indicated certain classes (construction sites, fruit trees and
berry plantations, coniferous forests, mixed forests) Identified from aerial photos and on the ground
but not represented In the result. In the case of ‘construction sites’, their nature and the difference in
the year of acquisition of SPOT data and aerial photos explain why they were not mapped. For three
other classes, the 1;100,000 scale of the map has certainly resulted in a simplification of boundaries
and omission of small-size zones by the mapping personnel. These unmapped classes contributed to
the percentage of commission errors for the units ‘continuous urban fabric, non-irrigated arable land,
decidous forests’. Moreover, the date of acquisition of the scene used for classification (17 May 1989)
was perhaps less favourable for identification of these themes.

□ Class readily identified by its specific shape

The theme ‘airports’ represents this case. The results Indicated 50% accords and 0% commission
errors. The 50% omission errors can be explained to be due to the map scale used.
 Table 19.6: Error matrix of the CORINE Land-Cover map vis-à-vis aerial photos and ground investigations

CORINE Land Cover

Aerial photos
111 1 1 2 121 1 2 2 124 131 133 141 142 2 1 1 222 231 242 311 312 313 324 511 512 Total Accord
(%)
.1 . 1 Continuous 1 16 1 1 1 4 1 25 4
urban fabric
.1 . 2 Discontinuous 33 1 13 3 3 2 2 57 58
urban fabric
.2 . 1 Industrial or 3 4 1 2 1 1 1 2 33
commercial units
.2 . 2 Roads and rail 1 1 1 1 2 6 17
networks
1.2.4 Airports 1 2 1 4 50
1.3.1 Mineral 1 7 1 9 78
extraction zones
1.3.3 Construction sites 1 1 1 3 0

1.4.1 Green urban areas 5 1 6 0

1.4.2 Sport and 1 1 4 1 7 57

leisure facilities
.1 . 1 Arable land 2 114 2 118 97
.2 . 2 Fruit trees and 2 7 1 1 0 0

berry plantations
2.3.1 Pastures 1 9 1 3 1 15 7
2.4.2 Complex cultivation 1 1 0

patterns
3.1.1 Broad-leaved 1 9 23 33 72
forests
3.1.2 Coniferous forests 1 7 8 0

3.1.3 Mixed forests 1 3 21 25 0

3.2.4 Transitional 1 1 1 0 0

wood-land scrub
5.1.1 Watercourses 1 14 15 93
5.1.2 Water bodies 1 2 5 8 63
Total 2 67 8 1 2 9 5 7 167 1 9 57 5 18 5 362
Percentage of 50 51 50 0 0 2 2 1 0 0 43 32 0 1 0 0 60 80 2 2 0

commission errors

Accords (number of correctly classified pixels).

342 Processing of Remote Sensing Data

□ Correctly mapped classes

The criteria of shape and location, combined with texture, enable correct visual identification, as
illustrated by the small number of omission errors. Such units are ‘mineral extraction sites’ (78% accords,
22% commission errors and 22% omission errors), ‘watercourses’ (97% accords, 22% commissions
and 3% omissions) and ‘non-irrigated arable land’ (97% accords, 32% commissions and 3% omissions).
The commission errors of this class are due to the scale of the map, which assigns to it zones smaller
than 25 ha in area belonging to other units, especially ‘discontinuous urban fabric, fruit trees and berry
plantations, pastures, broad-leaved forests’.

□ Classes incorrectly classified due to inappropriate map scale

Such classes are ‘continuous urban fabric’ (4% correct, 50% commission errors and 96% omission
errors), ‘discontinuous urban fabric’ (58% accords, 51% commissions and 42% omissions), ‘roads
and rail networks and associated land’ (17% correct, 0% commissions and 83% omissions), ‘pastures’
(7% correct, 0% commissions and 93% omissions) and ‘broad-leaved forests’ (72% accords, 60%
commissions and 28% omissions). The scale of map representation could be the source of omission
errors, as well as commission errors from other classes, as just mentioned above.

□ Classes incorrectly classified due to their definition

The concept of land use contained in the definition of classes is also a source of errors for some units
such as ‘sport and leisure facilities’ (57% correct, 43% commission errors and 43 % omission errors),
‘industrial or commercial units’ (33% accords, 50% commissions and 67% omissions), ‘green urban
areas’ (0% correct, 100% commissions and 100% omissions) and ‘water bodies (63% correct, 0%
commission and 37% omission errors).

□ Classes incorrectly classified due to unknown causes

A systematic verification was done in more than 362 points over the entire region of Ile-de-France for
the class ‘complex cultivation patterns’. Of the 88 map zones existing in the CORINE Land-Cover, 3
are correct and the other 85 correspond to villages.
An error matrix of qualitative accuracy assessment was established and applied to the CORINE
map (Table 19.7). It shows 82% of the total points classified in categories 4 and 5 and 13% in category 1.
Classes showing the smallest errors (75% points in categories 4 and 5) are ‘coniferous forests,
pastures, mixed forests, green urban areas and broad-leaved forests’.
Classes with the highest errors (more than 25% points in categories 1 and 2) are distributed in
several classes.
1. Errors (omission or commission) related to year differences between the SPOT data used for
CORINE Land Cover, aerial photographs and ground check-up. They are associated with the
classes amenable for rapid changes: ‘construction sites, discontinuous urban fabric. Industrial or
commercial units, transitional wood-land scrub’ and ‘non-irrigated arable land’.
2. Errors (omissions) related to scale of the CORINE Land-Cover map and small sizes of areas
mapped (25 ha and linear objects less than 100 m long). These are ‘roads and rail networks’ and
‘waterbodies’.
It follows from these various analyses and comparisons that automatic land-cover mapping does
not permit representation of certain classes of the CORINE Land Cover, since they correspond In fact
to land-use types but not land-cover units. Visual interpretation of satellite data enables mapping of
these classes by the shape and geographic location of objects. To improve the automatic mapping It will
be necessary to employ the criteria of shape and neighbourhood, as well as a cross-check with digital
elevation models for formulating decision rules (expert systems) for mapping these classes.
 Applications 343
Table 19.7: Matrix representation of the fuzzy relationship between CORINE Land-Cover map and qualitative
accuracy notations for comparison with control points.

Classes Qualitative accuracy notation

1 2 3 4 5

Continuous urban fabric 0 .2 0 0 .1 1 0.65 0.04

Discontinuous urban fabric 0.33 0 0 .0 1 0.30 0.36
Industrial or commercial units 0.30 0 0.05 0.40 0.25
Roads and rail networks and associated spaces 0.32 0.17 0.17 0.17 0.17
Airports 0.25 0 0 0.25 0.50
Mineral extraction zones 0.09 0 0.27 0 0.64
Construction sites 0.33 0 0.33 0.33 0

Green urban areas 0 0.09 0.09 0.82 0

Sport and leisure facilities 0 .1 0 0 .2 0 0 0.30 0.40

Non-irrigated arable land 0.32 0 0 0 .01 0.67
Fruit trees and berry plantations 0 .2 0 0 0 .1 0 0.70 0

Pastures 0.07 0 0.07 0.79 0.07

Complex cultivation patterns 0.44 0 0 0.56 0

Broad-leaved forests 0.06 0 .0 1 0.16 0.42 0.35

Coniferous forests 0 0 0 1 0

Mixed forests 0.04 0 0 .1 2 0.84 0

Transitional wood-land scrub 0.80 0 0 0 0 .2 0

Watercourses 0.26 0 0 0 0.74

Water bodies 0.37 0 0 0 0.63
For entire image 0.13 0 .0 1 0.05 0 .2 2 0.60

19.3 CONCLUSION
Land-cover mapping constitutes an essential information for environmental management. The CORINE
Land-Cover program produces maps with the same nomenclature for all the European countries.
Hence this is an important stage in enhancing information on environment. However, the scale of the
map is 1:100,000 and we should not search for more precise information than that permitted by the
rules applied In preparing this map. Moreover, zones of certain classes undergo very rapid temporal
variations, which diminishes the value of information. Visual interpretation enables mapping of land-
use classes, which Is not possible from computerised classification alone.

References
Bossard M. 1986. Production de la base de données CORINE Land Cover. Séminaire IFEN ‘CORINE Land Cover’.
11-13 décembre 1996, pp. 11-18.
Cornaert MH, Maes J. 1992. Land cover, an essential component of the CORINE information system on the
environment. GIS implications, Proc. Central Symposium‘International Space Year’ Conf., Munich, Germany,
30 March-4 April 1992. ESA SP-341, pp. 473-481.
344 Processing of Remote Sensing Data

Girard M-C, Yongchalermchai C, Girard C-M. 1992. Analyse d’un espace par la prise en compte du voisinage.
Gestion de l’espace rural et système d’information géographique. Séminaire de Florae, 22-24 Oct. 1991.
INRA, Paris, pp. 349-359.
Gopal S, Woodcock CE, 1994.Theory and methods for accuracy assessment of thematics maps using fuzzy sets.
Photogrammetric Engineering & Remote Sensing, 60; 181-188.
lAURIF. 1995. Les ‘ecozones’ d’îie-de-France, dans le cadre du programme européen CORINE Land Cover affiné.
D8-334, Région Île-de-France, Centre national d’études spatiales, 28 pp.
IFEN. 1995. Programme CORINE Land Cover, 8 pp.
Lillesand TM, Kiefer RW. 1994. Remote Sensing and Image Interpretation. John Wileys Sons, NY, 750 p.
Multiscope. 1993. Manuel d’utilisation, 150 pp.
 20
Herbaceous Formations and
Permanent Grasslands
Some applications of remote sensing in studies of herbaceous formations and permanent grasslands,
mainly of temperate zones and a few examples of intertropical zones, are illustrated in this chapter.
Humid grasslands and marshes will be discussed partly in this chapter and partly in Chap. 21 —
Wetlands.

20.1 TERRESTRIAL HERBACEOUS FORMATIONS

Botanical and biogeographical investigations have resulted in identification of several types of
herbaceous formations in the world. The first level of distinction is based on climax'* or other origin of
these formations, the second on perenniality of vegetation species and the third on major climatic
types.
Various types of herbaceous formations are represented by varied floristic compositions and
evolution dynamics, knowledge of which is imperative for managers (Table 20.1). In fact, while the
area of secondary herbaceous formations in intertropical zones Increases at the expense of tree
formations, an Inverse trend is observed in temperate zones (particularly in Europe), permanent
grasslands being in regression to the advantage of crops as well as natural or Induced reforestation.
Thus, in France, 3.5 Mha (25% of the initial area) of permanent grasslands had disappeared between
1970 and 1995, of which 67% were replaced by cereal crops, 10% by fallow lands and 3% by woods
(IFEN, 1996).
Various types of herbaceous areas, including temporary, artificial pastures and permanent
grasslands, in the nomenclature of the Service Central des Etudes Economiques et Statistiques
(SCEES) (Central Service of Economic and Statistical Studies), are given in Table 20.2.
Herbaceous formations are characterised by highly variable quantities of aerial or subsurface
biomass depending on their climax situation, degree of perenniality and mode of usage. Decrease (or
disappearance) of pasture pressure leads to large accumulation of aerial phytomass, dry fraction in
particular, accompanied by an Increase in litter, while quantities of root biomass are relatively stable.
Differences In the quantity of fresh phytomass (a) and percentage of dry matter in it (b) for three
types of permanent grasslands corresponding to three different levels of exploitation are shown in Fig.
20.1. Decrease in exploitation results in increase in the aerial phytomass (a). Contrarlly, phytomass of
a non-exploited rangeland Is less since it corresponds to a botanical unit other than the two grasslands.
This shows the importance of differentiating grasslands and herbaceous formations according to their
botanical and agronomic type. Increase in percentage of dry matter in aerial phytomass with decrease

^Climax: state of an ecosystem having reached a relatively stable stage of equilibrium (at least in the human
scale), conditioned only by climatic and edaphic factors (Delpech et al., 1985).
346 Processing of Remote Sensing Data

Table 20.1: Major types of herbaceous formations in the world

in exploitation or its absence is seen in Fig. 20.1 (b). This variation in percentage of dry matter with
level of exploitation plays an Important role In studying herbaceous formations and permanent
grasslands by remote sensing.

20.2 PROBLEMS
Herbaceous formations and permanent grasslands perform various functions:
1. Production of grass and fodder for feeding the cattle and wild fauna;
2. Protection and conservation of environment (protection areas in sloping basins for tapping potable
water, protection of soils against wind- or water-borne erosion, etc.);
3. Conservation of vegetation and animal species;
4. Contribution to aesthetic quality and diversity of landscapes, etc.
 Applications 347

Table 20.2: Categories of herbaceous areas (after SCEES)

Duration Species or type Category

HA: Cultivated grasslands Pure legumes or mixed legumes Artificial leys

Herbaceous area of 0 to 5 years
comprised in UAA*
Italian rye-grass or hybrid rye grass Short-duration
temporary leys
Pure graminaceous plants or Temporary leys
mixed graminaceous-leguminous
plants or mixed graminaceous plants

Outside UAA Cultivated grasslands Variable floristic composition Long-duration temporary

of 6 to 1 0 years grasslands
Uncultivated or cultivated Arable Arable PGA**
grasslands of more than
1 0 years Non-arable PGAC (compulsory)
Others (rangelands, shrubs,
heaths, high-altitude pastures,
marshes, etc.) Rough grazings

*UAA (Used Agricultural Area) = crops + HA (Herbaceous Area)

**PGA: Perennial Grassy Area

■ Exploited grassland ■ Less exploited grassland □ Unexploited grassland

Fig. 20.1: Effect of decrease or absence of exploitation on (a) aerial phytomass (green matter, GM, in kg/ha” ^)
and (b) percentage of dry matter, DM, contained in aerial phytomass.

Information sought by thematic specialists and managers is hence concerned with the following
aspects:
1. identification of vegetation species, vegetation groups and units or types of grasslands;
2. geographic distribution and estimation of areas;
3. quantity assessment of aerial phytomass (production and productivity);
348 Processing of Remote Sensing Data

4. quality evaluation of aerial phytomass;

5. facilitating prediction of dynamics of plant communities (spontaneous or induced by usage,
agricultural practices, variations in management, etc.).
The main difficulty in using remote sensing for such applications lies in the difference between
the parameters investigated and those measured by remote sensing. Hence, it is necessary to establish
relationships between various disciplines and the concerned typologies (Fig. 20.2). The latter are
based on additional observations and measurements and on records of acquisition of data different
from those of normal practice. These will be developed while discussing the applications.

Fig. 20.2: Model of use of remote sensing data for applications in grassland and pasture studies.

20.3 APPLICATIONS
20.3.1 Distribution of herbaceous formations and permanent
grasslands
Managers require information on distribution of herbaceous formations at a national level or at a more
local level.This information is essential for estimating surface variations overtime (see the beginning
of this chapter). In fact, this variation represents the impact of human activity, viz., clearing of forests
and creation of secondary grasslands in intertropical zones, abandoning permanent grasslands and
cultivation or afforestation in temperate zones, etc. These changes have a direct impact on climatic
 Applications 349

changes (global change) and quality of water resources (nitrate pollution in the case of conversion of
grasslands Into crops), as well as on soil erosion.
The objective of application of remote sensing in grassland studies consists of separating such
areas from other land-cover types. The European Union countries now are equipped with this
type of information thanks to CORINE Land-Cover data (see Chap. 19). However, the precision and
re-actualization of CORINE data are not necessarily suited to all applications investigated. Hence, we
present below a method of studying satellite data. An example on wetlands is given in Chap. 21.
Objective: Separation of herbaceous formations and grasslands from other land-cover types.
Principle: Selection of satellite data acquired on days when the herbaceous formations and
permanent grasslands exhibit very distinct spectral characteristics, i.e., specific physiognomic features.
Depending on the region, a single or multiple dates of acquisition will be necessary. Thus, in Lorraine,
a scene taken at the end of winter enabled identification of permanent grasslands and rangelands
from other land covers (Girard and Benoit, 1990), whereas in Ile-de-France a scene of end-winter and
another of spring were found necessary.
Methodology: Application of an unsupervised classification method (see Chap. 8) on the scene(s).
If accurate data on spectral characteristics of various land-cover types are available, supervised
classification can also be employed. The results can be refined by application of geographic or
radiometric masking (see Chap. 7).
Preparation of a map necessitates georeferencement of the results obtained (see Chap. 13).

20.3.2 Identification of vegetation species and groups

A vegetation group Is a collection of living species in a given environment. Some species are strictly
associated with well-defined media and others may occupy entirely different environments. If we accept
that vegetation groups are good indicators of environmental factors and that they react to modes of
exploitation or usage by man, we can constitute groups of indicator species for these factors (Fig.
20.2). It Is interesting to Identify such groups by remote sensing and thus indirectly acquire spatial
information on ecological factors.
A species is Identifiable by remote sensing if it exhibits on a given day a specific spectral
characteristic (reflectance values, digital numbers) enabling its identification. This spectral characteristic
is related to the phenological stage and physiological state of the species. In practice, unambiguous
identification of a particular species is rare. It is possible for some species and on limited days. For
example, in rough grazings, quasi-pure populations of Calluna vulgaris can be identified in satellite
data acquired in sprlng^ since this species has a spectral characteristic in the near infrared band (type
A, see Chap. 4, Fig. 4.6) entirely differing from that of neighbouring graminaceous species. The
graminaceous species Brachiaria dictyoneura can be distinguished from other forage and savannah
grass species using reflectance measurements made In a dry season (see infra: example of Llanos in
Colombia).
A vegetation group consisting of several species can be Identified due to specific physiognomy
corresponding to the combination of phenological stages and physiological states of the most abundant
species in the group. Again, this identification may be possible on a particular date and impossible on
another. Such identification assumes that the botanical composition of the group and an interpretation
model of spectral behaviour are available.

■ Spectral behaviour model

For multi-annual and multi specific vegetation with large coverage (usually > 70%), such as permanent
grasslands in temperate zones, various situations can be recognised: chlorophyllian, senescent or
florescent species. Other situations are the case (for example, after mowing) in which bare soil has a
350 Processing of Remote Sensing Data

small spectral contribution (sparse savannas or crops) or the case of large shadows. These various
behaviours are summarised in Fig. 20.3.
The spectral contribution of nonchlorophyllian plants In case 2 (Fig. 20.3) results in stronger
reflectance In the visible and middle infrared bands and weaker reflectance in the near infrared, relative
to case 1. A limited spectral contribution of soil (case 3) enhances the reflectance of vegetation in the
visible band and reduces it in the near Infrared. The moisture content of soil surface is assumed to be
close to that of vegetation; reflectance values in middle Infrared are similar. Pigments of many coloured
flowers (case 4) modify the reflectance only In the visible band and the spectral characteristic is
constant in the other two spectral domains. Lastly, contribution of shadows (case 5) drastically reduces
reflectance in the near and middle infrared bands. This effect is counterbalanced In the visible band by
increase in reflectance due to bare soil.

H Examples
Herbaceous formations and permanent grasslands are differentiated from crops due to the fact that
they are multiannual, multispecific and highly heterogeneous populations. As a matter of fact, temporal
succession of vegetal species according to their coverage and phenological stage adds up to the

1 to 4: 70 to 100% plant canopy

Senescent and chlorophyllian plants

Chlorophyllian plants with little Chlorophyllian plants and

bare soil visible coloured inflorescence

Crops with 30% bare soil,30% chlorophyllian plants and 40% shadows

Fig. 20.3: Types of spectral characteristics.

 Applications 351

effects of mode of exploitation. Spectral properties of vegetation covers in the visible and especially in
the near infrared band (see Chap. 4) enable differentiation of grasslands with dominance of
graminaceous species from grasslands with dominance of leguminous and dicotyledonous species. If
the spectral characteristic of a permanent grassland is taken in April (Fig. 20.4), rye-grass and colchicum
species dominate. In May, the yellow inflorescence of abundant Ranunculus increases the reflectance
values in the visible band and the abundance of standing green blqmass results in high values in the
near infrared. Mowing in June, which decreases the quantity of chlorophyllian aerial phytomass, leads
to lower reflectance values in the near infrared, whereas values in the visible band are close to those
of May. In July, vegetation dominated by Yorkshire fog again becomes chlorophyllian and dense: the
spectral characteristics are close to those of April.

__ July

__ June, mowed

__ May

._ April

Fig. 20.4: Ground reflectance measurements for permanent grasslands from April to July.

In a given region, knowledge of phonological stages and calender of various agricultural operations
is imperative (Table 20.3).
In the example given above, 4 satellite images acquired In end April, May-nJune, July and August-
September differentiated 7 out of 10 grassland types and only the types 7, 8 and 9 could not be
separated.
The large range of physiognomic variations created by combinations of states and modes of
exploitation enables identification of various herbaceous formations or grassland types, provided remote
sensing data acquired on precisely chosen days during a year are used. This is not always possible
due to frequent cloudiness of grassy regions. Lastly, the need for diachronic data enhances the financial
costs of study.
Identification of an IRC type colour composite or other combination of bands by visual interpretation
on computer monitor (or in printed hard copy) can be done for units of small extent (a few pixels).
Contrarlly, automated classification by supervised or unsupervised methods necessitates a larger
number of pixels (see Chap. 16). Since identification of vegetation species and groups requires an
accurate knowledge of the terrain, supervised methods are preferable over unsupervised ones (see
Chap. 8).

20.3.3 Mapping of units and estimation of areas

Geographic distribution of rangelands and permanent grasslands and estimation of corresponding
areas are necessary for determining the forage resources of a region.
352 Processing of Remote Sensing Data

Table 20.3: Example of calender of agricultural operations in permanent grasslands in Lorraine

(after Benoit et al., 1988)

Months Grassland
types
April May-June July August-September October

M ********
1
M M 2
-kicicicirkic-k-kic’kic-k*** ********
M 3
*★★ ★*** ** ** ★★★★** ★** ★** ********
M 4
M M 5
**************** *********
M* 6
★★*★**★★*★*★★*★* **★*****★★★★★★*★ **************** ********
7
** ★*★* *★★★* ★** ** **************** **************** ********
8
**************** ********
9
10

Recommended dates
of image acquisition

M: mowing, E: ensilage, ' grazing -no operation

An example of mapping the savannas and pastures in Llanos in Colombia illustrates the procedure
adopted. These savannas cover an area of 3.5 Mha and constitute environments for which relationships
between vegetation and ecological factors are not known. Land management in this region necessitates
a precise inventory of crops, natural vegetation formations and zones of degradation. Moreover, while
some natural plant communities and pastures have been accurately studied in highly localised sites,
generalisation and extension of the results obtained to large areas constitutes a problem for further
studies. The savannas and pastures in this region have been Investigated as follows:
— Spectral reflectance measurements were conducted on the ground with a SPOT simulation
Cimel radiometer (Girard and Rippstein, 1994) during January and late September to early October
1991. Simultaneously, height and density of plant cover, instantaneous production of aerial phytomass
and water content were measured in reference sites. The objective of these measurements was to
determine spectral characteristics of various savannas and fodder crops during dry and wet seasons,
respectively, and relate them to various plant parameters.
— Multispectral SPOT data were acquired in March (dry season) and September (wet season) of
1991. Meteorological constraints did not permit simultaneous acquisition of ground truth and satellite
data. It is generally difficult to obtain satellite and ground data simultaneously.
Ground truth measurement sites were used as reference zones for supervised classification of
SPOT data of each date. Moreover, statistical processing (analysis of variance, PCA, ARC, etc.) of the
ground truth data enabled establishing, for each period, groups of savannas and fodder crops
significantly differing in reflectance. These groups served as a basis for preparing the legend of the
map resulting from the two classifications (Table 20.4).
Relative estimate of areas Is given by the number of pixels corresponding to each class, on the
basis of a pixel of 400 m^ in area. An absolute estimation requires geometric correction and restitution
on a topographic base.
 Applications 353
Table 20.4: Legend for the map of savannas and pastures in Llaños of Altillanura (Colombia)

Highly exploited fodder Sparse savannas (degraded) Savannas on clay soils with
crops (degraded) Trachypogon vestitus
Fodder crops of <1 year Savannas on clay soils Savannas with Andropogon
gayanus
Fodder crops of >1 year Savannas on sandy soils Savannas with termite nests
Old fodder crops Savannas: burnt > 1 year Grassy lowlands
Savannas: burnt <1 year Shrubby lowlands

20.3.4 Aerial phytomass quantity estimation

Use of vegetation indices for estimating instantaneous production of aerial phytomass was mentioned
in Chap. 4. Such an evaluation for herbaceous formations and permanent grasslands is beset with
several difficulties:
— Difficulty in describing highly heterogeneous plant covers: erectophyll species (Gramineae,
graminoids) mixed with plagiophyll species (dicotyledons and the leguminous species);
— Very large quantity of chlorophyliian aerial phytomass for some grasses, resulting in saturation
of reflectance;
— Mixing of standing dry matter with green material for many types of plants, giving rise to variations
in vegetation indices (Fig. 20.5);
— Sparse plant covers (some natural herbaceous formations) with a spectral contribution of soil,
necessitating use of appropriate vegetation indices.

« 7 000

I
^ 6 000 o
5000 o
4 000
CO
E 3 000
o
“ 2 000 ♦ Green
1 000 O Green + dry
0
8 10
IR/R

Fig. 20.5: Effect of dry plants mixed with chlorophyliian vegetation on the ratio IR/R.

For this unexploited (measured In April to October) caldcóle rangeland (Fig. 20.5), a relationship
between IR/R and aerial phytomass could be established only for green vegetation areas.
Asymptotic relationships between leaf area index and aerial phytomass (Fig. 20.6), vegetation
index and leaf area index (Fig. 20.7) and vegetation index and aerial phytomass (Fig. 20.8) were
computed for a permanent pasture.
However, regular grazing of grass does not allow observation of large quantities of aerial phytomass
and reduces the validity of such a relationship (limited range of values).
Estimation of aerial phytomass production as a function of efficiency of interception of radiation,
determined from remote sensing (linear relation between vegetation index and efficiency), may be
envisaged. However, this assumes a temporal characterisation of ‘states of efficiency’ of the cover,
which is difficult to achieve for very heterogeneous natural or semi-natural groups.
354 Processing of Remote Sensing Data

Fig. 20.6: Relationship between leaf area index and aerial phytomass for a permanent pasture in
various stages of exploitation (/^ = 0.96).

Fig. 20.7: Relationship between leaf area index and IR/R for a permanent pasture {r^ = 0.95).

CO 3 000i
sz
O)
jsc : 2 500'

Û 2 000'
CO
CO 1 500'
CO
E
o 1 000' ♦
CÙ ♦
500-

0'
10 15 20 25
IR/R

Fig. 20.8: Relationship between IR/R and aerial phytomass for a permanent pasture ((/^ = 0.88).

Variation of fresh aerial phytomass (measured after extraction and weighing) for permanent
grasslands of Lorraine with NDVI values (see Chap. 4) computed from reflectance measurements on
the ground with a SPOT simulation Cimel radiometer Is shown in Fig. 20.9. 200 measurements were
made over the period from beginning-April to end-September for the years 1986 and 1987. A very
strong dispersion of values and an insignificant exponential fit can be observed.
Contrarily, after separation of the grasslands into various types according to modes of exploitation
and management (Benoit et al., 1988), the value of varies from 0.22 to 0.87 (Table 20.5) for an
exponential fit.
 Applications 355

Fig. 20.9: Relationship between NDVI and fresh aerial phytomass for various permanent grasslands.

Table 20.5: Correlation (/^ values) between vegetation index (obtained from ground Cimel radiometer
measurements) and aerial phytomass for seven types of permanent grasslands of Lorraine

Grassland type A6 A7 A8 A3 A1 A2 A9

Exponential fit for

phytomass 0.87 0.55 0.55 0.41 0.69 0 .2 2 0.56

Vegetation indices for estimating the quantity of aerial phytomass ought to be determined within
preliminarily defined grassland types.
Grasslands (more particularly A6) for which the relationship between the two variables Is strong
are those for which there Is no accumulation of standing dry nonchlorophyllian biomass due to regular
removal of grass. When grass removal Is large, i.e., 200 < UGB ha~^ yr^< 500 (A7, A8, A9 and A1) and
nitrogen fertilising more or less important, the range of measured biomass is relatively small, which
explains the rather low values of correlation coefficient. When grass removal is small (A2, A3), standing
dry matter accumulates and the value of becomes small.

20.3.5 Phytomass quality assessment

Evaluation of quality of phytomass, associated with identification of groups, is based on knowledge of
specific indices attributed to various vegetal species according to their pastoral or forage importance.
The 10- or 5-scale notation developed by Holland, New Zealand and French pastoral agronomists
enables computation of the pastoral value {PV) of grasslands according to a formula of the type (for
10-scale notation):

PV = E (C S ,x/S ,)/1 0

where CSj is the specific contribution of species / (frequency of occurrence or coverage) and ISjXhe
specific index of species /.
Pastoral values thus obtained must be considered only as a means of quality assessment and
comparison of grasslands. In fact, for the same vegetal species, the values of IS vary according to the
stage of development, animal race that consumes it, etc., whereas a single /S value is assigned to a
species once for all. These limitations notwithstanding, estimation of Pl/gives an approximate Idea of
the quality of biomass produced.
356 Processing of Remote Sensing Data

Types of perennial grasslands shown in Table 20.3, Identifiable by remote sensing, are botanically
defined by various combinations of the most abundant species. Mean P\/values of each type, computed
based on these species, differ significantly at 5% confidence limit. Hence, a spatial representation of
the quality of perennial grasslands can be obtained.
Another estimate of the quality of aerial phytomass is the water content of standing grass. Water
content can be related subsequently to the phenological stages. It serves as an indicator for fixing the
date of first mowing in the case of permanent grasslands targeted for agro-environmental measurements.
Various studies seem to indicate that water content can be estimated from the reflective middle infrared
data (Fig. 20.10).

80
A X
75
■
i A
X
l ì
e
<D
65
o
o

A Large height r^ = 0.64

■ Medium height r ^ = 0 .8 2
^ 55
X Small height r " = 0.72
50
7 9 11 13 15
NIR-MIR Digital numbers

Fig. 20.10: Relationship between NIR-MIR and water content for various permanent grasslands
(after Orth, 1996).

Water content of aerial phytomass of various marshes was determined using field samples and
double weighing method (fresh grass and after 48 h in a drying oven). A relationship could be established
between these values and the difference between ground measurements (NIR-MIR) of near- and
middle-infrared bands (CImel radiometer, SPOT-4 simulation).
The regression coefficient for all places of measurement is equal to 0.57, whereas it Is 0.82, 0.64
and 0.72 when grasslands are divided according to grass height. This shows the influence of the type
of canopy and the need to investigate correlation between factors within various grassland units.
An application of such results is in management of plots of permanent grasslands in marshes
under agro-environmental contracts. In fact, the date of first cutting is fixed according to the dates of
nesting of some bird species and applied to all plots. However, use and application of these contracts
would be facilitated if the date of first cutting were regulated. A map of marshy grasslands of Cotentin
and Bessin (Orth, 1996), prepared fromTM 3, 4, 5 bands of a LANDSATTM scene of 29 April 1990,
showed large differences in commencement of vegetation in end-winter. These differences, due to the
magnitude and duration of submersion, provide an objective basis for regulating the date of first cutting
depending on plots, according to their stage of growth.

20.3.6 Monitoring and forecasting of changes in vegetation

groups
Permanent grasslands almost everywhere in Western Europe and in Central Europe are subjected to
underexploitation or even abandoned, due to changes in agricultural practices. This results in increase
in dry matter and even in accumulation of standing dry matter, accompanied by changes in vegetation
groups: conversion from calcicole pastoral grasslands to rangelands dominated by Brachypodium
 Applications 357

pinnatum, gradually encroached by thorny and shrubby species (wild briars, blackthorns, hawthorns,
etc.). These modifications can be monitored by remote sensing (aerial photography or satellite data).
Thus, in a small zone of French Vexin (west of the Paris Basin), permanent grasslands and calcicole
rangelands could be identified in 1:25,000 aerial photos of 1968 and their status studied In 1:30,000
aerial photos of 1994. The following information could be inferred:
— certain level of maintenance of grassland areas in valleys (PGAC due to hydrological situation),
replacement of some of them by crops notwithstanding;
— large growth of underbrush in pastures and rangelands of mid- and high slopes;
— quasi-disappearance of grasslands in plateaus.
Based on these results of photo-interpretations, a map of grasslands and rangelands could be
prepared for western part of Ile-de-France using SPOT data of 1995. Comparison of this map with a
map of calcicole rangelands prepared In the 1970s revealed disappearance of a number of these
rangelands. This represents an adverse effect of agricultural community decisions. In fact, these
rangelands were ploughed just before compulsory set-aside land establishment: they were hence
counted in UAA and were left for natural colonisation (spontaneous fallow). Unfortunately, fresh growth
of rangelands is a slow process: ploughing these areas increases splitting of habitats of particular
fauna and flora, which are thus endangered.
Some fallow lands in the Paris region were also monitored by satellite data. For this, various types
of fallow lands were identified on the ground, resulting In preparation of the physiognomic typology
(bare soil, dominance of herbaceous species, mixture of grasses, shrubs and trees, etc.). Moreover,
position of these types In the dynamics of vegetation groups was studied.
A diachronic study was subsequently made using LANDSAT TM data of 30 April 1984 and SPOT
data of 24 April 1987 (Szujecka and Girard, 1990). A mask for eliminating land-cover types other than
various types of fallow lands was designed and applied on both scenes. Each scene was then subjected
to supervised classification into three classes, viz., young fallow land (sparse plant cover, bare soils
present at places), herbaceous fallow land and shrubby fallow land. After geographic restitution
accompanied by a resampling of pixels (imperative since the pixel size of TM is 30 x 30 m and that of
SPOT 20 X 20 m) the two scenes became comparable close to pixel level. An image showing changes
during 1984 to 1987 was prepared assigning a characteristic colour to each type of change. An example
of the observed changes is given in Table 20.6.

Table. 20.6: Example of changes in fallow lands In Ile-de-France region between 1984 and 1987 derived from
classification of satellite images

Classification In 1984 Classification in 1987 Interpretation

Non-fallow land Herbaceous fallow land Progressive change

Herbaceous fallow land Herbaceous fallow land Equilibrium within a stage
Shrubby fallow land Shrubby fallow land Equilibrium within a stage
Shrubby fallow land Young fallow land Regressive change: clearing before growth
Herbaceous fallow land Non-fallow land Change of land cover: construction of buildings

This example shows the role of remote sensing in land management, especially in the peripheries
of urban areas where disappearance of patches of rural space makes it difficult to monitor sites
characterised by multiplicity and geographic dispersion using ground data alone.

20.4 CONCLUSION
Remote sensing data provides valuable Information for management of land cover and especially
herbaceous formations and permanent grasslands. The most significant features are:
358 Processing of Remote Sensing Data

— capacity for spatialisation and generalisation of results acquired over limited areas;
— flexibility of processing digital data and possibility of analysing the results by geographic
information systems;
— complementary information provided by diachronic data to other reference sources.
On the other hand, their usage ought to take into consideration the following constraints:
— relatively recent existence of these data;
— difficulty in acquiring pertinent data in cloudy regions (see Chap. 26);
— necessity for collecting ground data, particularly more severe in precise thematic applications;
— information concerning mainly the upper part of the plant canopy due to bias of their spectral
characteristics;
— geometric resolution of data often Inappropriate for the study of phenomena under investigation.
Hence, development of classification methods based on the theory of fuzzy groups and artificial
intelligence seems very promising In this domain.

References
Benoit M, Girard C-M, de Vaubernier E. 1988. Comparaison du comportement spectral de prairies permanentes en
Lorraine avec leur type d’utilisation. Agronomie, 8:265-272.
Delpech R, Dume G, Galmiche R 1985. Typologie des stations forestières. Paris, Ministère de l’Agriculture, IDF,
243 pp.
IFEN. 1996. Les données de l’environnement: milieu, 25:4 pp.
Girard C-M, Benoit M. 1990. Méthode de cartographie des prairies permanentes: application à la Lorraine sur des
données SPOT. C.R. Acad. Sci., Paris, 310 (III): 461-464.
Girard C-M, Rippstein G. 1994. Utilisation de données SPOT HRV pour la cartographie de savanes et pâturages
dans les LIahos de Colombie. Bull. Soc. Française de Photogrammétrie et Télédétection, 133:11-19.
Orth. 1996. Typologie et caractérisation des prairies permanentes des marais du Cotentin, en vue de leur
cartographie, par télédétection satellitaire, pour une aide à leur gestion.Thèse INA-PG, 149 pp. plus annexes.
Szujecka W, Girard C-M. 1990. Cartographie et suivi diachronique des friches en Ile-de-France à partir de données
TM et SPOT. Photointerprétation, 90:1-3.
 _____________________^

Wetlands
A programme for investigating wetlands was launched in 1996 by the Ministry of Environment, water
agencies and a public interest group on hydrosystems (comprising the BRGM, CEMAGREF, CNRS,
IFREMER, INRA, ORSTOM, etc.). However, among the projects undertaken under this programme,
very few included remote sensing In their studies. We feel that it is useful to present some examples of
application of remote sensing for the study and management of wetlands.

21.1 NATURE AND IMPORTANCE OF WETLANDS

According to the law on water (No. 92.3 dated 3 January 1992), 'wetlands mean regions, exploited or
not, normally flooded or filled with fresh, saline or brine water permanently or temporarily; vegetation,
if it exists, is dominated in these zones by hydrophilous plants, at least during part of the yeaf. France
comprises about 3 Mha of wetlands distributed in the following main types:
— wetlands of alluvial valleys (examples: Loire valley, Rhine valley and Alsatian Reid, etc.);
— ^wetlands of inland plains (examples: Brenne, reservoirs and ponds of wet Champagne);
— regions rich in peat and/or small wetlands (examples: Jura, Morvan);
— Atlantic-English Channel littoral wetlands (examples: Bay of Mont-Saint Michel and marshes of
Bougeai, estuaries and marshes of the Seine, etc.);
— Mediterranean littoral wetlands (examples: Camargue, ponds and salt marshes of Languedoc,
etc.);
— Wetlands of DOM-TOM (example: marshes of Kaw In Guyana).
Implementing a combative policy particularly aggressive during the last 40 years, these wetlands
have been subjected to drainage and cultivation operations, which have greatly shrunk their area
(Interministerial committee on assessment of public policies, 1994). This has serious consequences
for natural heritage and economy of France. As a matter fact, wetlands play an important role In the
regulation of water flow and storage In watersheds and in the protection of water quality. They also
contribute to conservation of animal and plant diversity. Lastly, it has now become necessary to give
more attention to wetlands in land management. Various programs, such as Natura 2000, la directive
Habitat, the law of water accompanied by Schémas directeurs d’aménagement et de gestion des
eaux (SDAGE) and Schémas d’aménagement et de gestion des eaux (SAGE), are based on the
premise that the spatial distribution and nature of wetlands are known and that powerful tools for their
monitoring and assessment are available.
While many inventories, and specific studies are available, they generally pertain to zones of
limited areal extent, with no data about continuity between them, and provide no information on the
nature of habitats and functioning of non-inventoried zones. It is hence necessary to generate data
facilitating generalisation and spatial analysis of information acquired by other means and create
basic maps for locating changes produced in these zones and assessing their magnitude.
A few examples of application of remote sensing for some of these problems are described
below.
360 Processing of Remote Sensing Data

21.2 DELINEATION OF WETLANDS BY REMOTE SENSING

21 .2.1 Marshes
Before undertaking any complex studies in wetlands, it is necessary to delineate them. This application
is illustrated by an example pertaining to the marshes of Cotentin and Bessin, taken from a PhD (Orth,
1996).
The regional natural park of Marais, created in 1991, used the contours established in 1836 by
the Bas Ponds syndicate based on maximum floods observed. The corresponding boundary is not
very reliable and evidently obsolete, considering the management of these marshes for the last century.
Its delineation using satellite Images is possible only if correct Information is available about the nature
and functioning of marshes.
Thus, it was essential to know that these marshes resulted from filling of valleys about 10,000
years ago consecutive to a rise of sea level, followed by submersion under fresh water and major
erosion of soft formations (such as marls, clays) on slopes due to human activity. The marshes are
localised along many watercourses flowing into the English Channel at the foot of the Contentin isthmus.
These marshes constitute a group of digitate landforms, comprising a mosaic of large sized permanent
grasslands (one to several tens of hectares) flat, cut across by channels. Generally submerged during
winter, they form part of a conventional bocage landscape. The bocage, situated outside the marshes
in the ‘highland’, consists of permanent grasslands and crops, with medium size plots (2 to 8 ha)
outlined by a dense hedge system. In satellite Images marshes are continuous zones with homogeneous
texture and structure, whereas the ‘highland’ is a fragmented zone with a heterogeneous texture and
a cellular structure.
Few satellite data acquired during the period of submersion are available and an instantaneous
measurement of the latter is too punctuate to be representative of modes of waterlogging and to
facilitate tracing of marsh boundaries.
The marsh boundaries were initially determined by visual Interpretation of 1:100,000 paper displays
of IRC colour composites of 4 SPOT scenes (KJ 31 -250 and 32-250) of May 1990, May 1992, August
1993 and November 1990. Visual interpretation was based on colours, texture and structure of the
image (see Chap. 5). The effect of acquisition dates on delineation of boundaries was low (43 to 49%
marshes estimated in the same area depending on the date of acquisition). In this specific case,
autumn was found to be more favourable for visual delineation.
Moreover, the SPOT scene of November 1990, for which the hedge system was distinctly visible,
was subjected to automated classification. In order to evaluate differences in structure between marshes
and the bocage, a filtering technique (Sobel, see Chap. 12) was used on the first two PCA components
(see Chap. 7) of 3 bands.
A new plane was obtained by combining the two planes resulting from this filtering. The OASIS
classification (see Chap. 11), which takes the pixel neighbourhood into consideration, was applied on
this filtered plane using a 17 x 17 window and to bands b3 and b2.The reference nuclei were chosen
in marshes and the ‘highland’, respectively. The result thus obtained consisted of homogeneous map
zones of these two units. Attributing a zero value to the zones corresponding to ‘highland’, a mask was
prepared for selecting only marsh areas.
A comparison of the boundaries obtained by these two methods showed 62% boundaries classified
identically. The remaining 38% comprised partly valley bottom grasslands, correctly classified as
marshes by the automated classification whereas they were not identified in the visual interpretation,
and partly marshes , incorrectly classified as ‘highland’ by the automated method (22%). These were
the zones benefited by general water management and local agricultural development. These areas
also have a marshy structure but host grasslands close in quality to those of ‘highland’. The question
 Applications 361

of attributing these areas to marshes or not is beyond the scope of remote sensing data interpretation,
but concerns the very definition of the theme.

21.2.2 Flooded and wet grasslands

The objective here was to delineate flooded and wet grasslands, among others, of the Voire valley
situated in the Haute-Marne district of the wet Champagne region (Belluzzo, 1997). Two TM scenes
acquired in different seasons and for different crop years (1 April 1990 and 15 May 1992) were used
for this purpose. Demarcation of these grasslands would have been easier in two scenes acquired on
the same dates and during the same crop year. In fact, while the botanical composition of permanent
grasslands had no chance of varying between these two scenes, modes of exploitation of these areas
were likely to change. With respect to crops, the land cover for the same plot would be different in the
two scenes. Delineation of wet grasslands hence poses difficulties that would not have been there had
the two scenes acquired on the same dates and within the same year were available.
A few 1:17,000 panchromatic aerial photos of the Voire valley region acquired by the IGN were
used for verification of boundaries and classification. The work was carried out not on the entire scene
but on an image segment of 1500 x 1284 pixels, for each scene. The image of May was geometrically
restored using that of April in such a way that the two become comparable (see Chap. 13). The Image
segment comprises a portion of chalky Champagne in the north-west corner.This region is characterised
by a chalky plateau, with large size open-field plots occupied by Intensive annual crops, among which
alfalfa crops are sometimes observed. Contrarily, the wet Champagne region mainly consists of clayey
substratum, covered by smaller plots, scattered valleys and moist depressions mainly occupied by
grasslands and forests with no plots clearly visible.

■ Method of study
Before applying any processing technique, it was necessary to study the characteristics of the bands
of the two images for determining the means, standard deviations and coefficients of correlation with
other bands (Table 21.1). Use of bands pertaining to various uncorrelated spectral domains with high

Table 2 1 .1 : Data characteristics of bands for two images (bold: high correlation coefficients)

1 April 1990 15 May 1992

Correlation matrix Correlation matrix

TM1 TM2 TM3 TM4 TM5 TM7 TM1 TM2 TM3 TM4 TM5 TM7
TM1 1 TM1 1

TM2 0.947 1 TM2 0.939 1

TM3 8.832 0.947 1 TM3 0.847 0.949 1

TM4 0.461 0.411 0.217 1 TM4 0.390 0.300 0.092 1

TM5 0.783 0.876 0.905 0.384 1 TM5 0.811 0.854 0.879 0.292 1

TM7 0.705 0.846 0.945 0.124 0.938 1 TM7 0.748 0.842 0.934 0.009 0.933 1

Mean 63.03 27.69 28.60 53.95 54.58 24.47 Mean 66.37 31.20 30.41 85.64 59.15 24.51
Stand- 14.34 7.95 11.85 20.54 21.70 14.23 Stand­ 16.30 9.98 15.04 26.79 25.14 17.52
ard ard
deviation deviation
362 Processing of Remote Sensing Data

standard deviations provides more information than the inverse case. Bands TM 3 ,4 and 5, correlated
least with each other, were used for classification of grasslands. Band TM 7, less correlated with TM 4
than TM 5 but having a lower dynamic range, was not used for classification but employed for detecting
boundaries of agricultural fields. BandTM 2 was chosen for preparing IRC colour composites and the
results of their interpretation used for verification of classification results.
A mask for isolating the grasslands was applied as shown in Fig. 21.1.

T ‘ phase 2"“ phase S'*' phase

4 phase Analysis of masking accuracy

Fig. 21.1: Flow chart of mask for isolating perennial grasslands.

H Geographic masking: first stage

Manual delineation of a geographic mask is that much better Inasmuch as it is based on distinct
differentiation criteria. The objective here is to identify the boundaries of agricultural plots between
chalky Champagne and moist Champagne. To make the boundaries more apparent, a Laplacian type
8 filter (see Chap. 12) was applied on bands TM3 and TM7 of the images of April. For these two bands,
low digital numbers corresponding to permanent grasslands (high chlorophyllian activity and water
content) distinctly revealed the agricultural boundaries.
The new bands thus generated were used to prepare a three-colour composite in which zones
corresponding to these distinct agricultural fields were manually demarcated. This made it possible to
remove the entire chalky Champagne and also to isolate and eliminate cultivated areas in wet
Champagne. Evidently, the first mask could not eliminate all cultivated plots, for example, those planted
with a more or less chlorophyllian crop side by side with grassland plots, also chlorophyllian; the
boundary between these two is not revealed. Moreover, all forest areas still persist.
 Applications 363

■ Mask derived from an unsupervised classification: second stage

It was then necessary to eliminate land cover types other than grasslands, based on digital numbers.
To differentiate between statistically differing units, an ascendant hierarchic classification (AHC, see
Chap. 8) that uses a Euclidean distance was applied separately on the two image segments (for
bands TM4, 3, 2) after application of the mask prepared earlier.
The dendrogram was divided into two classes: a class comprising chlorophyllian canopies and a
class consisting of other non-chlorophyllian land cover types (verified with an infrared colour composite).
In April, the first class includes grasslands and some standing crops and in May, it corresponds to the
same land cover types with the addition of forests.
Diachronic study was accomplished by means of the ‘krosscolour’ function which combines two
coded bands (Monget and Roberston, 1992). Two-dimensional histogram analysis facilitated
investigation of relationships existing between the types of objects and set of vectors associated with
corresponding pixels. These values were represented as a cluster of coloured points in two-dimensional
space, with the separately measured values of the two bands as co-ordinate axes. Agricultural plots
not eliminated by the preceding graphic mask could thus be identified. These fields were chlorophyllian
on one date and non-chlorophyllian on the other, whereas the permanent grasslands were chlorophyllian
on both dates. Predominantly broad-leaved forests could likewise be detected. The resultant classified
band, which does not conserve the objects identified as crops and forests, is used to apply a second
mask.

■ Mask derived from thresholding of bands: third stage

The preceding mask does not eliminate all the crops but still retains chlorophyllian crops In April and
a few non-chlorophyllian ones In May. Criteria that ensure their differentiation vis-à-vis permanent
grasslands need to be Identified.
Crops are annual monospecies populations subjected to a specific type of maintenance, whereas
permanent grasslands are multiannual and multispecies plant communities, with a possible
accumulation of standing dry matter, subjected to varying modes of management year by year (see
Chap. 20). Chlorophyllian activity analysed In TM3 and the equivalent water thickness (water content
X aerial biomass), a parameter Identifiable in the reflective middle infrared band (Gao and Goetz,
1995) and analysed from TM5, are the pertinent criteria for differentiation. Crops in May show a higher
chlorophyllian activity (low digital numbers in TM3) and larger equivalent water thickness (low digital
values In TM5) than grasslands with small variations in these bands, whereas the large diversity of
modes of maintenance of grasslands gives rise to a large variance In their values.
After application of the preceding masks, bands TM3 and TM5 of both dates were separately
segmented into three classes: high, medium and low values of digital numbers. Threshold values of
the three classes were selected such that the corresponding zones are in the best possible agreement
with the boundaries of plots Identified in the aerial photos. Comparison of the results of the two dates
using the ‘krosscolour’ program enabled identification of the last crop plots and creation of a mask for
uniquely isolating the permanent grasslands (Fig. 21.2).

■ Mask accuracy assessment: fourth stage

As aerial photographic data was available only for 1990, accuracy of masking and grassland
classification could be assessed only for that year. The accuracy assessment consisted of a systematic
sampling of aerial photos at 72 points with a 4 x 4 cm grid (680 x 680 m on the ground). Since these
photos were earlier used only for verification of plot boundaries but not for any Information on their
content, their use for such analysis was considered valid. The two classes retained for this analysis
were the mask and the grasslands.The resultant error matrix between the aerial photo and the classified
image is given in Table 21.2.
364 Processing of Remote Sensing Data

Table 21.2: Error matrix between aerial photo and classified image

Classified image
Mask Grasslands Total % correct

Mask 42 6 48 87.5
Grasslands 0 24 24 100
Photo Total 42 30 72
% commission 0 20
errors

Fig. 21.2: Final mask isolating permanent grasslands.

The small number of control points warrants caution in commenting on this table. A high Kappa
value, 0.86, was obtained (see Chap. 17). The mask used for isolating permanent grasslands seems
to be reliable: 87.5% correct with 0% commission errors. Contrarily, while 100% of the grasslands
existing in the photos were identified, a 20% commission error was observed. This would have an
influence on the geographic accuracy of grassland mapping carried out subsequently.

21.3 CLASSIFICATION AND MAPPING OF WETLANDS

The typology of continental wetlands according to their organisational level can be described as
follows (Table 21.3). Each class of a given level is made up of a mosaic of lower-level units. For the
local level, the CORINE Biotopes map has been used as reference, retaining it at level 2 for legibility.
Some examples are given to illustrate the mapping methods employed depending on the geographic
level of organisation.
 Applications 365

21.3.1 Mapping at national or regional level

Major types of wetlands (national level) will obviously be mapped at small scale. As they are defined
mainly from geographic criteria, this mapping can be envisaged only under GIS, comparing the altitude
and slope data obtained from a digital elevation model with the satellite data classified according to
land cover types. This is illustrated by an example of classification of a mosaic of LANDSAT TM
images (for covering the entire region) of the Champagne-Ardenne region. The land-cover types used
as criteria for differentiation of wetlands are based on the definition given In the Introduction. They are
Identified from an unsupervised classification (AHC) according to spectral characteristics (see Chap.
7). These land-cover classes are then compared with altitude and slope data (altitudes ranging from
70 to 100 m, slopes less than 1%).
Wetlands are often associated with permanent or temporary presence of water on or near the
surface. Water is readily discerned from digital numbers (especially in near and middle reflective
infrared bands) In satellite images provided the dimensions of water bodies are greater than those of
pixels. Lastly, artificial water bodies (gravel pits, ballast pits, reservoirs) are distinguished by their
rectangular forms and proximity to highly reflective mineralised surfaces, in addition to their spectral
characteristics. Identification of water tables is useful for locating flooded and waterlogged zones by
the criterion of ‘proximity to water’. This criterion can be used under GIS by means of a ‘buffer’ whose
dimension is chosen according to the phenomena to be detected.
Wetlands associated with watercourses and water expanses are characterised by proximity to
water tables, a fairly homogeneous continuous pattern depending on valley shapes and a varied land
cover consisting of peat, fen, forest and herbaceous vegetation, crops and bare soils.
Exploited peat bogs not covered by vegetation can be identified in any season due to their low
values of digital numbers. Vegetation groups of peat, sedge, reeds, etc., can be differentiated from

Table 21.3: Wet zone units of various levels

National level (*) Regional level Local level (extract of CORINE Biotopes)

Wetlands associated with 2.1 Lagoons

a given water expanse Water bodies 2.2 Standing fresh water lakes, ponds, pools
Pond regions Poplar plantations 2.3 Standing brackish and salt water lakes,
ponds and pools
2.4 Running waters

Marshes and wet moors Riverine forests and 3.1 Temperate heath and scrub
of plains permanent grasslands 3.7 Humid grasslands and tall herb communities

Slope wetlands Peat moors, 4.4 Temperate riverine and swamp forests and brush
wooded moors 5.1 Raised bogs

Mixed wet plains Hydromorphic bare 5.3 Water-fringe vegetation

associated with soils 5.4 Fens, transition mires and springs
watercourses
Wetlands of water­ Mosaic of crops and 8.1 Improved grasslands
courses and wooded wet soils 8.9 Industrial lagoons and reservoirs, canals
borders
Mosaic of grasslands
and crops
Permanent grasslands

(*)CIEPP typology (1994)

366 Processing of Remote Sensing Data

grasslands and crops by their very large heterogeneity (combination of dry and more or less
chlorophyllian vegetation) represented by large variation in digital numbers. They are also characterised
in spring by delay in growth (colder microclimate), i.e., non-chlorophyllian spectral characteristic.
Alluvial wet forests (riverine forests, poplar plantations, alder and ash trees) are difficult to distinguish
from other broad-leaved forests, except by a delay in spring vegetation. They can be identified only by
a combination of altitude and slope criteria. The same is true of grazed or mowed grasslands that are
more or less moist and intensively exploited, as well as cultivated drained zones. When bare soils are
visible, they are most often of dark colour (low value of digital numbers).
The wetlands of the Champagne-Ardenne region were mapped on this basis by regrouping various
land-cover classes. Two categories were distinguished according to the degree of submersion and
waterlogging. An example of constituent elements of the wetlands of the Saint-Gond marsh is given in
Table 21.4.

Table 21.4: Wetland units identified in satellite images for the Saint-Gond marsh

National level Regional level Local level

Very wet More or less wet

Ponds
Marshes and wet moors Wet marshes and moors Peat bogs
of plains (Saint- Alluvial Sedge-fens
Gond marsh) forests and More or less wet grazed or Reed-fens
thickets mowed grasslands Rushes
Riverine forest Mesohydrophilous Shortgrass
Flooded grasslands grasslands, crops and Ash trees
Hydrophilous grasslands broad-leaved trees Poplar plantations
Very wet grasslands Willows
Very humid bare soils Alders
Grasslands

Boundaries thus obtained surround the NZEFFP (only wet environments), SPZ and IZBC
boundaries of the BRIDGE database, as well as those of inland wet zones (inland swamps and peat)
of the CORINE Land-Cover database (see Chap. 19). As a matter of fact, the definition of these units
is more restrictive than that used in our method. This verification shows the importance of such a
methodology for small-scale spatialisation of medium-scale limited data (1:100,000 accuracy for
CORINE Land Cover).

21.3.2 Mapping at local level

■ Marshes
It is important to know to what accuracy can remote sensing data be used for identifying and mapping
wetland units. The rightmost column in Table 21.4 consists of the units classified by a supervised
classification method using a LANDSAT TM subscene of 15 May 1992 in the Saint-Gond marsh
region. Classes and their names were chosen based on a (1:10,000) habitat map of these marshes
(CPN Champagne-Ardenne, 1996).The accuracy of results varied depending on the units: the ground-
truth data, albeit very precise, was not compatible with the resolution of satellite data and made
delineation of training polygons difficult. Causes of confusion between classes were varied:

IN z e f FI: Natural Zone of Ecological, Faunal and Floristic Importance; SPZ: Special Protection Zone; IZBC: Important
Zone for Bird Conservation.
 Applications 367

— heterogeneity of plant communities: this was mainly the case for sedge-fens and peat bogs
which are confused with sedge-fens and when they are wooded, with thickets;
— confusion with physiognomically similar plant communities: such was the case with reed-fens
and rushes, grasslands and shortgrass and willows and poplar plantations.
These confusions can be partly reduced If a scene acquired on another date is used to take Into
consideration the phenological offsets between certain communities. In such a detailed study, aerial
photos enabled mapping of various units; however, this necessitates ground-truth verification, on the
one hand, for establishing reference data for visual Interpretation of vegetation communities and, on
the other, for assessment of Interpretation accuracy. For this reason, costs of mapping plant communities
at the local level are necessarily high, requiring purchase of aerial photos, time for interpretation,
ground check up, etc.

■ Wet and flooded grasslands

The example of grasslands of the Voire region (whose delineation was discussed earlier) illustrates a
classification based on diachronic data. The Image segments of April and May, masked to retain only
perennial grasslands, were classified separately in order to avoid radiometric corrections that introduce
a bias in digital numbers of pixels (see Chap. 16).The method consisted of a supervised (parallelepiped)
classification by thresholding bands T M 3,4 ,5 in three or four levels using the spectral characteristics
model of grasslands (see Chap. 20). Threshold values were chosen such that they determine zones
conformable with the shapes of grassland plots identified In aerial photos.
Three classes were defined: (i) least chlorophyllian grasslands with low levels of biomass quantity,
chlorophyllian activity and water content, (ii) highly chlorophyllian grasslands with large values of all
these parameters and (Hi) moderately chlorophyllian grasslands with intermediate values (Fig. 21.3).
For the May data, one more class was added, viz., low-biomass grasslands with moderate values of
chlorophyllian activity and water content, corresponding to grasslands that were probably subjected to
an early ensilage operation.
The preceding classifications need to be superposed over one another for preparing a map and
legend of permanent grasslands. The legend should also take into consideration variations in the
physiological state of grassland classes (Table 21.5). It should also incorporate the possibilities and
duration of floods which affect the flora— predominance of rushes and sedges (Carex)— and the speed
of recovery of chlorophyllian activity in spring. It may only represent an instantaneous effect of exploitation

Physiological state Physiological state

TM4
TM4 Biomass

April 1990 May 1992

Fig. 21.3: Classification of permanent grasslands according to their physiological state, using manual
thresholding of bands TM3,4, 5 of April 1990 and May 1992.
368 Processing of Remote Sensing Data

Table 21.5: Legend for diachronic classification of permanent grasslands using variations in their
physiological state

April
Very low Medium High

Very low Reed-fens or flooded Wet grasslands, not Well-drained grasslands

grasslands with flooded ensiled since a more or
predominance of Carex less long time
Low
May Medium Grasslands flooded Grazed grasslands Grazed grasslands
for a more or less with rushes and well-drained mowed
High long time grasslands

modes: only the grasslands cut before the acquisition of the scene will appear with a low physiological
state and those mown one day later have strong vegetative growth. Some specific species populations
such as reeds or sedges have a spectral characteristic that facilitates their identification.
Flooded grasslands are mainly found In the Voire valley whereas reed-fens are essentially localised
in the borders of lakes in the Orient forest and many ponds of wet Champagne, as well as in the Aube
valley. A field survey conducted subsequent to this classification in the south-western zone, used as
reference for classification, revealed the accuracy of mapping (Table 21.6). The number of reference

Table 21.6: Accuracy assessment of grassland mapping (In number of plots)

Classification

Reed-fens Flooded Grazed Non- Grazed Ensiled Total %

or flooded grass­ grass­ flooded and grass­ omission
grasslands lands lands with wet grass­ mowed lands errors
with sedges rushes lands grasslands

Reed-fens or 8 8 0
flooded grass­
lands with sedges
Flooded grasslands 5 4 9 45
Grazed 9 2 11 18
grasslands
3
V. with rushes
■D
C Non-flooded wet 3 7 10 30
3
2 grasslands
0
Grazed and 4 3 48 7 62 23
mowed grasslands
Ensiled grasslands 2 10 12 17
Others 7 20 27

Total 15 32 9 16 50 17 139
% correct 53 16 100 44 96 59
% commission 47 84 0 56 4 41
errors
 Applications 369

fields varied depending on the classes (see Chap. 17). About 19% pixels classified as grasslands
corresponded to other land-cover types: sometimes crops, often a combination of grasslands, road
edges, trees, etc.This corresponds to 20% commission errors pertaining to the mask used for isolating
grasslands (Table 21.2). They represent ‘mixels’ due to the 30 m x 30 m resolution of LANDSATTM
which combines the little or non-chlorophylllan objects In April and the more or less chlorophyllian
ones in May. They are hence confused with classes ‘reed-fens’, ‘flooded grasslands with sedges’, on
the one hand, and ‘flooded grasslands’, on the other.
Errors vary depending on the classes. The ‘grazed grasslands with rushes’ are correctly classified
(0% commission errors and 18% omission errors) and ‘grazed and mowed grasslands’ somewhat less
correctly (4% commissions, 23% omissions). ‘Non-flooded wet grasslands’, contrarily, seem to be
incorrectly classified (56% commissions, 30% omissions) but more control points are needed to assess
the accuracy. For ‘ensiled grasslands’ (41% commissions, 17% omissions), the mediocre quality of
results is explained by the gap between the dates of acquisition (1990 and 1992) and the date of
verification (1997). In fact, decision to ensile a field or not varies considerably from year to year depending
on the needs of farmers.
The accuracy of local level mapping of wet and flooded grasslands can be improved by using the
reflective middle infrared data of higher geometric resolution such as those of SPOT-4 (see Chap. 2)
or infrared colour aerial photos.
Other examples pertaining to marsh grasslands are given In Chap. 20.

21.4 STUDY OF PLANT COMMUNITY DYNAMICS

In large river valleys (such as Loire), fluvial dynamics (modification of the major course due to erosion
of banks, water leakage, cutting of meanders, etc.) influence the dynamics of plant communities.
Remote sensing data can be used for several objectives in this case^ :
— Mapping variations in fluvial geomorphology and land cover of some sites based on visual
interpretation of aerial photos (1:25,000 to 1:30,000). This map will be useful in selecting profiles of
phytoecological sampling and their positioning;
— Mapping natural environments and spatialisation of functional units using satellite data.
These examples clearly illustrate the importance of remote sensing for local as well as regional
studies. Moreover, these data are equally useful at the beginning and at the end of an investigation, as
the modes of processing and Interpretation are chosen according to the objectives of study.

21.5 CONCLUSION
The examples of wetlands demonstrate the need for correct ground truth data about botanical
composition and phenological stages of plant communities, on the one hand, and practices and modes
of exploitation by man, on the other. Moreover, information furnished by remote sensing constitutes
the basis for selection of sites and preparation of sampling plans, for generalisation and spatialisation
of field data and for study of dynamics.
These examples also indicate the need to adapt, as far as possible, remote-sensing data to the
objectives pursued (see Chap. 16).The data available today can be used for applications ranging from
local to regional or national level studies and from small to large-scale Investigations.

^Project ‘Determination of free space for the Ligerian fluvial system: identification and spatialisation of functional
morphodynamic and ecological units in open and impounded valleys of Loire— social concerns and factors’ as a
part of the National Program of Research of Wetlands.
370 Processing of Remote Sensing Data

Since wetlands are venues for complex interactions between static and dynamic phenomena,
involving biotic, abiotic, natural and anthropogenic factors, applications concerning them essentially
necessitate use of geographic information systems facilitating management, comparison and synthesis
of data of various sources and periods.

References
Belluzo G, Girard C-M. 1997. Identification et classification de l’occupation du sol à partir de scènes Thematic
Mapper: application aux prairies d’une région de Champagne humide. Bull. SFPT, 146:22-32.
Comité Interministériel de l’évaluation des politiques publiques. Premier Ministre, Commissariat général du Plan.
1994. Les zones humides: rapport d’évaluation. La Documentation Française, Paris, 391 pp.
Conservatoire du patrimoine naturel de Champagne-Ardenne. 1996. Cartographie des habitats naturels et des
espèces de la Directive Habitats-Marais de Saint-Gond.
Gao BC, Goetz AFH. 1995. Retrieval of equivalent water thickness and Information related to biochemical components
of vegetation canopies from AVIRIS data. Remote Sensing of Environmnent, 52:155-162.
Monget J-M, Robertson YC. 1992. Two-variable mapping applied to remote sensing data interpretation: a software
implementation. Remote Sensing from Research to Operation. Proc. 18th Annual Conf.The Remote Sensing
Society, pp. 571-580.
Orth D. 1996. Typologies et caractérisation des prairies permanentes des marais du Cotentin, en vue de leur
cartographie par télédétection satellitale, pour une aide à leur gestion. Thèse INA-PG, 150 pp.
 22
Crop Inventory

22.1 ECONOMIC ASPECTS

Since the launching of the first LANDSAT satellite in 1972, civil remote sensing has been extensively
applied for crop inventories. The American research programs LACIE (Large Area Crop Inventory
Experiment) and Agristars v^exe devoted to this problem. In the European context, the most important
program of application of spatial remote sensing is MARS: Monitoring Agriculture with Remote Sensing.
Crop yield assessment (Meyer-Roux and Girard, 1997) Is, in fact, a very important economic
operation and its procedures are difficult to control owing to the inherent variability of yields due to
human but more often, meteorological factors.
In the United States, the emphasis has been on estimation of crop yields of foreign countries.
Within the European Union, requirements are related to controlling production, far exceeding
consumption, leading to high costs for the Community’s agricultural policy. A number of mechanisms
such as land management are set up. Their efficiency mainly depends on reliable forecasting of yields
and stringent controls.
During the period 1986-1992, the cereal production, considered maximum, was 165 Mt for the
European Community of twelve nations. This yield has been exceeded since then (Fig. 22.1). An
earliest possible forecast enables estimation of expenses and eventually corrective measures for the
following season. It is Important to note that collecting of grain is carried out gradually after harvesting
due to storage In farms and that effective statistics for market operations are thus known very late. The
low yields in 1994 and 1995 associated with low global stocks led to a mandatory decrease in set-
aside lands in the European Union in 1995 and 1996.

Million tons

Years

Fig. 22.1: Variation of cereal production in the twelve European nations from 1976 to 1996. Starting from 1990,
production in East Germany is included (source Eurostat).
372 Processing of Remote Sensing Data

22.2 OPERATIONAL USE OF SPATIAL REMOTE SENSING

Spatial remote sensing is used operationally for assessment of foreign crop yields by USDA (United
States Department of Agriculture) and by private companies who sell their services essentially for
speculation on market prices. The Earth Satellite Corporation with its Cropcasf service is widely known.
The European Community also uses remote sensing operationally to obtain neutral information,
independent of agricultural lobbies or member states. The Community Research Centre (ISPRA)
supplies monthly information whereas the Direction General of Agriculture finances operational
subcontracts such as satellite image analysis.
Remote sensing, however, remains marginal among the various techniques used for such
estimations for several reasons:
— It only partly covers the required Information on agriculture and a simultaneous inventory of
animal and vegetation products and economic structure is often preferable;
— Remote sensing does not identify crops, and hence cannot provide Information on their areal
extents, except when they are well in place, whereas a cultivator can foresee his intentions and give
precise information starting from the sowing stage;
““ Remote sensing can contribute to forecasting yields but cannot really give their exact values;
conventional methods are hence needed to obtain definitive results and often such forecasts are
considered a luxury not strictly necessary;
— Costs of enquiries with cultivators are very low, especially when they are made by
correspondence. The real cost corresponds to the time spent by the cultivator but it is not billed whereas
remote sensing costs are;
— Many technical limitations hinder an efficient usage of remote sensing. For example, in Europe,
cereal crops reveal high aerial biomass and vegetation indices for very dense chlorophyllian canopies
get saturated.
Remote sensing Is hence mainly used when conventional techniques cannot be applied or when
they are difficult to employ harmoniously, as In Europe. Their more or less extensive usage thus
depends essentially on the quality of results by remote sensing compared to the corresponding cost
and quality of conventional methods.
Operationally, remote sensing is employed mainly in the United States for crop yield assessments
in areas where such yields are generally low, since in this case remote sensing information is significant
as vegetation Indices are not saturated.
The European Commission uses remote sensing essentially for estimation of areas and indirectly
for verification of agrometeorological models of yield prediction. Direct application for yields Is still in
the research and development stage In the European conditions of high yield and sensors and satellites
presently available.

22.3 AREAL ESTIMATION IN EUROPE BY REMOTE SENSING

The method employed in project MARS consisted of representative sampling of 60 sites, 40 x 40 km
in area, for the European Union (Fig. 22.2). A series of SPOT images, about 3 to 4 per year, were
interpreted to extract the desired information.
For estimations conducted during a year no ground-truth information Is needed. However,
retroactively some ground-truth data enable accuracy assessment of computer aided visual
interpretation.
Choice of sampling sites was not aimed at achieving national representation or obtaining a reliable
absolute estimation. It constituted an early variation from the preceding year, as the data of the preceding
year were obtained through Eurostat using conventional statistics. A number of constraints such as
overlapping with the SPOT grid or an adequate percentage of arable land, led to the choice of a
systematic type of sampling rather than random.
 Applications 373

Fig. 22.2: Location map of 60 sampling sites in Europe.

Setting up of operational procedures enabled reduction of delay between Image acquisition and
its availability to the interpreter (3 days). In practice, any image taken before 20^^ of a month is included
In estimations at the end of the month.
Visual interpretation was carried out over randomly selected segments (about 50 ha in size) of
sites in the image. Interpretation results of a segment were then extended to the entire site by automated
classification. These site segments were used for ground-truth verification at the end of the year.
A fairly complex model of statistical Interpretation was employed to take into consideration the
fact that during certain early stages many crops cannot be distinguished and that some sites may not
have Images. This model uses relative percentages of crops recorded in the preceding year and, in
the case of lack of images, the results obtained from adjoining images of the same relatively
homogeneous soil strata (Dallemand and Vossen, 1995).
The quality of results varies depending on the years and crops. Since 1992, when the method
was developed, the results are better than those of conventional methods up to the end of August or
September, afterwards data received from member states are better. Reference data generally
corresponded to Eurostat statistics, but for two years the data Issued by CCR, viz., areas and yields,
were supplied to member states who used them In their own estimations and then sent them to
Eurostat; hence, the methods are no longer independent.
The example presented below (Fig. 22.3) gives, for the year 1992 the results of this method for
wheat compared to two datasets for 8 March and 14 October supplied by Eurostat. The results were
especially interesting for the period 1992-1994 when changes in arable land were very high following
374 Processing of Remote Sensing Data

Eurostat 8 March 1992 Eurostat 14 oct 1992

Action 4 Agrom et

Fig. 22.3: Estimation of variation in wheat area in Europe during 1992/1991.

ground freezing .These limitations notwithstanding, the method was the only tool available for estimation
of pre-existing agricultural fallow lands before setting in of set-aside lands In the Iberian Peninsula,
The variations recorded subsequently are of the same order of magnitude as is the precision of the
method, viz., about 2% variation from one year to another for important crops.
The site Investigation technique has been extended since 1997 to the central European countries
under the program PHARE and project MERA, ‘MARS and Environmental Related Applications’, and
the SCOT society conducted a preliminary study for extending such a method to Russia.
Attempts are in progress to modify this system facilitating national representation to the most
important countries. However, its cost is already high, about 1.6 million Euros per year for the European
Union of 15 nations and cannot be increased significantly to achieve these new objectives.
The same method is applicable for verification of agriculturists’ statements. In this case sites are
not selected statistically but by the member states depending on the zones to be verified. The method
provides a check of validity of agriculturists’ statements vis-à-vis the results of image interpretation.
Land registration and reference maps of plots necessitating subsidies are used for comparison between
image and statement. Positive validation by remote sensing is judiciously acknowledged by community
leaders and negative results are subsequently validated by ground-truth verification. Remote sensing
thus provides a preliminary sorting on computer monitor and limits field Investigations to 1% of the
statements, instead of 5% required in the conventional method.

22.4 AGROMETEOROLOGICAL MODELS OF YIELD

ESTIMATION
22.4.1 Yield estim ation methods
There are many methods of crop-yield prediction; some are operational and some in research stage.
The most common operational methods are statistical or regression methods. In practice, they consist
 Applications 375

of a statistical relationship between objective observations In the growing period of crops, i.e., before
harvesting, and the final yield. This relationship Is applied to the current year.
Observations may be related to crops (number of plants and heads, length of head, etc.) or to
meteorological nature. In the latter case, decennial or monthly variations of meteorological parameters
are used. In general, these methods are based on analysis of trends while objective field measurements
incorporate technological changes. Statistical correlations based on field observations are more precise
but cannot be routinely employed in Europe because too expensive.
Results of agrometeorological statistical methods are considered deceptive. A rigorous analysis
of the ‘AGROMET’ model of ‘Eurostat’, used for more than ten years, showed that significant correlation
between meteorological variables and yields cannot be obtained, once trends are deduced (Anonyme,
1997).
Crop-functioning models are equally operational but only in controlled conditions when many
input parameters are measured. They are generally more useful for studies such as investigation of
new crop varieties, resistance to cold, some diseases, increasing yield potential, etc., than for yield
prediction, which is only one of the output parameters. The method used by the project MARS is
based on a growth model (Vossen and Rijks, 1995), but uses it only for defining stress that is
subsequently included in a statistical regression for final yield. Thus it forms a mixed model.

22.4.2 MARS method or CGMS model

The CGMS (Crop Growth Monitoring System) model is based on a large geocoded database covering
all of Europe, whose parameters determine operation of a growth model for the important annual
crops of Europe using meteorological data as input. This database essentially comprises crop
characteristics, dates of sowing, harvesting, water needs, temperature constraints, etc., and soil data.
The former are essential for the crop growth model and the latter for the hydrological model. Major
work on this model consisted of creating this homogeneous database. The soil map, for example, took
more than 8 years of study by the Soil Map Service of INRA of Orleans (King et al., 1994; King et al.,
1995).
Usage of the method is associated with constraints of a geographic information system In which
diverse types of data coexist: data related to geographic location such as meteorological information,
natural boundaries as in soil maps or some agronomic data and administrative boundaries connected
with crop varieties or yields. A fictitious meteorological station is created by interpolation along a 50
km X 50 km square grid (module 1) and a crop growth model is applied to homogeneous soil zones
inside this square with a time interval of a day (module 2). These models, producing varying output
data, are then combined according to administrative zones for which previous information on yield
estimation exists, leading to prediction of yield (module 3) (Fig. 22.4).
Output data of module 2, such as delayed or early growth of vegetation, foliar coverage, can be
related to remote sensing data for verification of functioning of the model (Dallemand and Vossen,
1995).

22.4.3 Remote sensing as a complement to models

Crop yield data operationally supplied to the Direction General VI at Brussels are not a direct output of
the model. These raw data are controlled by trading organisations. The latter are equipped with crop-
monitoring maps acquired by NOAA-AVHRR. These data are based on vegetation Index map with the
same grid of 50 km x 50 km. The indices are an indicator of activity and advance or delay in vegetation
and hence are not directly transformable to yields. They do, however, give important complementary
information (Dallemand and Vossen, 1995). The monthly MARS bulletins for the period April to
376 Processing of Remote Sensing Data

Module 1 Module 2 Module 3

Fig. 22.4: CGMS model.

September, sent to the DG VI, Agriculture, are a synthesis of surface aspects and yields derived from
a model controlled by NOAA-AVHRR indices.

22.4.4 Perspectives
Application of remote sensing has already become very important with respect to yield estimations
and verification of agriculturists’ statements (Meyer-Roux, 1996). Methods employed are relatively
simple but demand considerable effort in organisation and auxiliary data to provide results at continental
level. Their application is hindered mainly by three factors:
— ^technical constraints: revisit capability and resolution;
— quality and suitability of sensors for the phenomenon under study;
— cost of operations.
Resolution Is not really a problem for yield predictions and In the case of area determinations if
multispectral SPOT data are used. It Is a problem mainly in verification of aid statements and in
particular for determination of exact areas of fields. Acquisition of three or four SPOT images during a
survey does not represent adequate repetition for defining the state of crops and hence for predicting
crop yields. Repetition may be resolved by using radar images of sufficient resolution, about 30 m, but
the corresponding methods are not yet operational and it Is not certain that they will be. The ERS-1
instrument has been used for verification in the case of lack of images of visible band but such attempts
represent only experiments. Moreover, present-day costs do not permit frequent coverage with high-
resolution data, even limited to a representative sampling of images, irrespective of visible or microwave
data. The essential tool used for these Investigations corresponds to low-resolution meteorological
satellites. However, the quality of sensors and production of products is adequate only for extreme
vegetation conditions. The Instrument VEGETATION (see Ghap. 2 and CD SPOT System) operating
on SPOT-4 ought to provide a significant improvement for yield estimations.
 Applications 377

22.5 CONCLUSION
Remote-sensing techniques applied to agriculture have been extremely important and played a strategic
role during the change in the European Community’s agricultural policy for the years 1991,1992 and
1993 whereas national administrations had no adequate conventional system for verifications. A lack
of readily available methods could have led to serious consequences.
Methods of yield estimation In Europe are also significant but economically justifiable only when
the conventional statistical data, less expensive, are not reliable or available too late. Conventional
statistical methods have improved for the last ten years but remote-sensing technology has likewise
progressed. It Is especially Imperative to conserve and Improve these methods for strategic aspects
such as, for example, negotiations with the central European countries at the time of membership in
the European Union. Lack of reliable and neutral data can, in fact, spell trouble for the European
Community and the member countries. Hence, a great effort is underway to rectify the situation.
France has played and continues to play a major role in these operations, primarily through
spatial data and the SPOT program, well suited for the European agricultural requirements and
constituting the main source of data. Methodologies developed by national research centres, and in
particular the INRA Bioclimatology of Avignon, have been significant, in addition to those of teaching
and research institutions.The Soil Map Service of INRA of Orleans has taken the responsibility of soil
aspects and associated geographic information systems. Service societies such as SCOT or computer
Information societies have adapted themselves to meet the demands In the European context, mostly
in agricultural matters. Amalgamating of investigations at various levels of organisation, ranging from
detailed to continental, is no easy task but necessary if the true potentialities of remote sensing are to
be realised.

References
Anonyme. 1997. Prévisions de rendement agricole. Textes présentés lors du séminaire de Villefranche-sur-Mer,
France, 24-27 Oct. 1994. Office for Official Publications of the European Communities, Luxembourg.
□allemand JF, Vossen P. 1995. Agrometeorological models: theory and applications in the MARS project. Proc.
Workshop for Central and Eastern Europe, ISPRA, 21-25 November 1994. JRC-PHARE, EUR 16008 EN.
King D, Daroussin D, Tavernier R. 1994. Development of a soil geographical data base from the soil map of the
European Communities. Catena 21:37-56.
King D, Daroussin J, Le Bas C, Jones RJA, Thomasson A.K. 1995. Estimation des réserves en eau potentielle des
sols à partir de la carte de sols de communautés Européennes.
Meyer-Roux J. 1996. Le programme Spot au service de l’agriculture Communautaire. In: D’une décennie de
réalisations ... à une décennie de promesses, CNES, 15-18 avril 1996.
Meyer-Roux J, Girard C-M. 1997.Télédétection et estimation des récoltes. C.R. Academie d’Agric. Fr., 83 (3); 149-
162.
Vossen P, Rijks D. 1995. Early crop yield assessment of the EU countries: the system implemented by the Joint
Research Center. JRC, EUR 16318 EN.
 23
Soil Mapping

23.1 REMOTE SENSING AND SOILS

The term ‘soil’ refers to the pedological cover: upper layer situated between the atmosphere, biosphere
and fresh rock. In this flat volume (thickness of about a metre to tens of metres and area hundred to
tens of thousands of km^), several layers, known as soil horizons, are differentiated, most often occurring
one above the other.
Obviously, remote sensing can give information only on the topmost part of soil. The penetrating
power of an electromagnetic wave is of the order of magnitude of its wavelength and varies from a
micrometre in the visible band to about a few centimetre in microwave frequencies. Thus, when clouds
or plants do not conceal soil, certain soil parameters can be directly detected only from its surface
state. However, a rain that moistens the surface, a wind that dries it, a slaking crust or efflorescence of
a few mm In thickness suffices to completely modify the spectral characteristics of soils.
Remote sensing is useful for direct acquisition of soil information on multiple dates without altering
the surface state and continuously in geographic space. The multitemporal capability permits acquisition
of dynamic properties of soil, such as its temperature, water content, variation of surface roughness
due to meteorological events or soil works (directly sensed parameters).
It Is possible to obtain information on deeper part of soil cover from data on vegetation whose
spectral characteristics are partly determined by Its roots that penetrate soil (indirectly sensed
parameters).
Studies on surface state of soils have been carried out by several groups; the important among
them are Condit (1970), Goetz and Huete (1991) and Baumgardner (1981) In USA, Bialousz (1978)
and Ciernlewski (1993) in Poland, Epema and Mulders (1987) in the Netherlands, King (1993) of
BRGM, Girard (INA-PG), Courault and Baret (1992) of the INRA, Escadafal, Mougenot and Pouget
(1998) of the iRD and Cervelle (1977) of the Paris University in France. The following discussion Is
based on these papers, but it is not necessary to give all the references In this book, since most of
them are found In a limited number of publications (see list at the end of the book).

23.2 DIRECTLY SENSED PARAMETERS

23.2.1 Spectral characteristics of soils in visible and near
infrared bands
Reflectance curves of soils are uniform, continuous and intersecting; they do not show peaks or
distinct absorption zones, but for water. These curves can hence be compared taking only one or two
wavelength bands. No spectral zones requiring specific interpretation occur as in the case of vegetation
(see Chap. 4, Fig. 4.13).
 Applications 379

23.2.2 Surface states of soil cover

The concept of ‘surface state of so/7’ was proposed in 1978 by Aubert and Girard and developed by
Escadafal in 1981 and 1989. In a wider context it incorporates information on agricultural activity,
slaking crusts, salt encrustations and efflorescences, etc. A surface of soil is defined as ‘the composition
and organisation of a soil surface at given instant \ It is transitional zone between the atmosphere and
soil volume; it is a kind of first or zero horizon of the soil profile. Escadafal characterises it with the
following parameters: colour, texture, effervescence, coarse soil components, vegetation and animal
organic matter and its cover, algae, moss and other plant material; surface features such as slaking
crusts, efflorescences, pores and cracks, microrelief, etc. Laboratory analytical data and microscopic
descriptions can be added to this list. Measurements made for determining the surface state of soils
concern only with parameters occurring on the surface and viewed from above. This definition is
hence correctly well adapted to the spatial view of remote sensing, although this concept was developed
independent of the latter.
A surface state of soil Is a complex entity comprising various soil elements exposed to the sun or
shadow and more or less covered by vegetation.The main factors of spectral characteristics of soil are
those that influence its surface state, viz., colour, roughness (type of surface: slaking crusts,
efflorescences, coarse elements, texture, structure and shadows), calcareous or organic matter, iron,
moisture and chemical composition. The term ‘soil’ in the following text must be understood as ‘surface
state of pedologica! cover’.

23.2.3 Colour
Soil colour is measured in the field with reference to the Munsell code. Hues for soils vary from red (R)
to yellow (Y), passing through all the intermediate spectral bands (YR). In some cases, colour may
tend to blue. In most cases, the brightness value ranges from 0 to 8 and chroma from 0 to 8. Dark soils
have value/chroma codes of 2/2 and 3/2 and bright soils 8/3 or 7/2. The most common soils in France
are characterised by hues of 10 YR, 7.5 YR and SYR when red or 7.5 Y if greenish-yellow. If they are
plotted In a Cartesian reference defined by their reflectance values in the 450 nm and 750 nm bands
(Fig. 23.1), soil colours vary from darker to brighter along a linear cluster of points (Courault et al.,
1998). Soils with the same chroma (/2, /4, /5, /6) are aligned along straight lines constituting a cluster
and ranging between 5% and 10% reflectance for 750 nm band. For soils on these straight lines,
reflectance increases with Increase In brightness value (4/, 5/, 6/, 7/, 8/). Samples with hues of 2.5 Y,
more yellow, occur together with those of hue equal to 10 YR, but with a slight shift in chroma: they
appear duller. Practically no difference Is observed between samples of 7.5 YR and those of 10 YR.
This is conformable with results of colour analysis (see Chap. 3, Fig. 3.9).
It was seen in Chap. 3 that Munsell colour of soils can be related to their reflectance and digital
numbers of satellite Images (Fig. 23.2). Thus soil-colour maps can be prepared if a satellite image
acquired over bare soils is available.
Colour of soil is identified from reflectance curves (Fig. 23.3). When a soil is dark (Munsell value
2 or 3), the curve is concave up to about 800 nm; if the brightness Is medium (Munsell value 4 or 5),
the curve is nearly linear; when brightness is high (Munsell value 6, 7 or 8), the curve is convex. If the
colour is pure (Munsell chroma 6 or 8), the slope of the curve is steep between 400 nm and 600 nm;
If it is dull (Munsell chroma 2 or 3), the slope is small.

23.2.4 Roughness
Surface roughness of soils depends on:
— intrinsic soil factors, such as salt efflorescence, ferruginous crusts, cracks, soil movements
(Gilgai of vertisols), structure and porosity, slaking crusts, etc.;
380 Processing of Remote Sensing Data

2 /,4/, 6/....Munsell brightness

/2, /3,/4.... Munsell chroma

Fig. 23.1: Relationship between Munsell colour and reflectance at 450 nm and 750 nm for 84 soil samples
(after Courault, 1989).

— ^factors extrinsic to soil, such as microtopography or agricultural activity, etc.

Roughness manifestations are generally produced during specific times of the year depending
on meteorological conditions (slaking crusts, efflorescences) or agricultural activities. Soil surface
roughness is hence a temporary state in most cases.
Roughness may also be associated with microrelief created by sands, gravel, pebbles or blocks
(see infra) permanently situated on the soil surface. Roughness In such cases is determined mainly by
shadows produced by these coarse features on the soil surface. It must hence be related to the sun’s
height above the horizon and therefore with season, time of acquisition of scenes and latitude of the
zone under study.
 Applications 381

% CIMEL
reflectance • «

20

• •

15

SPOT
digital 60 50 40
#
•r % spectroradiometric
reflectance
numbers
• •

SPOT
digital _70
numbers

Fig. 23.2: Comparison between 18 radiometric values measured in laboratory (spectroradiometric), reflectance
values (Cimel) and digital numbers in a SPOT Image.

Fig. 23.3: Forms of spectral curves of soils according to colour.

For all types of roughness, the main response in terms of radiance is due to the shadow detected
by the sensor. In fact, sensors are most often vertical whereas the sun Is oblique. Hence, the sensor
perceives a part of illuminated area and a part of shadow. Various Indices have been proposed for
measuring roughness, viz., standard deviation of heights relative to a mean plane, tortuosity index
and shade index (Courault et al., 1993), and a number of models developed (Cierniewski and Courault,
1993).
When the proportion of shadow increases, reflectance decreases (Fig. 23.4). However, this effect
is more important for a bright soil than for a brown or dark soil (Yongchalermchal, 1993).
382 Processing of Remote Sensing Data

In agronomy, Boiffin (1984) defined several phases of soil degradation, ranging from FO: non-
degraded state to F2: degraded and slaked surface. In the last case, soil surface is smooth with no
shadows whereas for the FO state, it is full of mounds. It is observed that the greater the degradation,
the stronger the reflectance (Courault, 1989) (Fig. 23.5). On an organic soil (with manure), with a
stable structure, reflectance is lower than on a limed, brighter and less organic soil (Table 23.1).
Strongest and most variable reflectance is observed for sodic soils with unstable structure.
Agricultural operations also influence the surface roughness of soils. Thus, over a given type of
soil, ploughing (very rugged surface) gives rise to lower reflectance than sowing (less rugged surface)
and movement of a roller makes the soil surface almost smooth, leading to strong reflectance (Fig.
23.6),
Studies of SPOT images have shown that it is possible to differentiate calcareous soils from
slaking soils, since the latter indicate less distinct boundaries than limestone outcrops (Courault and

Fig. 23.4: Relationship between proportion of shadow and reflectance for three soils (590-680 nm)
(after Yongchalermchai, 1993).

Fig 23.5: Reflectance of three different surface states of soil according to their roughness (after Courault, 1989).
 Applications 383
Table 23.1 ; Characteristics of three surface states of soils

Analysis of surface Munsell colour Organic matter (%) pH Structural Clay

state of soils stability content (%)

Soil with manure 10YR4/4 21 6.4 Strong 20

Limed soil 10YR5/6 5.3 7.9 Strong 18
Sodic soil 10YR7/4 6.2 5 Low 20

Fig. 23.6: Reflectance curves of three agricultural fields of various roughness conditions: ploughed, sown and
rolled plots (after Courault, 1989).

Girard, 1990). More recently, three to four surface states of more or less slaking soils could be
distinguished in loams of Caux County (Burlot, 1995).

23.2.5 Carbonate
Carbonate content Influences reflectance of soils by giving very high values starting from the blue
band. The reflectance curve is convex, brightness values very high but less than soils with salt
efflorescence. For total carbonate contents less than 10-20%, carbonate seems to have no influence
on reflectance. Contrarily, beyond this limit, the stronger the reflectance, the larger the carbonate
content, other factors being constant. For values exceeding 60-70% carbonate, there is little difference
in reflectance and saturation occurs in high brightness values in images and aerial photos. Special
processing is often required to improve the signal.
Correlation between reflectance at 400 nm and total carbonate content for 84 soil samples Illustrates
this relationship (Fig. 23.7). However, the significance of this relationship, for Image processing, lies in
graphically deducing total carbonate content from reflectance value. It is seen that when reflectance is
less than 15%, the total carbonate content may vary between 0% and 50%. This is because only a
single-factor analysis is presented here. Thus, a calcareous soil may have the same reflectance as
that of a non-calcareous slaking soil.
In sedimentary regions, highly calcareous zones usually correspond to outcrops of limestone,
chalks or bright marls. Ploughing also often brings these materials to the surface. Consequently, one
384 Processing of Remote Sensing Data

R{% )

Fig. 23.7: Relationship between total carbonate content and reflectance at 400 nm. Munsell values are shown
(Courault, 1989).

way of Identifying them is to analyse contrast between the calcareous zone and its surroundings.
Since contrast is generally high for calcareous zones, boundaries can be readily delineated.

23.2.6 Organic matter

Organic matter reduces reflectance in all bands of the visible spectrum (Fig. 23.8).
The reflectance curve of organic soils remains concave for the entire visible range. For organic
matter contents less than 1.5%, there seems to be no effect on the reflectance curve. Contrarily, for

Organic
matter content

1.3
1.5
2.3
2.8
4.7
37.5

MSS Spectral bands

Fig. 23.8: Reflectance curves (obtained with an Exotech radiometer) of more or less organic soils.
 Applications 385

contents greater than 8% and up to 20%, reflectance decreases. However, since the brightness values
are very low in the visible bands, it is preferable to pay attention to the red and near infrared bands.
At 650 nm, the relationship between reflectance and organic matter content is exponential (Fig.
23.9). In this case also, organic matter content cannot be readily Inferred from reflectance, since other
soil factors interfere.

Fig. 23.9: Relationship between organic matter content and reflectance at 650 nm. Munsell values (4/) are
shown (after Courault, 1989).

In temperate climatic regions, organic zones in bare soil are visible in satellite images or aerial
photos only when they are cultivated. Others are covered by vegetation. Thus, the range of variation of
organic matter content Is much more reduced. Starting from a certain weight ratio of organic matter,
the preceding relationship becomes invalid. In fact, if all soil components of a soil surface are entirely
covered by organic matter only over a few microns (as in A horizon of Chernozem soils— Soil Reference
Manual, Baize and Girard, 1995), its reflectance will be very low, although its organic matter content
may be only 5 to 8%. An organic horizon, whose content Is very high because all components are
made up of organic matter, will not give higher reflectance. In fact, reflectance in the visible and near
infrared bands depends only on surface conditions.

23.2.7 Iron
Since iron is characterised by low reflectance in the blue band, slope of the reflectance curve in the
visible range is high. There is no stronger reflectance in the red band for Iron-rich soils than for other
soils. Contrarily, an absorption zone occurs between the wavelengths 860 and 910 nm. Hence, an
386 Processing of Remote Sensing Data

acquisition band as closer to 400 nm as possible is needed for detecting red soils and two others
around 850 nm and 900 nm for detecting the absorption zones. Unfortunately, such bands are not
presently available In satellite sensors.
It has been observed from laboratory measurements that an exponential type relationship can be
established between reflectance at 900 nm and iron content (Fig. 23.10). It may be noted that the
signal is not affected by iron content below 2 to 3%. When iron content is more than 15%, the reflectance
is small and brightness values are low. A proportional relationship exists between these two limits. Soil
complexity due to mixing of soil components is again observed.

R (%)
900 nm

Fig. 23.10: Relationship between iron content and reflectance at 900 nm. Munsell values are indicated, for
example as 2/ (after Courault, 1989).

Interpreting Iron contents in soils from satellite images is not common. In intertropical regions,
where Iron-rich soils with variable contents frequently exist, coatings of algae or those produced from
brushwood fires are often observed on Iron pans. In deserts, ‘desert varnish’ covers all surfaces so
that iron is not directly perceived (see Chap. 24). In the case of savannahs, dense vegetation canopies,
more or less dry, do not permit perception of soil. Flence, detection of differences In iron content of
soils assumes special importance for cultivated fields in these regions and in Mediterranean zones.

23.2.8 Moisture
Absorption bands for water situated at 950,1150,1450,1950 and 2350 nm are visible in soil reflectance
curves and more so as these are the principal modifications affecting such curves. Flowever, absorption
peaks at 950 and 1150 nm are fairly small, whereas those at 1450 and 1950 nm are generally distinct
since water is always present in the soils, at least associated with salts, oxides or hydroxides contained
In soil. Flowever, no Information is acquired by satellite sensors since radiation of these bands is
strongly absorbed by atmosphere.
 Applications 387

In general, the more the soil moisture, the lower the reflectance (Fig. 23.11) when other soil
components are similar. When the moisture content increases, colour becomes darker and reflectance
lower for all wavelengths.

R (%) H (%)

Fig. 23.11: Variation of reflectance of a soil with moisture (weight percentage).

Moisture thus detected obviously corresponds exclusively to the surface moisture. Information
about moisture content can be obtained up to a depth of 10-30 cm If microwave band is used and
about water movement if thermal infrared data is used (see Chaps. 1 and 26).
Field experiments under controlled conditions have shown that when soil moisture decreases
(from 28.8% to 10.6% by volume), points plotted on a graph with values of reflectance in red and near
infrared bands as co-ordinates fall on two linear segments (Fig. 23.12). An offset and a rotation between
the two linear segments are observed. Scatter of points is smaller for the higher moisture state (colour
10 YR 4/4) than for the lower one (colour 10 YR 7/3). Regression relationships between the two
wavelength bands for the two states are obtained as follows:

Molststate: //?= 1 .6 4 /? -5 6 .5 8 r= 0 .9 8

Drystate: /f?= 1.33 f ? - 65.38 r= 0.94 .

Similar results are obtained from laboratory studies. Between moist and dry states, samples of
organic soil show a very small variation in reflectance (2 to 3%), while for ferruginous, organo-ferruginous
and organo-calcareous soils the variation is larger (10 to 15%); maximum variation (20 to 30%) is
observed for calcareous and gypsum soils.
In reflective middle infrared band, spectral characteristics of soils are dependent on their moisture
content: dry soils have higher reflectance than the same soils with moisture. This spectral region is
much less sensitive to organic matter than the visible. However, as these various parameters are
related with each other (peat soils have higher water contents than sandy soils), it is difficult to isolate
the parameters that influence spectral characteristics in natural conditions.
388 Processing of Remote Sensing Data

Fig. 23.12: Variation of reflectance with moisture content {%) for the same soil (with its colour given in Munsell
code) (after Courault, 1989).

23.2.9 Grain size

Soils are divided into two groups according to their grain size with a limit of 2 mm. Above 2 mm, soil is
referred to as coarse: gravel (2 mm to 2 cm), pebbles (2 to 7.5 cm), stones (7.5 to 25 cm) and blocks
(more than 25 cm). Below 2 mm, it is known as fine soil: clay (< 2 pm), silt (2 to 50 pm) and sand (0.05
to 2 mm). Soil texture Is a characteristic resulting from assemblage of particles of various sizes. In the
visible range of wavelengths no Interaction of radiation with particles takes place since sizes of clay
particles are too large for these wavelengths, but they influence reflectance in the near infrared band.
Silt and sand particles are too large to affect reflectance in the visible or near infrared bands. Some
authors reported variations In reflectance according to their clay or sand contents, but then they are
dependent on the water content of these components.

M Clays
Relationships between grain size and moisture content are indirect. In fact, for a given moisture content
in the ground, clays contain more water than sands. Moisture content hence is not a good criterion for
evaluation. Hydraulic potential or water retention capacity should be taken for a soil sample. Clays are
often associated with organic matter. Moreover, as clays are usually of low brightness, reflectance of
clayey soils is relatively low.

■ Sands
Sands are drier since they retain no water in their porosity, often very coarse (see Chap. 24). On the
other hand, most sands are of high brightness and hence sandy soils are relatively strong reflectors.

9 Silts
In the case of non-calcareous silts of low organic content, the surface structure is modified by rains
and roughness changes to smoothness. This is the slaking process. They then appear very bright in
 Applications 389

aerial photos and satellite Images. However, this high reflectance, in fact, ought to be attributed to the
low roughness of silts, the cause of slaking, and not the nature of particles. The same silty soil, just
ploughed or tilled, has much lower reflectance.

■ Coarse components
Surface abundance of all coarse components in soil is referred to as ‘stoniness’. Most often reflectance
increases with stoniness. This is mainly due to the fact that the energy reflected by these coarse
components, which usually act as specular surfaces is greater than the shadows they produce. Moreover,
in quarries, extraction of material gives rise to a surface coating by fine powders, most often white.
Hence, quarries are generally characterised by very strong reflectance.

23.2.10 Salts
An integrated study of saline soils was carried out by Mougenot et al. (1993). Laboratory measurements
(Mougenot, 1990) (Fig.23.13) revealed that absorption zones of salts, except that of NaCI, are clearly
differentiated ih the infrared band. Therefore, bands 5 and 7 of Thematic Mapper or the reflective
middle infrared band (b4) of SPOT must be used for detection of salts by satellite images.
In laboratory measurements, halite Is detected mainly from very bright hues in the blue band of
the visible spectrum. Gypsum, very sensitive to water, shows three absorption peaks at 1450, 1950
and 2350 nm. Jarosite, common in mangrove zones, depicts absorption peaks at 1950 nm and, if iron
is present, at 600 and 900 nm. Absorption peaks for calcite occur between 1600 and 2500 nm, with an
apparent peak at 2150 nm.

Wavelength (nm)

Fig. 23.13: Spectral characteristics of various salts: Halite (NaCI), Gypsum (CaS0 4 , 2 H2 O), Alunite [KAl3 (S0 4 ) 2
(OH)g], Jarosite [KFe3 (S0 4 )2 (0 H)g], Calcite (CaC0 3 ); (A) Laboratory and (B) Field measurements (after Mougenot,
1990).
390 Processing of Remote Sensing Data

Field measurements are more difficult to interpret since an association of several salts mixed with
organic matter, various minerals and water is often observed. Measurements are mainly affected by
surface states of soil such as encrustations and efflorescences. However, it is necessary that the
image resolution be very good or that these crusts cover large areas as in the case of dry and saline
zones.
Saline soils can often be detected indirectly if plants also exhibit salinity.
It is also necessary to use multi-date images, since characteristics of saline soils change rapidly
depending on the seasons. It may be reasonably expected that detection of saline soils can be improved
using new thermal and microwave sensors.

23.2.11 M ultifactorial analysis

Since soils generally comprise varied quantities of organic matter, carbonates, iron, etc., of various
particle-sizes (soil texture) and distributed variedly (soil structure), multifactorial analysis becomes
necessary. A factorial correlation analysis was carried out using reflectance and soil parameters of 84
samples (Fig. 23.14). The analysis revealed several classes in a plane of axes 1 (45.5% information)
and 2 (33.5% information).

Fig. 23.14: Plane (1,2) of multifactorial analysis for carbonate, organic matter and iron for 84 soil samples,
defined by four wavelengths: 380,450, 500 and 600 nm (after Courault, 1989).

Axis 1 indicated three groups according to organic matter content: OM1 more than 45%; O M 2,20
to 45% and O M 3,10 to 20%. Axis 2 differentiates in its positive part two groups by their total carbonate
content: C al more than 70% and Ca2, 50 to 70%. In the negative part of axis 2 and positive part of
axis 1, samples rich in iron are distributed in two groups: Fel more than 30% and F e 2 ,15 to 30%. In
the centre of the plane, two groups containing less than 10% organic matter and less than 6% iron can
be identified, one without carbonate (OMFe) and the other with less than 10% carbonate (OMCaFe).
Another group containing carbonate and organic matter but no iron (OMCa) is also noticed.
This example does not represent all possible cases of the large number of soil types. Nevertheless,
it shows that soil parameters of a region can be modelled by their multifactorial analysis with reflectance
and digital numbers from the image under study.
 Applications 391

Similar investigations on moisture and surface roughness resulted in the following conclusions:
— dry organic soil has a lower reflectance than that of a less organic and moist soil. Organic
matter content hence dominates over water content.
— A smooth soil has a stronger reflectance than that of a rough soil, irrespective of the moisture
content. The effect of roughness is also greater than that of water content.
— However, on the ground, roughness and moisture content of soils are not absolutely independent
of each other. Too moist a soil cannot be ploughed; evaporation and hence moisture content of soil
surface and the sun’s height (hence shadows) are likewise not independent.
In a Thematic Mapper image of southern Tunisia (Fig. 23.15), from the blue-green (TM 1,450 to
520 nm) and infrared (TM 7:2.1 to 2.35 jim) bands, Escadafal (1989) identified various surface states
of soil. These are shadowed (1), gypsum (2, 3), carbonate (4 to 7), sandy, silty and calcareous (8 to
11), and quartz (12 to 14) soils. This demonstrates the significance of reflective middle infrared data.

Fig. 23.15: Separation of gypsum, carbonate and quartz soils from TM 1 (blue-green) and TM7
(infrared) bands of LANDSAT (after Escadafal, 1989).

The magnitude of influence of soil parameters on reflectance can hence be given qualitatively.
Colour constitutes the primary indicator since it is related to many parameters and physically inferred
from radiometric measurements in the field or from aérospatial platforms. Surface roughness seems
the next most important factor for signal variation. Next in importance are total carbonate and organic
matter, which have opposite effects, followed by iron and moisture content in the last position.

23.2.12 Linear clusters of soils

When the proportion of shadows changes, soil reflectance values for two wavelength bands, blue
(400-500 nm) and near infrared (800-900 nm) or red (600-700 nm) and infrared, plotted in a Cartesian
diagram, cluster along a linear direction. This Indicates that reflectance as well as digital numbers of
images for the two bands are positively correlated (Table 23.2). This is referred to as a linear cluster of
392 Processing of Remote Sensing Data

Table 23.2: Parameters of five surface states of soils

Reference index of Colour of dry soil Colour of moist Slope of regression Ordinate at Correlation
surface state soil line origin coefficient

VB 10YR 8/1 10YR7/2 0.75 18.7 0.98

MB 10YR7/1 10YR5/3 0.87 9.5 0.99
BS 10YR6/4 10YR5/4 0.98 6.7 0.97
MD 2.5 Y 5/2 2.5 Y 4/2 1.04 4.6 0.98
VD 2.5 Y 5/4 2.5 Y 4/4 1.24 2.2 0.95

soils (Fig. 23.16). Such a cluster consists of straight-line segments defined by the slope of regression
line and intercept on the ordinate.

NIR (%)
55 •— « Very bright soil (VB)
O— O Moderately bright soil (MB)
50 ^ __ X Brown soil (BS)
It--.» Moderately dark soil (MD)
o— c Very dark soil (VD)

40

30

VB :NlR = 0.75 R + 18.74 - r = 0,98

20 MB :NIR = 0.87 R + 9.52 - r = 0.99
BS :NIR = 0.98 R + 6.71 - r = 0.97
MD ;N1R = 1.04 R + 4.59 - r = 0.98
10 VD :NIR = 1.24 R + 2.18 - r = 0.95

10 20 30 40 45 R (%)

Fig. 23.16: Five linear clusters of soils (after Courault, 1989).

It is seen that surface states of very bright (VB) soils have an intercept greater than that for darker
soils. The slope of the regression line for bright soils Is smaller than that for dark soils.
These results show that the soil line Is dependent on its colour but not on the quantity of shadow.
As the slope of these straight lines varies, a soil may not have a single soil line. It is observed that soil
lines are distributed as a cluster which widens towards lower values of reflectance.

23.3 MODEL OF IMAGE INTERPRETATION

Various surface states of a soil, represented by digital numbers on a visible-infrared plane, occur as a
cluster globally oriented in the direction of the first bisectrix. The preceding data plotted on such a
graph constitutes a model for interpretation of soils using images (Fig. 23.17). Most often, digital
numbers are projected onto two axes representing red and near infrared bands, since these are most
explicit for vegetation. For soils, a projection on infrared and green or blue is preferable.
 Applications 393

Fig. 23.17: Model for soil interpretation in images.

Very dark soils (0) are associated with the lowest values of green (G) and infrared (IR). These
may be confused with pure waters or cloud shadows. Shapes of map zones most often enable
differentiation of an extended form of a water body or a cloud shadow from a very dark soil surface.
When a cloud Is not observed in the image, identification of a soil with a cloud shadow may be difficult.
If soils are brighter (1), they are situated in the diagram at higher values for both the axes. Brightest
soils occur at the extremities of the axes (2) and may be confused only with clouds or snow. For these
two latter features, ambiguities can be eliminated from knowledge of geography or examination of the
two images, If no map is available. It may be noted that waters charged with suspended particles (9)
are generally situated outside the cluster, to the right (in the model used in Fig. 23.17).
Rough bare soils (4) are fairly close to very dark soils. They often correspond to ploughed lands
for cultivated fields. When roughness decreases, points representing surface states of soil move away
from the centre. They may be confused with some bright soils (3).
Highly organic bare soils (5) are close to dark soils. Calcareous soils are close to bright soils
(between (1) and (2)).
Iron-rich soils (6) are not specifically distinguished in this diagram, since they are situated to the
left of the cluster and are confused with vegetation.
The effect of soil moisture causes a displacement ((7), (8)) in an inclined direction relative to that
of the cluster.
This cluster can be readily detected in an Image by analysing a two-dimensional histogram (see
Chaps. 9 and 8, Figs. 9.3 and 8.4). A principal component analysis can also be carried out, in which
the first axis most often corresponds to the linear clusters of soils.
Obviously, this model needs to be modified to suit each situation. Such modifications are related
to the geographic region, date of image acquisition and spectral bands available. An example is given
In Figs. 7.4 and 9.4 (see Chaps. 7 and 9).
However, the linear clusters for soils given here assume that the soil is not covered by vegetation,
which is common in arid zones and which happens only during half a year for cultivated soils. Hence,
spectral characteristics of soils must be analysed when they are not covered by vegetation.
394 Processing of Remote Sensing Data

23.4 REMOTE SENSING AND SOIL MAPPING

In France, soil scientists have been using aerial photos for delineating boundaries between soil units
since 1950s and since the beginning of soil mapping (Boulaine, 1957; Jamagne, 1967). Usage of
satellite images in normal soil mapping took longer and is less common, since their interpretation
necessitates computerised data processing and hence a specialised group. Also, images are not as
readily available as aerial photos.
Escadafal and Pouget (1987) prepared 1:100,000 maps of surface state of soil for Tunisia using
satellite images. The legend of the maps clearly explains all soil parameters used, viz.,
— predominantly gypsum soils, with shadows, stones, pebbles and encrustations;
— predominantly calcareous soils, with shadows, stones, pebbles and crusts, silt and gravel;
— predominantly quartz soils, with more or less fixed sands, sandy covers on limestone, mobile
sands and sandy covers on gypsum;
— soils mostly covered by dry vegetation, dense or sparse.
At present, satellite images are found to be very useful In the programme of soil Inventory,
management and conservation (Ministry of Agriculture, INRA) In 1:250,000 scale for the programme
regions and national territory. Concept of soil landscape is used in particular for mapping at these
scales.

23.4.1 Soil landscape and remote sensing

Soil landscape is defined as follows:
'Combination of soil horizons and landscape elements, viz., vegetation, effects of human activity,
geomorphology, hydrology, substratum or bedrock, whose spatial organisation determines a (ora part
of) soil cover in its entirety’. Subsystems can be defined as soil landscape units and soil landscape
elements.
A soil landscape unit corresponds to a soil system, is based on geomorphology and can be
represented by a map unit. A soil landscape element corresponds to one or several spatially connected
typological units, connected spatially, with a simple spatial organisation (for example, a sequence of
soils).
Satellite images can be used to define most of the landscape elements such as vegetation,
effects of human activity, geomorphology, hydrology, substratum or bedrock, if necessary aided by:
— vegetation maps (1:200,000),
— geological maps (1:320,000, 1:80,000 and, ultimately, 1:50,000),
— topographic maps (1:250,000,1:100,000 and 1:50,000).
Visual interpretation (Girard et al., 1993; Bertrand, 1994; Girard and Gllliot, 1997) and computer
processing (Burlot, 1995; Gllliot et al., 1998) of images provide a synthesis of various landscape
elements. Each of these elements must be defined by using a systematically coded format. All such
formats can then be statistically analysed and interpreted.
A similar type of interpretation can be made when only maps are available but not satellite images,
but synthesis between various maps becomes more complicated (Antoni and Girard, 1996).

■ Interpretation with limited soil data

For each map zone or map unit, assumptions can be made about chorological laws relating landscape
elements with existing soils, since soil differentiation factors are then available. These assumptions
will be supported by ground truth data or by maps. If exist. It must be then verified by GIS whether all
the assumptions are mutually coherent and spatially coherent with chorological laws (Francoual, 1997).
Lastly, ground truth verifications should obviously be made. In fact, several factors of soil differentiation
cannot be interpreted using satellite images or aerial photos (Yongchalermchal, 1993).
 Applications 395

Interpretation with abundant soil data

If several maps or ground data exist, we can preliminarily define chorological laws relating soil cover,
differentiation factors and environmental features seen in Images. Using these laws, guidelines for
interpretation of satellite images are formulated.
Importance of satellite Images lies in obtaining an instantaneous view over a large area (3600
km^ for SPOT) and hence comparing and, consequently, interpreting map zones situated at tens of
kilometres from each other. The resolution of satellite images, 10 to 30 m, is adequate for soil landscape
interpretations in 1:250,000 and 1:100,000 or even 1:50,000 scales (Gilllot and Girard, 1996). When
interpretation is carried out In GIS and Images restored on a topographic base, the results of
interpretation can be directly used for GIS analysis.

23.4.2 Method of soil landscape elaboration

■ Interpretation of soil landscapes
Satellite images provide a possibility of interpreting various soil genetic factors in an integrated manner.
A single map effectively gives the overall general structure of the physical environment of a region,
including morphology, hydrology and land cover. Interpretation is facilitated since information is acquired
at a single moment in a homogeneous manner over a vast field. Already published thematic maps,
such as topographic, geologic and vegetation maps, can be used for enhancing the semantic content
of each unit delineated In the image. These maps, which may not have been of the same scale and not
prepared on the same dates, must obviously be used for extracting semantic information and transferring
it to the map units based on a graphic model coherent with that of satellite images. These maps cannot
be directly scanned and superposed over each other, since such superposition results in a meaningless
map with no semantic coherence. Boundaries between soil units are drawn in satellite images by
incorporating data from other maps.

■ Geomorphology
In temperate zones and for interpretations aimed at constructing graphic soil databases in 1:250,000
scale, morphology is one of the main factors of differentiation of soil landscapes. Morphology and
hydrology are readily Interpreted from satellite Images. In fact, as the field of view is large, a synoptic
perception of the entire geomorphology is obtained, which is rarely the case with aerial photos or
ground observations; its form as well as its position vis-à-vis other entities is readily analysed. An
entire watershed and distribution of talwegs in it can be easily mapped. Since zooming is possible
without difficulty, any morphological feature can be analysed In detail and a more general picture
obtained. Thus, fronts of talwegs, extending beyond ‘temporary or permanent watercourses’ indicated
on topographic maps, can also be traced. It Is possible to further improve this Interpretation when a
digital elevation model is available.

■ Winter images with bare soils

Satellite images, especially those acquired in winter when soils are bare revealing the hue of surficlal
horizon, give correct information about soil landscape elements.
\ Thus clayey soils are differentiated from marly and calcareous soils that are very stony, as well as
coarser elements on soil surface.
As deciduous trees lose their leaves, various talwegs running through forests can be Identified.
Impact of drainage patterns on rocks and hence their resistance to weathering and soil erosion can
thus be interpreted. This enables assumption of soil thickness and types of soil genesis.
396 Processing of Remote Sensing Data

Permanent grasslands as well as very humid or even peat zones can also be readily distinguished,
leading to interpretation of differences in hydrological regime of soils.
Lithological interpretation of geological maps gives information on decomposition of rocks, from
which, data pertaining to various soil parameters such as acidity or content of carbonaceous elements
is obtained. Probable texture of material in which soil Is formed can also be inferred.

■ Objectives of soil landscape maps

Objectives of preparing soil landscape maps are numerous.
1. Landscape maps prepared from a satellite image facilitate integration of various data sets already
acquired at various scales and making a graphical synthesis. Images furnish scale transfer functions
since they have a vast spatial field. Moreover, when interpretation is directly made under a
geographic information system, zooming can be used to delineate boundaries and hence
incorporate various precision levels of boundaries. Thus, information corresponding to maps of
various scales can be taken into consideration.
2. Importance of a soil landscape map lies in the fact that it facilitates a clear understanding of the
soil environment through various factors and their comparisons and enables definition of
chorological laws by multi-criteria analysis, leading to evaluation of spatial organisation of soils.
3. Once a sketch is ready, a legend is prepared which must obviously be corroborated and explained
by ground observations and soil analyses.
4. When no map exists for the entire region under study, such a sketch is useful to classify soil
environment for the purpose of gathering data from specific places and hence for planning field
itineraries and observation profiles. Such field studies are aimed at characterising soils defined
by soil landscape analysis and thus verifying the chorological laws proposed. The latter may be
confirmed or rejected. In the second case, other relationships need to be investigated. Field
studies must also provide verification of questions raised during preparation of legends for soil
landscape maps.

23.4.3 Example
A Study was conducted in Lorraine for the regional chamber of agriculture (Girard and Gllliot, 1997) for
testing the image interpretation of a 1:250,000 map.
Chorological laws used In this study were based on morphology, geology and land cover types.
The soil cover seemed strongly related to the geology comprising limestone and marls, since suriace
formations are less extensive. The essential soil constraints are the shallow soil cover and excessive
water. People had to necessarily adapt themselves to these constraints of soil use. Consequently, the
land cover types, clearly sensed in images, were also used for soil studies.
As sun’s elevation was low (18° 06') at the time of image acquisition (10 December 1987), the
small level differences between the two sides of a talweg could be distinctly seen, since one was
Illuminated and the other not. Numerous talwegs in the Woevre region could thus be delineated.

■ Criteria used
Three main criteria, viz., land cover, morphology and geology, were used for this study. The following
modes were used for the three criteria:
— Land cover: forest, permanent grasslands, crops, fallow lands and orchards.
— Morphology:
® hydrology, lakes, talwegs, secondary rivers, main valleys and low terraces;
• relief forms: plains, slopes (zero, small, moderate and steep), undulating morphology, low valley
landscape, plateau, ledge, mounts/peaks.
 Applications 397

— Lithology:
• Surface formations: alluvium, recent or ancient, colluvium and talus, sands, gravel, grits, plateau
silts;
• Calcareous sedimentary material: clayey marl, marl, calcareous marl, hard marly limestone,
gryphaeate limestone, oolitic limestone, sublithographic limestone, coral limestone and sandy
limestone;
• Noncalcareous sedimentary material: clays, schists, sands.
Soil references (Baize and Girard, 1995) interpreted from soil landscapes in the zone understudy
comprise (in alphabetical order): Brunisols, Calcisols, Calcosols, Colluviosols, Fluviosols, Lithosols,
Luvisols, Pelosols, Redoxisols, Reductisols, Rendisols, and Rendosols.
Most reference soils have not been defined by precise characteristics, since the latter cannot be
identified by remote sensing.

■ Some chorological relationships

Some chorological relationships can be identified for calcareous, clayey and loamy substrata as well
as for talwegs in the region studied.

□ Limestone
If the limestone is hard, disintegration takes place slowly and decalcified clays (Rendisols) remain. If
erosion is Intense, very little material remains and hence results in thin soil (less than 10 cm in thickness;
Lithosols). If the limestone is softer and oolitic, Rendisols or Calcisols are obtained depending on the
distribution between disintegratlon/alteration of limestone and erosion. If the limestone Is marly or if
the bedrock is marl, Calcosols or Calcisols are produced.

□ Clays (and schists)

Clays and schists give rise to clayey Brunisols, more or less thick, which may be saturated. Pelosols
may also be observed in such zones. The Oxfordian clays on Woevre plateau produce redoxic
Neoluvisols and clayey Redoxisols.

□ Silts
Silts are covered by forests or crops and produce Neoluvisols, redoxic or clayey. Under forests, silty
zones could not be Interpreted. They can be identified with difficulty by a variation in the species,
which is rarely evident if forest exploitation has modified the environment, or by the vigour of plants
which can be eventually Identified in images acquired at the beginning of summer or in aerial photos.
Winter images do not reveal this differentiation.

□ Talwegs
Talwegs are numerous in the region between Woevre and Moselle. Clayey talwegs are distinguished
from hydromorphic talwegs which constitute the flow of ‘lakes’ In the Woevre region. These are continued
by flat bottom talwegs, which are fresher and enter into the Bajocian limestone, creating several
meanders. V-shaped talwegs in the Bajocian join them, most often under the forest.

■ Soil landscape map

A soil landscape map comprising 54 units (CD 23.1) was prepared using visual interpretation and
aided by topographic, geologic and vegetation maps. Each unit is described by 15 variables: 2 soil
references, morphology, 2 types of land cover, 2 slopes, 2 geological stages, 2 lithologies and 4
characteristics for soil types: coarse elements, texture, redoxic nature and calcification. This description
398 Processing of Remote Sensing Data

is accompanied by comments indicating possible accuracy of these variables. All this information was
included In a database for analysis of maps under GIS.
Obviously, It is difficult to represent all these units by allotting a separate colour to each such that
they can be readily identified. A simplified soil landscape map can be prepared based only on the
most common reference in the unit and defining a dozen groups.
Graphically, some units are more compact than others that are more distorted. This Is characteristic
of visual image interpretation. When the chorological laws are based on distinct morphological data,
such as talwegs, boundaries can be readily traced very accurately. Contrarily, when they are based on
data indirectly sensed from images, boundaries are more rounded or less distorted.
It Is important to draw all graphic boundaries In order to Integrate them with other geographic
maps such as topographic map, digital elevation model or any other digital map. It is thus possible to
add boundaries, to modify them and to highlight graphic and semantic information.

□ Contrast
Lastly, contrasts need to be mentioned; contrast between two sites is equal to the ratio of semantic
distance to geographic distance. Semantic distance can be computed synthetically by a mathematical
distance between soil references or analytically by a mathematical distance between various soil
parameters such as coded observations, measurements, soil analyses, etc. Contrasts vary in the
case of boundaries. Some boundaries have high contrast and probability of their detection is very
high. Others may have less contrast for two reasons:
1. Difference between soil references (numerator) is small, indicating that the chorological laws are
not clear to ensure delineation of boundaries from the data used. Such may be the case of
boundaries between Rendosols and Calcisols, for example.
2. Geographic distance between soils (denominator) is large and transition between two units is
gradual or it cannot be correctly sensed in satellite images. Such is the case of silty soils, which
cannot be delineated under forest for want of establishing chorological laws. If they exist, we may
not be able to locate them. In such a case, another map is used and the boundaries determined
by other means are drawn on it, but without being able to justify them from satellite data.
Contrast on a map is estimated by the area of coloured zones and by differentiation of colours by
the eye.

□ Validation and factorial maps

A better method of validating a map is by constructing factorial maps from the base map. As the map
discussed here is a soil landscape map, it can be validated partly by thematic maps corresponding to
the criteria used for chorological laws, viz., land cover, morphology and geology (CD 23.2). It is also
possible to draw maps using the 15 criteria used to describe each unit. A new thematic map Is thus
prepared: for example, a ‘soil genetic ambience map’. The latter makes use of redox and carbonate
characteristics (CD 23.2).
The following soil groups were identified: cognate group of ‘reductive-redoxic’ soils (blue, magenta,
green); cognate group of ‘carbonate-saturated’ soils (yellow, green); cognate group of ‘noncarbonate’
soils (red, pink, magenta).

■ Conclusions on soil landscape mapping

Soil landscape maps are more informative than the earlier small-scale soil maps. It Is desirable to
compare the time required for conventional methods and remote sensing methods of mapping virgin
areas.
Such a comparison was made (Girard, 1995) between a 1:250,000 soil map, prepared by a
combination of soil maps of various scales for the region west of Montpellier (Bornand et al., 1988),
and a soil landscape map obtained from interpretation of two SPOT stereoscopic images of two
 Applications 399

seasons. In the case of soil units that are not under forests (the latter being difficult to interpret by
remote sensing), 73.8% units were found common for both the maps. The image interpretation and
data processing took not more than one month, which is smaller than the time required for integration
of existing maps.
The soil landscape map constitutes a very good tool for classification of soil cover in zones for
which no ground data are available. Moreover, it is a practical tool for synthesising soil data derived
from maps of various scales. In fact. It provides a geographic framework for such a synthesis, either
from existing maps (Antoni and Girard, 1996) or from satellite images (Yongchalermchal, 1993; Bertrand,
1994; Francoual, 1997).
Use of remote sensing and chorological laws, developed around the concept of soil landscape,
enables flexible and rapid production of Integrated soil maps of 1:250,000 scale for various objectives.
Diachronic images are useful for preparing thematic maps pertaining to erosion, soil movements,
floods, determination and classification of wetlands and hence of hydromorphic soils.
However, for better utilisation of remote sensing, ground surveys are necessary in all cases for
identifying the proper functioning of soil cover and obtaining necessary information on mapping, soil
genesis and various thematic applications.

23.5 CONCLUSIONS
A fairly efficient model is now available for determination of some soil parameters from reflectance
and digital number data. This model ought to become more powerful with use of microwave and
thermal infrared data.
Obviously, the position of pixels in their neighbourhood must be taken into consideration, since
soil cover is a continuum. Textural processing methods are used for this purpose. In the case of
mapping, remote sensing and the concept of soil landscape have proven useful: it is currently employed
for preparing reconnaissance soil maps in the national programme ‘Soil inventory, management and
conservation’, aimed at covering the entire country by 1:250,000 soil maps. The main Importance of
remote sensing In soil surveys lies in providing systematic and repetitive geographic information.

References
Antoni V, Girard M-C. 1996. Reconnaissance régionale des sois du Tarn-et-Garonne. Établissement d’une carte
des pédopaysages à l’echelle du 1/250,000, Adeprina, 37 pp.
Aubert G, Girard M-C. 1978. Vocabulaire de l’environnement pédologique tropical, ACCT, Paris, 81 pp.
Baize D, Girard M-C. 1995. Référentiel Pédologique. INRA-AFES, Paris. 332 pp.
Bertrand P. 1994. Élaboration d’une base de données localisées sur les agropaysages à partir d’images satellitaires:
application à l’étude des organisations spatiales et à la segmentation du département de l’Yonne, Mémoire
de Mastère SILAT, INA-PG, 45 pp.
Bialousz S, Girard M-C. 1978. Les coefficients de réflexion spectrale des sols dans les bandes de travail de Landsat,
Fotointerpretacja w geograffi, Katowice, III (13): pp. 96-109.
Boifffin J. 1984. La dégradation structurale des couches superficielles du sol sous l’action des pluies. Thèse de
riNA-PG, Paris, 320 pp.
Bornand M, Barthes J-P, Bonfils P, Legros JP, Conventi S. 1988. Region Languedoc-Rousillon. Carte régionale des
sols et des contraintes du milieu à l’échelle du 1/250 000. Essai méthodologique. Convention PIM CEE
Régions.
Boulaine J. 1957. Étude des sols de la plaine du Chelif, Thèse Doctorat d’État, Université d’Alger, 582 pp.
Burlot F. 1995. De l’interprétation visuelle à l’interprétation automatique des images satellitaires: application aux
pédo- et hydropaysages. Mémoire Mastère SILAT, Grignon, 65 pp.
Cervelle B, Malezieux J, Caye R. 1977. Expression quantitative de la couleur liée à la réflectance diffuse de quelques
roches et minéraux, Bull. Soc. Fr. Minerai. Cristall., Université de Paris, 100:185-195.
400 Processing of Remote Sensing Data

Cierniewski J, Courault D. 1993. Bidirectional reflectance of bare soil surfaces in the visible and near-infrared
range. Remote Sensing Reviews, 17:321-339.
Condit HR. 1970.The spectral reflectance of American soils. Photogrammetric Engineering & Remote Sensing, 34:
955-960.
Courault D. 1989. Étude de la dégradatior> des états de surface du sol par télédétection, Thèse de TINA-PG,
Grignon, SOLS, 17:239.
Courault D, Girard M-C, EscadafaI R. 1988. Modélisation de la couleur des sols par télédétection, 4® coll. Int.
“Signatures spectrales d’objets en télédétection”, Aussois, Esa SP-287, pp. 357-362.
Courault D, Girard M-C. 1990. Éléments d’interprétation des marques d’érosion intra-parcellaire à partir d’images
SPOT. Photo-interprétation, 1 (2): 11-20.
Courault D, Girard M-C, Bertuzzi P. 1993. Monitoring surface changes of bare soils due to slaking using visible and
near infrared reflectances. Soil Science Society of America Journal, 57 (6 ): 1595-1601.
EscadafaI R. 1981. L’étude de la surface du sol dans les régions arides (sud-tunisien). Recherches méthodologiques,
ORSTOM, Bondy. ES. 187, 60 pp.
EscadafaI R. 1989. Caractérisation de la surface des sols arides par observation de terrain et par télédétection.
Thèse de pédologie. Université Paris VI, 317 pp.
EscadafaI R, Huete AR. 1991. Improvement in remote sensing of low vegetation cover in arid regions by correcting
vegetation indice for soil ‘noise’, C.R. Acad Sc., Paris, 312 (2 ): 1385-1391.
EscadafaI R, Pouget M. 1987. Cartographie des formations superficielles en zone aride (Tunisie méridionale) avec
LandsatTM, Photo-interprétation, 4 (2): 9-12.
Francoual T. 1997. Détermination des agropaysages du département du Rhône par interprétation visuelle de données
satellitaires SPOT. Réalisation d’une base de données spatialisée, Rapport INA-PG, Chambre régionale
d’agriculture de Rhône-Alpes, 34 pp.
Gilliot J-M, Girard M-C. 1996. Étude de la vulnérabilité des sols à l’érosion, dans la vallée de l’Orvin, Chambre
d’agriculture de l’Aube, 15 pp.
Gilliot J-M, Girard M-C. 1997. Le programme Inventaire, gestion et conservation des sols et les pédopaysages.
Application à une zone de la Lorraine, Chambre d’agriculture régionale de Lorraine, 30 pp. + ann.
Gilliot J-M, Girard C-M, Girard M-C. 1998. Cartographie et analyse spatiale des zones humides de Champagne-
Ardenne, Direction régionale de l’environnement. Champagne-Ardenne, 29 pp.
Girard M-C. 1995. Apport de l’interprétation visuelle des images satellitaires pour l’analyse spatiale des sols. Un
exemple dans la région de Lodève, Étude et gestion des sols, 2 (1): pp. 7-24.
Girard M-C, Soyeux E, Bornand M, Yongchalermchai C. 1993. Structuration de l’espace régional et protection des
ressources naturelles. C.R. Acad. Agric. Fr., 79:37-50.
Goetz AFH. 1991. Overview: Imaging spectrometry for studying earth, air, fire and water. EARsel, Adv. Rem. Sens.,
11 (1):3-15.
Jacquemoud S, Baret F, Hanocq J-F. 1992. Modeling spectral and bidirectional soil reflectance. Remote Sensing of
Environment, 41:123-132.
Jamagne MX. 1967. Bases et techniques d’une cartographie des sols, Ann. Agro., N® hors série, 18:142.
King Ch, Delpont G. 1993. Spatial assessment of erosion: contribution of remote sensing, a review. Remote Sensing
Reviews, 7:223-232.
Mougenot B. 1990. Caractéristiques spectrales de surfaces salées à chlorures et sulfates au Sénégal, 2®journées
de télédétection: Caractérisation et suivi des milieux terrestres en régions arides et tropicales. Colloques et
séminaire, ORSTOM, Bondy, pp. 49-70.
Mougenot B, Pouget M, Epema GF. 1993. Remote sensing of salt affected soils. Remote Sensing Reviews, 13:
241-259.
Mulders MA. 1987. Remote sensing in soil science. Development in Soil Science, Elsevier, 15:379.
Pouget M, Mulders MA. 1988. Description of the land surface for correlation with remote sensing data. Proceedings
of the 5^^ symposium of Working group remote sensing, ISSS, Budapest, pp. 153-158.
Stoner E-R, Baumgardner M-F. 1981. Caracteristics variations in reflectance of soils. Soil Sci. Soc. AM. J., 45:
1161-1165.
Yongchalermchai C. 1993. Étude d’objets complexes, sol/plante, à differents niveaux d’organisation: de la parcelle
au paysage, Thèse de l’INA-PG, Grignon, Sols, 19:232.
 24
Mining Geology
Although the basic objective of mineral prospecting lies in the discovery of mineral deposits, often
very small in size and almost always concealed at depth, a task almost impossible for a surface
technique such as remote sensing, the latter has emerged as a valuable tool for the prospector. This
is due to the fact that it forms an integral constituent of a coherent approach guided by métallogénie
models and is complementary to other techniques of investigation such as geophysics or geochemistry.
It has hence become an effective component of synergistic data, leading to decision making of a
prospector who is generally guided by ground observation data.

24.1 METALLOGENY AND TYPES OF DEPOSITS

Any problem of mineral exploration invariably reduces to two types of questions: What to search and
where to find it? The first question is often guided by current economic conditions and by the price of
a particular metal In the international market. Once a choice is made, métallogénie models are studied.
This science of ore deposits is based on their geology and geochemistry and on the chronology of
facts leading to the formation of mineral concentrations. Solution of the second question, viz., location
of mineralisation, will be guided by more global concepts related to tectonic evolution of entire regions
of the globe, which thus define preferential zones, métallogénie provinces, where deposits may be
concentrated.
This approach is imperative for employing prospecting techniques and especially remote sensing.
As the latter can only indicate indirect surface effects of mineralisation, success of its utilisation depends
on identification of precise métallogénie objectives before any Investigation. We can locate a thing
only when we know what to search for! Experience of mining geologists is imperative for defining the
interpretation keys which are essential for application of remote sensing in this domain.

24.1.1 Models of mineral deposits

Three fundamental questions are posed in metallogeny (Pelissonnier, 1976): where do the constituent
elements come from {source), in what manner do they arrive at the point where they are presently
found (transport) and, lastly, by what mechanisms are they deposited or concentrated (deposition)?
Based on these three criteria, a volcano-sedimentary hydrothermal deposit, for example, refers to a
mineral concentration originating from an eroded volcanic source and deposited in a sediment in
which hydrothermal fluids develop later, probably due to reactivation of tectonic events.
Such theoretical models often correspond the definition of type systems based on observations
on existing deposits. For example, the ‘Kuroko’ or ‘Chypre’ type of deposit is a ‘gossan’ model of pyrite
accumulations (Fig. 24.1).This volcanic type of deposit with hydrothermal concentration is manifested
on the surface by an alteration aureole rich in haematite, which is particularly favourable for detection
by multispectral remote sensing in desert and semi-arid regions (Kruse et al., 1990).
402 Processing of Remote Sensing Data

Fig. 24.1 : Volcano-sedimentary ‘gossan’ type of Pb-Zn deposit. It is the case of sulphide accumulation
followed by erosion.

24.1.2 An examples uranium deposits in France

The uranium deposits in France, for example, can be classified as three types: vein deposits in granites,
mineralisations associated with sedimentary formations and those, rarer, associated with metamorphic
rocks (Devilliers et al., 1980).
The intragranite vein deposits occur mainly In the Hercynian formations of the Massif Armorican
(Vendée), especially in the Massif Central. They are associated with a granite rich in potassium
(leucogranite), light In colour due to Its acidic composition. Mineralisation occurs as vein fillings in
fractures and as columnar-shaped bodies resulting from hydrothermal alteration. The accompanying
mineralisation is dark-coloured pitchblende In vein deposits and bright-coloured alveolar episyenite in
larger mineralised bodies. In the latter, quartz grains are completely replaced by uraninite with
concentrations reaching up to 10 to 100 kg of uranium per ton (U/t) In exceptional cases.
The most classical sedimentary ore deposits are situated in the peripheries of granite massifs,
which constitute the source. These deposits often occur In sandstone formations, red In colour, which
are, however, buried under more recent sediments. They are hence referred to as concealed deposits.
Mineralisation in them is pitchblende or coffinite in disseminated form (1 g to 3 kg U/t). Lastly, some
low-grade marginal deposits occur in the Palaeozoic schistose metamorphic rocks (in Bretagne) affected
by Hercynian granitic intrusions.
In other parts of the world, this classification extends to other types of deposits such as ‘roll-fronf
of Wyoming andTexas (Raines et al., 1978).The important point to note is that the specific characteristics
associated with leucogranitic vein deposits are also observed everywhere outside France. The contrasts
In geochemical composition, structural state and colour may hence constitute key guides for
interpretation of remote-sensing data, which are applicable in all latitudes when vegetation cover is
low or negligible.

24.1.3 Regional metallogeny and metalliferous provinces

As mentioned above for uranium, the conditions favourable for occurrence of source-transport-deposition
system for any mineral deposit are not the result of purely random processes. They are always logically
related to a particular type of geology whose tectonic history per se is related to the major episodes of
evolution of the Earth’s crust.
 Applications 403

A study of all such causes and their effects, combined with a systematic inventory of previous
discoveries, leads to a model of spatio-temporal distribution of deposits and delineation of boundaries
of regions in which they occur. Existence of such boundaries gives rise to the notion of metalliferous
province and the basic concept of regional metallogeny.
The regional disposition of the French uranium deposits (Fig. 24.2; CD 24.1 ), for example, follows
a west-east curved zone produced by Hercynian orogeny. The axis of this structure corresponds to the
Averno-Vosgien cratonic nucleus. However, association between métallogénie province and ore deposits
is not unique. Most often, similar favourable conditions may correspond to any metallic mineral
assemblage.

Fig. 24.2: Distribution of uranium deposits in France, compared to the geometry of the Moldanubian (Variscan
orogeny) and Vendéocénévole (Mesozoic) provinces.

For example, a geographic relationship similar to that observed for uranium can be drawn for the
Pb-Zn deposits associated with the Hercynian chain. The latter can be extended divergently towards
south-west, combining the Armorican and Iberian peninsulas Into a single metalliferous province,
Including the Pb-Zn deposits of the Zamora region in Spain (Marconnet, 1987). It thus follows that
techniques of application of remote sensing are similar in both cases, the apparent geographic and
climatic diversities notwithstanding.

24.1.4 Metallotect; an operational concept

Although the concepts of metallogeny are essential for guiding systematic mineral exploration, they
only provide theoretical models that are often least related to the practice of prospecting.
In fact, prospecting methodology is more a combination of multiple techniques of investigation
such as geophysics, geology, geochemistry and remote sensing, than really a theoretical approach.
The prospector analyses a sequential series of diagnostic features based on a large amount of data
gathered from the field. The concept of metallotect is introduced to answer to this empirical approach
of mineral prospecting (Routhier, 1969).
404 Processing of Remote Sensing Data

The term metallotect is defined as any geological, geochemical, tectonic, etc. object or phenomenon
which favours or contributes to the formation of a mineral deposit or concentration. This concept
particularly facilitates proceeding from pure genetic models to more flexible models such as those
introduced by geostatistics. In geochemistry, this leads to the concept of paragenesis, assemblage of
chemical elements accompanying the metallic compound explored depending on the type of deposit.
In remote sensing, the term lineament-metallotect is used to denote the factors that may relate
the density of lineaments traced on an image and localisation of deposits in the zone of observation.
Thermal metallotects can be constructed if certain properties of fluid circulation in mineralised zones
are related to variations in their surface temperature. Remote-sensing metallotects are generally based
on the one hand, on the principle of synergy of forms (ex. lineament-lineament or lineament-contact
intersections) and on the other, on the detection of ‘transparent zones’ that reveal the object under
exploration If it is masked by plant cover, soil or other surface perturbations (Marconnet et al., 1982).
This transparency can also be Inferred from correlation with other factors such as geochemical or
geophysical.
The work of the prospector thus lies in combining the metallotects to guide his decisions of
exploitation or exploration. Systematic methods of decision making based on the principle of integration
have also been developed (Leymarie et al., 1982).

24.2 MULTISCALE APPROACH IN MINERAL PROSPECTING

Mineral prospecting is essentially a multiscale operation. It is organised in successive phases that
tend to focus on mineralisation. It starts at regional level based on purely métallogénie reasoning,
gradually descending Into detail and finally ends with the discovery of an object which most often is
very limited in volume. Remote sensing, by its synoptic view from space, constitutes one of the best-
suited techniques for this approach.

24.2.1 Multiscale structural model in remote sensing of vein

deposits
Remote sensing is often suitable for understanding the multiscale structural phenomena, in particular
in studying lineaments and their association with localisation of deposits.
Most bedrock mineral deposits are fractured as in the case of uranium deposits of Vendée and
the Massif Central and the Pb-Zn deposits of the Zamora province in Spain mentioned earlier. Several
successive episodes have led to mineralised concentrations, which make the deposits economic.
Mineralisation traps often occur in major shear zones such as in Margeride, the Vendée massif or
in the uraniferous district of La Crouzille. Sequences of permeable breccia, lamprophyres forming a
screen and open fractures serving as drains— conditions favourable for occurrence of episyenite— are
observed In such zones. Remote sensing provides a wide scope for efficient exploration of structural
traps due to its multiscale approach, ranging from satellite images to field Investigations (Monget,
1982).
The wide range of scale of all such observations enables classification and characterisation of
significant tectonic objects according to their geographic association with mineralisations:
— Orbital Images ranging from 1:1,000,000 to 1:50,000, with a resolution of 30 m for LANDSAT
satellites and up to lO m forS P O T or even betterfor 1RS, Indicate faults, fractures and lineaments, 100
km to less than 1 km In size. They provide coverage of large parts of a métallogénie province and
reveal major fault zones often organised around transform faults (Campredon et al., 1980).
— Aerial photos, side-looking radar and high-resolution orbital observations enable structural
investigations in 1:50,000 to 1:5000 scales, with a resolution of 1 m to 50 cm. They facilitate detection
 Applications 405

of joints and fractures, varying in size from a kilometre to about 10 m, and echelon or curved patterns
that mark tectonic nodes or lineament networks. Together with lineament-density maps, interpretation
aids in locating favourable zones (Gilli, 1985).
— Detailed geologic mapping, in scales from 1:10,000 to 1:1000, clearly reflects faults and fracture
zones of about 100 m in size. They enable precise location of metre-size objects and detection of trap
structures (Rousselin, 1985).
— Lastly, opening of a mine or quarry in a 1:200 or higher scale facilitates mapping of objects of
a decimetre or even a centimetre in size. Detailed sketching of joints, complex and braided joint systems,
braided faults and open fissures is done at this stage. It enables working at the minute geometric level
down to mineralisation traps.
Geometric comparison between lineaments and structural features indicated by these various
observations depicts similarities as well as lack of correlation from qualitative, quantitative and directional
points of view. In particular, major lineaments in satellite images often have no direct equivalents on
the ground but their directions may indicate regional-scale fault structures. This demonstrates the
utility of an extensive range of investigations.

24.2.2 Example of regional application of structural

interpretation
We have seen earlier that the uranium deposits in France are distributed in narrow provinces associated
with geographic localisation of Hercynian massifs. It is however possible to construct indicator
metallotects at this level of global perception based on detection of linear and narrow features, i.e.,
lineaments, which distinctly differ from their surroundings.
This approach can be developed In particular based on observations by TIROS-AVHRR type
meteorological satellites. In the example of CD 24.2, thermal and near-infrared data are used to
delineate lineaments extending over distances exceeding hundred km. Changes In these contrasts
depending on thermal inertia of formations and variations in the vegetation coverage during a year
can be investigated based on diachronic comparison between summer and winter images.Thus maps
depicting geographic relationships between lineaments and existing mineral deposits can be prepared
to delineate zones favourable for occurrence of yet to be discovered mineralisations (Fig. 24.3; CD
24.1) (Durandau et al., 1982).

24.3 REMOTE SENSING: AN INDIRECT METHOD

Geological phenomena are essentially subsurface and with a few exceptions, mineralisation per se is
rarely detected on the surface. Therefore, in mineral prospecting remote sensing is an Indirect
geophysical technique.
It is no less significant that remote-sensing data can delineate indicator metallotects either as
structural forms or paragenetic systems accompanying mineralisation, whose expression may manifest
on the surface or below a thin cover. Important remote-sensing parameters in this context are reflectance
and temperature. Use of microwave imagery for this purpose is also described in Chapter 26.

24.3.1 Reflectance properties

■ Direct paragenetic detection: hydrothermal alteration
Hydrothermal alterations represent modifications in physicochemical properties of minerals and hence
of rocks, due to the hot-water flows associated with volcanic eruptions or crystallisation of magma.
406 Processing of Remote Sensing Data

Mille\faches plateau
\ ^ .

Day/night thermal lineament \ m

^ \

Day thermal lineament \ *

> / \

m Known uranium deposits \

-* " s Boundaries of major granitic massifs

Fig. 24.3: Relationship between regional thermal lineaments and uranium deposits in the western
region of the Massif Central.

Hydrothermal solutions contain various metals such as iron (Fe), titanium (Ti), copper (Cu), lead (Pb),
zinc (Zn), tin (Sn) and uranium (U). Various forms of hydrothermal activity are recognised with specific
composition of metal assemblages, and iron in the oxidised form often plays the role of an indicator
element. When conditions are favourable, these metals may concentrate as economic deposits.
The exposed zone of such a metallic deposit Is often oxidised and forms a ‘gossan’. On the
surface it depicts the form of an aureole and yellowish to brownish-red hues characteristic of the
presence of haematite. Thus such zones can be recognised in multispectral remote-sensing data
containing bands sensitive to these oxide compounds (band 7 of TM) or very narrow bands (see
Chap. 4).
Alterations of crystalline or volcanic host rock at the contact of veins occur in the form of friable
clays and distinct colours. Formation of white kaolinite In the alteration of leucogranites rich in potassic
feldspars and Impoverished in ferromagneisan minerals is a typical example. Kaolins of the Brame
massif around La Crouzille in north of Limoges are well known for their use in fabrication of porcelains.
Most geological materials produce their maximum reflectance at 1.6 pm (Hunt et al., 1979). The
suite of minerals often associated with hydrothermal alterations is characterised by a maximum
reflectance at 1.6 pm and a marked absorption at 2.2 pm. This paragenetic association includes
dioctahedral phyllosilicates (kaolinite, montmorillonite, etc.), some sulphates (gypsum, alunite) and
aluminium oxides (diaspore, gibbsite).
 Applications 407

Absorption in these minerals is caused by vibration processes and inner electrons of the molecular
structure. Between 0.35 and 1.3 pm, electron transitions of minerals rich in iron, such as haematite,
goethite, montmorillonite and jarosite, show characteristic features marked by minima in the wavelength
bands of 0.43, 0.65, 0.85 and 0.93 pm. The minima in the interval 1.30 to 2.5 pm are more particularly
related to the molecular structures of alunite, kaollnite, montmorillonite, pyrophylllte, potassic micas,
diaspore, gypsum, jarosite and carbonates (see Chap. 23, Fig. 23.13).
Detailed observations in the position of absorption or reflectance peaks enable identification of
Indices on the content of alterations. For example, occurrence of a very narrow but low-amplitude
band near 0.43 pm Indicates jarosite. An intense minimum centred around 0.85 pm represents haematite
whereas a similar characteristic centred on 0.93 pm indicates the presence of goethite. Thus, if the
sample observed contains haematite or goethite, a band will occur with a minimum between 0.85 and
0.93 pm. On the other hand, development of an additional band appearing as a shoulder of the
reflectance maximum at 0.75 pm and centred on 0.65 pm is more pronounced for samples rich In
goethite rather than for those with high haematite content.
For clay minerals and alunite (Fig. 23.13; CD 24.1), characteristic absorption bands are situated
more distinctly in the near infrared zone, at 1.4 and 2.2 pm respectively. However, an entire range of
variants exists, some of which can be used to recognise constituents of alteration clays. Intense
minima, narrow and single, at 1.4 and 2.2 pm indicate presence of pyrophylllte and potassic micas
whereas dual peaks of absorption at the same position rather indicate existence of kaollnite (Fig. 24.4;

Fig. 24.4: Reflectance curves of minerals associated with hydrothermal alteration of rocks.
408 Processing of Remote Sensing Data

CD 24.1). For montmorillonite one more distinct minimum appears at 1.9 pm; it is also known as the
‘water band’. Gypsum has a characteristic similar to that of alunite, but at 1.75 pm.
The spectral contrasts of pure samples are not generally observed in real conditions of existing
mineralisations. For example, if hydrothermal alterations associated with gold deposits of Beaver Creek
(Colorado) and Goldfield (Nevada) are considered, alunite is encountered as a major constituent of
most altered samples collected. It is hence often difficult to associate a sample with a particular type
of alteration (sllicified, opalised, argilitic and phyllltic) based only on the form of its spectrum even If
high-resolution observations, such as those of airborne hyperspectral imagery, are available. This Is
an obstacle in determining from distance the paragenesis of the anomaly observed and hence an
uncertainty on the type of concentration. The latter can be established only from a visit to the field
accompanied by precise geochemical measurements.
The observations mentioned above on the form of spectra have led to development of ratioing
techniques or principal component analysis, which enable quantitative estimation of constituents (Lee
et al., 1990). Absorption properties specific to such alteration products have guided the choice of band
7 (2.08 to 2.35 pm) for the LANDS AT Thematic Mapper.

■ Indirect spectral detection of geologic formations

Mineralisations, or the formations that host them, are not always associated with well-defined spectral
characteristics, especially when they are covered by a sufficiently thick soil or dense vegetation.
However, favourable conditions may exist wherein a particular association of geologic bedrock-
soil-vegetation has spectral contrast and thus enable indirect mapping of a formation (Siegal et al.,
1977).
In the leucogranitic massif of Saraya In eastern Senegal (Blot, 1980), ferruginous soil masks the
nature of underlying formations and renders their direct mapping by remote sensing difficult.
Nonetheless, when the observations of the region by LANDSAT and CZCS (CD 24.3) are analysed, a
distinct contrast Is noticed between the soils covering the massif and those on its periphery. In the
case of CZCS, highest contrast is observed for the first spectral band (blue— 433 to 453 pm). This
band is particularly sensitive to the concentration of quartz, which is very high in the surface soil (Fig.
24.5; CD 24.1), and reflects the influence of a granitic body situated at depth.These ‘colour’ contrasts
are almost not visible on the ground which is flat in most cases. Thus an effect of ‘transparency’ is
observed.

0m

1m

2m

3m

Granite

A-horizon Iron pan I Gravelly-clay horizon

Fine gravel | | Granite I Porphyroblast

Fig. 24.5; Schematic cross-section of surface formations over Saraya granite.

 Applications 409

At a higher level of detail in the LANDSAT MSS images (CD 24.3), structural passages can be
noticed running across the massif, which have lower values of digital numbers. These zones correspond
to dolerite veins, known for their association with vein-type uranium mineralisation of economic size.
On top of dolerite veins, which are more clayey, soil is less silicious, thus indirectly making these
formations transparent.
Although it is not possible to replace deep-level studies by surface investigations, it is seen that
the interpreter of spectral contrasts should always be on the lookout for anomalous zones because
they may indicate deep causes. This, of course, is possible only in zones where the medium Is In
natural equilibrium without anthropogenic disturbances. Examples of such unfavourable anthropogenic
phenomena are land patches of annual burning in Africa.

24.3.2 Use of thermal band

Measurement of surface temperature of rocks or soils (see Chap. 26) by airborne or spatial observations
are rarely used instantaneously. At least, its interpretation always takes into consideration the regime
of excitation energy to which the body under observation is subjected, in view of diurnal variation of
solar radiation. Moreover, knowledge of temperature variations of a rock during day and night enables
determination of Its thermal inertia. This volumetric property is of the same kind of intrinsic nature as
other geophysical measurements such as electric conductivity, density, etc.
Computation of thermal inertia necessitates on the one hand, sequential observation of
temperatures during a diurnal cycle and on the other, use of a thermodynamic model to describe
thermal phenomena inside the solid medium (see later, CD 24.4).
Diurnal temperature variation of water is of low amplitude since It has very large thermal inertia.
Soils, on the opposite, show large amplitudes of day/night temperature variation. For rocks and dry
soils, inertia is small and amplitude is as high as the soil Is porous. As the moisture content of soil or
fracturing of rock increases, it gets saturated with water. Its thermal inertia increases and amplitude
decreases (Fig. 24.6). In the case of chlorophyllian vegetation, self-regulation by évapotranspiration
(see Chap. 1) also reduces temperature deviations.
Contrasts between a moist fractured medium, such as a granitic massif, and a dry and less
consolidated medium, such as altered crushed zone where quartz sand is abundant, are very important

Fig. 24.6: Theoretical curve of diurnal temperature variation of soil and water body for
continuous diurnal insolation.
410 Processing of Remote Sensing Data

in mineral prospecting. Such contrasts are especially more distinct during night. Zones of high thermal
inertia in such cases appear hotter than their adjoining areas (Rousselin, 1985).

H Use of night thermal contrasts at regional level

Night observations can be primarily made by small-scale satellites such as AVHRR or HCMM and
used to delineate entire massifs such as the leucogranites of La Crouzille district, which are situated
in North-Limousin.The granitic massif of this region consists of two main formations, the Brame granite
in the west and the Saint-Sylvestre granite in the east. The Brame granite is a biotite-granito-gneiss,
bounded on the west by the Nantiat fault. Its general orientation is NNE-SSW. The Saint-Sylvestre
granite is a leucogranite with two micas. It occupies the entire zone south and east of La Crouzille (Fig.
24.7; CD 24.1).
Mineralisation is localised in subvertical tectonic zones, about a metre in width, filled by crushed
granite. It occurs essentially in the form of autunite and pitchblende, that is, as phosphates, silicates
and oxides. Three principal factors are responsible for mineral concentration. Firstly, dense tectonic
activity associated with extreme alteration of contact zones results In stockwork deposits, often workable
as opencast mines but of low grade. Secondly, fault intersections with lamprophyre veins give rise to
localised mineralisations (veinlets) of high-grade pitchblende and, lastly, fault intersections with
micaceous episyenite lenses, whose voids are filled with pitchblende, localise accumulations of high-
grade ore.

JT 7T Bioiite-granite [ | Fine-grained granite with two m i c a ^ ^ ^ Faults

XXXX . Granite-gneiss □ Leucogranite Metamorphic formations

Fig. 24.7: Schematic geologic map of Saint-Sylvestre massif (after Marquaire and Moreau, 1969).
 Applications 411

Considering all these characteristics, uranium prospecting in the La Crouzille sector is guided by
two main principles: firstly, strategic exploration of favourable source rocks, viz., leucogranites, and
secondly, delineation of fracture zones favourable for mineralised fillings. Remote sensing in this sector
followed this approach (Durandau et al., 1982).
Satellite observations of night time thermal contrasts by HCMM (CD 24.4) showed that the Saint-
Sylvestre leucogranite is marked by distinctly higher temperature than that of its surroundings. This is
an indication of higher thermal inertia. This factor is probably due to a more open fracturing of this
formation capable of containing larger quantities of water.
Satellite measurements were confirmed in detail by airborne thermal measurements (CD 24.4).
The latter showed identical regional thermal contrasts and indicated details of vein formations of
about a metre in width. They depicted differences in night temperatures between massive, cold,
lamprophyres and fissure zones in which fluids circulate, which are indicated as hotter zones.

H Night time aerial thermal measurements

Utility of night time aerial data with ARIES scanner (Tabbagh, 1973) obtained with a ground resolution
of 10 m is demonstrated with a few examples (CD 24.5). All the images of our example pertain to
Quercy In the southern boundary of the Permian Brive basin, zone known for its prospects of uranium
deposits. Major faults such as Meyssac or Padirac control and delimit potential zones of mineralisation
in the Permian sedimentary formations (Daudon, 1987). One can particularly notice the Argentât fault
zone occurring in the cold and dark region, the red sandstone of high thermal Inertia in Brive region
appearing as bright and the Liassic formations, clearly delineated. Lastly, many sinks and avens
indicating the Gramat karst plateau, which are detected as cold zones at night, appear as numerous
dark spots.
Interpretation of night time thermal Images, however, is also beset with some problems. Special
mention may be made of meteorological effects leading to accumulation of cold air In valleys. Such an
effect is observed in the thermal images of the Martel region (CD 24.5). It can mask lithological contrasts.
Night time thermal scanning is one of the most important tools for mineral prospecting (Scanvic,
1983) and detailed geologic mapping since it reveals inertial properties of rocks, which result from the
nature of rocks and may also indicate effects of heat transport from deep sources.

M Monitoring complete thermal cycle and thermal inertia mapping

Determination of thermal inertia, which requires acquisition of day as well as night data, limits practical
utility of these techniques to only meteorological satellites. The low resolution of the latter, 1 km for
AVHRR and 4 km at nadir for METEOSAT, narrows down their applicability to only regional metallogeny.
Let us take the Bandiagara massif in Mali (Daveau, 1959) as an example. This sandstone formation
constitutes a slab sandwiched between the Niger delta in the west, towards which it dips with a mild
slope, and the sandy depression of Gondo in the east over which It stands as a cliff.
The LANDSAT MSS data Indicate (CD 24.5) striking spectral homogeneity of the massif. This is
due to the uniform coverage of ail outcrops by ^desert patina’ rich in iron oxide. Observation of contrast
between day and night temperatures (CD 24.5) for the same formation clearly shows, the low resolution
notwithstanding, a zone situated in the northern part of the massif for which thermal inertia is much
lower than that of plateau sandstones.
This thermal-inertia map depicts the true phenomenon of transparency across the layer of desert
varnish that masks any lithological difference in the visible and near-infrared bands. Referring to the
geological map (Fig. 24.8; CD 24.1), the zone of high inertia corresponds to an iron rich pan covering
dolerite intrusions, probably mineralised. The complementary nature of spectral data in geological
mapping is evident here (Bardinet et al., 1982).
412 Processing of Remote Sensing Data

Sandstone plateau

Sandstone plateau covered

by iron pan

Sandstone plateau
overlying schists

Sandstone plateau with

sandy cover

Iron pans over dolerites

Iron pans

Fig. 24.8; Geological map of Bandiagara plateau (after Blanck, 1982).

24.4 CONCLUSION
Apart from a few results obtained in mapping mineralised alterations in desert zones, it can still be
considered that remote sensing in general is not a tool for direct detection of minerals. Its principal
advantage, however, lies in providing a synoptic base map of major lithological units, morphological
forms and structures.This information becomes very useful in mineral prospecting when it Is integrated
with other types of ground data such as geological, geophysical or geochemical.
The quality of geological information given by remote sensing depends on the mode of
instrumentation, type of terrain observed (plant cover, soil, rock outcrops, climate), method of processing,
techniques of image manipulation adopted and, lastly, degree of understanding of the physical
phenomena Involved in the measurements. Hence It is a method in which the expertise of the interpreter
plays a major role. Remote sensing can succeed only in a multidisciplinary structure in constant
liaison with ground observations. It Is no less significant that Its capability in covering large zones of
exploration makes it an essential variable for identifying regional metallotects used in economic models
of prediction and development of natural resources.
In the future, enhancement in spatial and spectral resolution of remote-sensing instruments ought
to enable more precise detection of mineral components of rocks if only they are exposed. The
techniques of interpretation will then become similar to those of geochemistry and probably emerge
as more quantitative methods. It would then be possible to achieve direct detection of mineral deposits
when vegetation cover is not too dense.
 Applications 413

References
BardinetTC, Monget J-M, Patoureaux Y. 1992. ‘Combined use of daily thermal cycle of METEOSAT imagery and
multispectral LANDSAT data: application to the Bandiagara plateau, Mali’. Proceedings of an EARSEL-ESA
Symposium, Igis. Austria, Apr. 1982, ESA Technical Publication SP 175: 95-101.
Blot A. 1980. L’altération climatique des massifs de granite du Sénégal, Travaux et Documents de i’ORSTOM, 114:
434.
Campredon R, Monget J-M, Simon P. 1980. Mise en évidence par télédétection d’un accident à rejeu récent dans
le bassin permien de Luc-en-Provence. Comptes rendus de l’Académie des Sciences, Paris, Série D, 291:
55-57.
Daudon P. 1987. La fracturation du socle en Bas-Limousin et Quercy et ses répercussions dans la couverture
sédimentaire Thèse d’État, Université d’Orléans, 215 pp.
Daveau S. 1959. Recherches morphologiques sur la région de Bandiagara, Mémoires de l’Institut français d’Afrique
Noire, IFAN, Dakar, 56:119.
Devilliers JP, Ziegler V. 1980. L’industrie minière en France. Sa situation au début de 1980, Annales des Mines, 7 -
8:123-134.
Durandau A, Monget J-M. 1992. ‘Interpretation of satellite and airborne thermal measurements gathered over
hercynien granites (France) in relation with uranium deposits’. Proceedings of the International Symposium
on Geology of Granites and their Metallogenetic Relations, Nanjing, China, Oct. 1982, pp. 611-619.
Gilii J. 1985. Analyse numérique de l’image radar du secteur Port-Gentil Azingo Lambaréné Ouest: contribution à
l’étude géologique du bassin sédimentaire gabonais. Thèse de 3® cycle en Géologie, Géochimie et Techniques
Avancées, Université de Nice.
Flunt GR, Ashley RP. 1979. Spectra of altered rocks in the visible and near infrared. Economic Geology, 74:1613-
1629.
Kruse FA, Kieren-Young KS, Boardman JW. 1990. Mineral mapping at Cuprite, Nevada with a 63-channel imaging
spectrometer Photogrammetric Engineering and Remote Sensing, 56 (1): 83-92.
Lee JB, Woodyatt AS, Berman M. 1990. Enhancement of high spectral resolution remote sensing data by a noise
adjusted principal components transform. Geoscience and Remote Sensing, 20 (3): 295-304.
Leymarie P, Baelz-Maniere S, Durandau A, Monget J-M, Sinding-Larsen R. 1982. ‘Un système d’aide à la prospection
minière en prospection de l’uranium’. C.R. du Symposium’ AEN/AIEA de R&D: Méthodes de prospection de
l’uranium, Paris, Juin 82, pp. 889-907.
Marconnet B. 1987. La télédétection spatiale, une méthode pour la prospection minière stratégique et tactique.
Application aux gisements stanno-wolframifères de l’Ouest de la péninsule ibérique. Thèse en Pétrologie
Structurale et Métallogénie, Université de Nancy, 244 pp.
Marconnet B, Gagny C, Leymarie P, Monget J-M. 1992. ‘Phénomène de transparence d’un leucogranite à étain-
tungstène sous couverture métamorphique (province de Zamora, Espagne)’, Actes du Symposium International
de la Commission VII de la SI PT. Toulouse, Sept., 82, Éd. GDTA, Int. Archives of ISPRS, 24-VII, 1, pp. 523-
531.
Monget J-M. 1982. ‘Télédétection multiscalaire des provinces uranifères françaises’, C.R. du Symposium AEN/
AIEA de R&D: Méthodes de prospection de l’uranium, Paris, juin 82, pp. 121-128.
Pelissonnier H. 1976.‘Classification par types en métallogénie’, In Métallogénie et prospection minière. Mémoire
Hors Série de la Société Géologique de France, 7:277-283.
Raines GL, Offield T, Santos ES. 1978. Remote sensing and subsurface definition of facies and structure related to
uranium deposits. Powder River basin, Wyoming, Economic Geology, 73:1706-1723.
Rousselin T. 1985. Télédétection thermique des granites uranifères du Nord-Limousin (France), Thèse de 3® cycle
en Géologie, Géochimie et Techniques Avannées, Université de Nice.
Routhier P. 1969. Essai critique sur les méthodes de la géologie, Masson, Paris, 202 pp.
Scanvic J-Y. 1983. Utilisation de la télédétection dans les sciences de la terre, BRGM, Manuels et Méthodes, 7:
158.
Siegal BS, Goetz AFH. 1977. Effect of vegetation on rock and soil type discrimination, Photogrammetric Engineering
and Remote Sensing, 43 (2): 101-196.
Tabbagh A. 1973. Essai sur les conditions d’application des mesures thermiques à la prospection archéologique.
Annales de Géophysique Fr., 29 (2): 179-188.
 _______________________________________________________________________________________________________^

Remote Sensing and Coastal-zone

Management
The objective of this chapter is to give several examples specific to coastal environments, based on
the Investigations carried out in France and in particular at the IFREMER.

25.1 INTRODUCTION: GENERAL PROBLEM

Land cover and land use of coastal regions are today undergoing rapid changes, more particularly in
countries producing and exporting petroleum or mineral products. Other modes of coastal exploitation
have also become preponderant. Aquaculture is fast developing as is evident from the fact that 200,000
ha of ponds have been constructed in fragile zones on the Philippine coast. Similarly, tourism has
increased in the Ancient World and some island states of the Caribbean, Pacific or Indian Ocean, in
the near future tourist activity will become world’s premier industry in terms of volume of business.
In almost ail countries with a maritime front, there is migration of populations to the coasts: 2/3 of
world’s population lives within 400 km from the sea. The impact of such activity Is represented by:
— development of infrastructure: ports, airports, industrial, mining and energy complexes, offshore
petroleum production, agro-nutritlonai transformation, urban complexes, hotel and tourist Infrastructure,
aquacultural Installations, etc.
— transformation of natural zones: destruction of wet zones, reefs, natural vegetation in back
lands, rectification and protection of coastline, various types of pollution of marine environment, increase
in ecological hazards, etc.
Increasing awareness of such hazards and the need for more efficient management are
experienced at all levels and are concretised by co-ordinated management procedures which most
often are supported by legislation:
— in France, schemes for economic management of seas (SMVM), more recently territorial
management directives (DTA),
— in the United States, the Coastal Management Act, Middle Atlantic Coastal Resource Council,
— at international level, programmes such as COMAR (UNESCO), Regional Seas (PNUE), etc.
The European Union, concerned with this problem very early, launched specific experiments on
information and integrated studies on coastal zones.
Such legislations and procedures necessitate employment of planning and management products
derived from multiparameter Information systems for which remote sensing constitutes a source of
data.
Coastal zones present the difficulty of linear type of extension with a land-sea interface and Its
adjacent zones in all countries, independent of borders. This interface is under the jurisdiction of
several administrations whose prerogatives are often opposed to one other. Further, like in applications
described in other chapters, information furnished ought to be multispatial, multlparametrlc and
 Applications 415

multitemporal. Thus in metropolitan France for example, the basic geographic maps pertaining to
littoral zones are of various types, various scales and adopted to various planes of projection: ex.
topographic maps of the National Geographic Institute, marine maps of the Naval Oceanographic and
Hydrographic Service.
In many tropical-zone countries, very few precise and up-to-date cartographic systems are available,
particularly for sensitive and inaccessible regions such as mangroves and coral reefs.

25.2 SPATIAL OBSERVATION OF COASTAL ENVIRONMENTS

25.2.1 Littoral and coastal-zone objects
It is difficult to define what constitutes the littoral zone since the phenomena pertaining to this interface
are both terrestrial and marine.
From an administrative point of view, the extension of the coastal zone on the upstream, coastland,
side can be limited to the border of littoral communities and on the downstream (sea) side, to 12
nautical miles. Based on environmental criteria related to water exchanges and flows. It can be
considered to extend from the Inland boundaries of watersheds up to the zone of extension of these
flows into the sea. On the geographic and topographic basis, it can range from a land zone of a
defined elevation (200 m for example) to the border of continental plateau (bathymetric limit of -200
m). For the following discussions, we limit ourselves to the ‘marine littoral zone’, i.e., to the zone
bounded on the land by high-tide mark. Two types of environment can be distinguished in this zone:
— subtidal, situated below low-tide level; venue of influences, often combined with one another,
emerging: (a) from the surface (sea swell, waves, currents, winds, pollutants such as hydrocarbons,
ice, etc.), (b) from the water column (turbidity, temperature, discharge plumes, water blooms, plankton,
etc.), (c) from the bottom (shallow areas floors observed in Clearwater in optical domain, themselves
consisting of various substrata, hard or soft, with or without vegetation, emerging from animal
constructions such as coral reefs and with influence of the sea bottom on surface currents, detectable
in the microwave band, etc.).
— intertidal, situated in the balance zone of tides, comprising bare soils (sands, mud, rocks, etc.),
vegetation (plants in swamps such as mangroves, marine halophytes, algae, microflora, etc.), animals
(mollusk banks, coral reefs, etc.), managed zones (oyster parks, basins, canals, breeding ponds, salt
pans, etc.).
In addition to a general linear disposition, the littoral zones are characterised by small-size features
constituting highly varied landscapes whose structure results from natural or artificial processes.
Lastly, the constituents of the littoral zones are highly fluctuating with varied periodicity, random
(storms for example) or periodic (hydrological regimes, phenological regimes, tides, etc.).
Considering these ¡features, it is pertinent to examine whether the Earth-observation systems
possess adequate spatial resolution and observational periodicity to provide, on operational basis,
data necessary for correct characterisation of surface phenomena in littoral zones, as complementary
to conventional ground measurements.

25.2.2 Littoral objects and specifications of aérospatial

observation systems
Preliminary analyses of adequacy between characteristics of coastal environments and technical
performance of aérospatial observation systems (Klemas et al., 1980; Gierloff-Emden, 1982; Loubersac,
1983; Klemas et al., 1987) have shown that present-day remote-sensing systems are not correctly
416 Processing of Remote Sensing Data

suited for coastal-zone studies. There is incompatibility between high spatial resolution and high temporal
resolution required for these observations. This is a major drawback hindering the application of high-
resolution remote sensing to coastal Investigations that do not involve dynamic phenomena.
This incompatibility is Illustrated in Fig. 25.1. An analysis of the diagram (logarithmic scales) of
temporal resolution expressed in days (abscissa) and spatial resolution expressed in metres (ordinate)
offered by major Earth observation systems leads to the following inferences.

Spatial
resolution (m)

M/T/OSAT(IR)

NOAA-TIROS
SEAWIFS

ENVISAT/MERIS

LANDSAT(MSS)

LANDSAT(TM)

SPOT 1-4 (XS) ERS (SAR PR1)

SPOT 1-4 (P)ADEOS ■

IRS-SPOT 5

ORBVIEW3

QUICKBIRD

Fig. 25.1: Possibility of detection of marine and coastal phenomena by major sensors and platforms of present-day
Earth observation satellites (adapted, completed and updated after Klemas et al., 1987, courtesy IFREMER).

The spatial resolution (tens of metres) of systems such as SPOT or ERS enables with difficulty
operational characterisation of some of the human activities in littoral zones and certain natural
phenomena such as coastal erosion (except the specific case of strong sedimentary dynamics). On
the other hand, the very high resolutions of SPOT-5 (5 m) or Orbvlew-3 (4 m in multispectral and a
metre in panchromatic) and Quickbird (3.3 and 0.8 m respectively) are closer to the resolution
requirements for most applications in coastal zones.
 Applications 417

The optimal temporal resolution, extremely variable depending on the application, is not satisfied
by the presently existing satellite systems (better than 3 days with ERS for typical coastal studies).
Thus operational applications of remote sensing to tidal phenomena, monitoring of dispersion due to
currents and winds (dispersion of wastes, resulting pollutants, etc.) or navigation control are most
often limited.
Aerial platforms are better suited for coastal studies since they provide ground resolution and
revisit frequency higher than those of today’s high-performance satellites and permit a greater number
of onboard sensors. However, the disadvantages of aerial systems are limited synoptic view, high
costs of acquisition with Increasing repetition, atmospheric constraints and tedious procedures of
geometric processing of data.
Operational characteristics of the major satellite or airborne sensors for various studies of coastal
zones are summarised in Table 25.1 (adapted after Klemas et al., 1987). Use of certain platforms for
some sensors is not technically feasible and hence not indicated. It follows from the Table that optimal
observation for monitoring coastal phenomena necessitates use of almost all the spectral bands and

Table 25.1: Operational characteristics of remote-sensing systems for coastal studies

c 0 0 CO
c o c CO
0
Ü 0
C E c 0
O c 3 0 o
■D
c c
o

Ì
0 c '0 _C0 0
O o 0
o
-Q o E 0
Sensor 0 C T3 0
U) CL
o 0 E O 0
3 C
Ò 0 C T3 E c
E Ü
T3 o 10
CO 0
0
0
> 0
i5 o C ■¿0
0 C 0 o T3 ■O 0 CO
CL
CD .C o CO 0 c Ü
CL >> o T3 0
CL CO CO
TJ O 0 I o S0 C 0 t:
0 o 3 O) CL 0 13
■o O 0 CO CO
JZ
o > 3 E
O CO o
13 t5
CO

H m ■ B
m0
Radar altimeter S 0 0 0 0 0 0 0 0 0 0

Profiling lasers A 0
m 0 0 ■ 1 0 1 0 0 0

Laser fluorescence A 0
B 1 1 B 1 1 0 0 1

a ^ ■ m B B B mm■
Wm B
Photography A 1 2 0 0

S 1 2
mm 0 0

Multispectral A 1
mm m mB a mmQ 0 0

scanners S 1
m m B B mm a 0 0

1 a m m mm ■
m
Imaging A 0 1 0 0 1 1

radar S 0 1 0
mm 1 0 1 1 1

IR thermal A a m m
B
1 1 0 0 0 1 0 1

scanners S 1 1 0 0
■ 0
m 1 1 0 1 0

Passive microwave A 1 1 0 1 m 1 1 1 0 0 ■
radiometers S 1 0 0 0 1 1 0 0 0 0 0

Radar scatterometer S
B 0 0 0 0 0 0 0 0 0 0 0

A— Aeroplane; S— Satellite; 4— Operational; 3— Truly potential but experimentation required; 2— Potential utility;
1 — Limited utility; 0— Not usable
418 Processing of Remote Sensing Data

all types available sensors, without however enabling presently reliable measurement of an important
hydrological parameter such as salinity. An optimum but technically difficult solution would be to combine
a multispectral scanner, a thermal scanner, a radar (especially Imaging one) and a microwave
radiometer, all with a spatial resolution of less than 5 m and a revisit capability of a day or less, on a
single platform or on different platforms operating in phase.
The presently available multispectral data of optical and thermal bands and the SAR type radar
offer great potential for a number of applications, some of which are given below.

25.3 SPECTRAL CHARACTERISTICS OF LITTORAL

OBJECTS
Spectral responses of a water layer are given in Chapter 4 for optical domain and In Chapter 1 for
thermal infrared and microwave bands. Examples of application of thermal and microwave data are
discussed in Chapter 26. Spectral characteristics of objects of the maritime region in intertidal and
subtidal zones are described below.

25.3.1 Intertidal and subtidal littoral environments

The intertidal littoral zone generally constitutes a ‘mosaic’ of various biological communities, substrata
of varied geology and grain-size, moisture gradients associated with inundation and exudation, all
disturbed by human activity (marine cultures, salt pans, various management activities, etc.). Varied
pixels and mixels represent these features.
The subtidal environments correspond to littoral zones permanently inundated by brine or sea­
water. The spectral response of the sea depends on the content of sediments or pigments (water
colour), nature of sea floor, action of surface wind, storms, currents, pollution, etc.

25.3.2 Mineral targets of intertidal zone (optical domain)

‘Pure’ mineral targets such as sands, muddy sands, mud and rocks exhibit monotonous spectral
characteristics between blue-violet and near-infrared (see Chaps. 4 and 23). Only level variations are
detectable, which result from a number of factors such as natural colour of the substratum, water
content, organic matter, grain size, etc. Contrarliy, the spectral response of sediments may be
considerably modified by surface deposition of biological matter such as diatoms, indicating the presence
of microphytobenthos. Gulllaumont et al. (1988) have shown that modification in spectral response in
the presence of a microflora is directly related to the content of pigments (chlorophyll and
phaeopigments). Such a result Is significant since microphytobenthos, with phytoplankton and
macrophyte algae, contributes to the primary production of coastal ecosystems, in particular in areas
where muddy-sandy stretches exposed to low tides are vast. Their temporal variation is very high and
the conventional methods of observation and measurement are difficult to employ.

25.3.3 Vegetal targets of intertidal zone (optical domain)

Spectral characteristics of intertidal vegetation are similar to those of terrestrial vegetation discussed
in Chapter 4. Three classes of Intertidal vegetation are distinguished: macrophyte algae, marine
phanerogams and swamp halophytes.
 Applications 419

Macrophyte algae
Compared to other higher forms of vegetation, algae show a large pigment diversity adapted to the
variability of their luminous environment (Levavasseur, 1986). Studies of various algae groups (green,
red, brown) show a low response between 400 and 500 nm (absorption by chlorophyll) and absolute
values of reflectance, beyond 700 nm, dependent on the conditions of light, cover and degree of
immersion. In the region 500--700 nm, reflectance indicates the diversity of pigments present (Fig.
25.2).

Fig. 25.2: Example of spectral reflectance curves of various algae groups (after Guillaumont, 1991).

Although the range of variation is small, Viollier et al. (1985) have shown that this diversity of
pigments can be detected in wide bands of a HRV type sensor, which give the index XS1/XS2 (see
Chap. 4). This index can be used to differentiate major vegetal populations, provided it is verified that
the pixels analysed fulfil the condition of ‘pure’ target.
Using airborne remote-sensing data of high spatial (pixel < 5 m) and spectral resolutions and
supported by detailed ground spectroradiometric surveys, Bajjouk et al. (1996) have shown that the
13 spectral bands of the CASI airborne imaging spectroradiometer, programmable for position and
width of spectra (see Chap. 2), can be optimally selected for discriminating major macroalgae of the
North Bretagne tidal flats. These positions representative of the principal species are shown in Fig.
25.3, with superposition of ground spectral curves.
On the other hand, specific growth of vegetal fronds and especially the horizontal disposition of
thallus and large density of leaf canopy of algae hinder penetration of sunlight into depth. Thus, for
brown algae, Ben Moussa (1987) has shown that normalised vegetation index (NDVI) reaches its
maximum from a thallus density of 4 (density attained quasl-systematically in situ).
A review by Guillaumont et al. (1997) gives more Information on principles and methods of remote
sensing of macrophyte algae.

H Marine phanerogams
These flower plants colonise sandy and muddy-sand zones, sometimes constituting large areas in
tropical regions (turtle grass). They show In emergence a planophylic posture like macrophyte algae.
In the case of seaweeds, a significant correlation is observed between biomass and normalised
420 Processing of Remote Sensing Data

X (mn)

Ulva -- L. digitata H.elongata

Z. marina - A. nodosum
Q Spectral band
R palmata . -. R serratus

Fig. 25.3: Superposition of the CASI bands and reflectance curves obtained from field measurements
(Bajjouk et al., 1996).

vegetation index. On the other hand, spectral reflectance of a marine phanerogam plant shows variations
depending on the cover, which per se varies seasonally depending on the development of the plant.
Further, Guillaumont (1991) has shown that correlation can be established between a normalised
vegetation index and an estimate of biomass. Diachronic analysis by remote sensing is one of the
methods of monitoring variations in these plants.

H Temperate maritime swamps

The halophyte plants of maritime swamps of the temperate (and cold) zones are erectophyles, generally
small in size (unlike the plants described above). Their spectral characteristics indicate phenological
variations (perennial species, annual species, etc.) and the spectral response fluctuations may be
rapid and large depending on natural (earing, efflorescence) or artificial (mowing, grazing) factors
(Fig. 25.4).
In-situ plant populations can hence be characterised only by diachronic monitoring of spectral
responses.
Significant results have been obtained In the United States (Budd et al., 1982) In favourable
cases such as large stretches of practically monospecific vegetation of Spartina alterniflora and little
modification of the landscape by man. They led to mapping and quantification of biomass of plant
species of swamps, particularly through formulation of significant laws of correlation between biomass
and vegetation Index.
 Applications 421

Reflectance (%) Vegetation index

0.8

0.5

1987

Fig. 25.4: Seasonal variation of reflectance of Schpus m aritim us — Cimel radiometer (after Guillaumont, 1991).

For the European maritime swamps, Caillau et al. (1987) have shown that one should
simultaneously take into consideration the natural seasonal fluctuations of each component of
ecosystem and the calendar of activities corresponding to local characteristics (sometimes ancient
ones) of land use (dams, polderland, salt works, oyster culture, fish farms). Variation of these swamps
depends directly on the management of water resulting from these activities in full development.

■ Tropical maritime swamps

Tropical maritime swamps generally correspond to mangrove formations, fragile forest ecosystems
since their genetic diversity is very much reduced (about sixty ligneous species constitute their flora)
and their ecological tolerance is very low (salinity gradient and daily duration of immersion in particular).
These are one of the most productive populations; dry aerial biomass by weight Is of the order of 300
t ha“ "* in wet regions and primary productivity in the dense mangroves of Malaysia reaches up to 12 to
1 5 1 ha""' y“ ”*.
Several important studies have been conducted, especially in the optical domain, to prepare
Inventories and cartographic zonings and to determine the characteristics of plant communities
(Aschbecher et al., 1995). Some of the investigations In France are those of Blasco et al. (1983),
Populus et al. (1986), Mougenot (1990) and Cuq et al. (1993).
In areas where a dry season exists, the mangrove zones are associated with dry brine marshes
situated behind them and are called ‘salitrals’. The succession of landscape units of a tropical maritime
swamp in New Caledonia is shown in Fig. 25.5. Zones 1 and 2 correspond to mangrove forest proper;
zone 3 Is a transition region between forest and bare soil colonised by one (or more) species whose
spectral characteristics are similar to those of temperate swamp plants (see the preceding section).
Zones 4, 5 and 6 correspond to bare muddy-sand sediment their radiometric response is similar to
that of mineralised zones but can be affected by algal cover of vegetal origin.
Combined use of the normalised vegetation index XS3 - XS2/XS3 + XS2 and the brightness
index (XS1 + XS2)^^^ (see Chap. 4), based on spectral measurements, enables classification of the
principal units of this type of landscape: shrubby mangrove, transition zone and bare soils proper
(Loubersac, 1987). An application in pre-selection of aquaculture sites is given below.

25.3.4 Subtidal zone: hydrocarbon pollution

Spectral response of a water layer in the optical, thermal infrared and microwave bands is discussed
in Chaps. 1, 4 and 26. In this section, we will consider only surface pollution and, in particular.
422 Processing of Remote Sensing Data

Fig. 25.5: Swamps of Mara, New Caledonia: landscape units in the intertidal zone.

hydrocarbon spilling and modification of remote-sensed signal at the sea surface due to the presence
of such products.
Hydrocarbons, generally floating as a thin layer on the ocean surface can be detected by passive
as well as active remote-sensing techniques covering almost the entire electromagnetic spectrum.
From the beginning of research activities in remote sensing, numerous studies have been devoted to
this subject; Stewart et al., 1970 (visible and near-infrared); Munday et al., 1971 (multispectral including
thermal); Hollinger, 1974 (passive microwave); Guinard, 1971 (radar); Fantasia et al., 1974 (Laser).
Integrated analysis of these works has resulted in dividing the electromagnetic spectrum into several
parts in which hydrocarbons show sufficiently stable spectral characteristics relative to the sea surface,
enabling their discrimination.
In passive remote sensing, the division of wavelength {X) bands is as follows:
— X between 300 and 400 nm; Reflectance of hydrocarbons is greater than that of the sea, but
can be used only in day with clear sky.
— X between 420 and 550 nm; Solar reflection on hydrocarbons Is masked by scatter due to
water.
— X between 650 and 900 nm: Hydrocarbon reflection is greater than that of sea.
— X between 3 and 5.5 pm and 8 to 14 pm: In this last portion of the spectrum (see Chap. 1), the
emitted radiation is proportional to the temperature and surface emissivity of the body under observation.
As the temperatures are assumed to be equal and emissivity of oil Is lower than that of sea water,
hydrocarbons appear ‘colder’: detection possible in day as well as night.
— X between 3 mm and 3 cm: Hydrocarbons show higher radiant temperature than that of sea
and variable with oil thickness: detection of petroleum pollution possible in day as well as night and
estimation of the volume of oil through estimation of oil-layer thickness.
In active remote sensing, division of bands is as follows:
— X between 250 and 600 nm: Hydrocarbons under light excitation (laser) can re-emit a signal at
a wavelength, offset relative to that of excitation (fluorescence). As the backscattered wave and excitation
wavelength depend on the type, density and age of hydrocarbons, it Is theoretically possible to identify
the type of pollution detected.
— X between 3 and 30 cm: Since hydrocarbons have the characteristic of attenuating capillary
waves of quite large amplitudes produced by wind (> 4 knots), a (lateral) radar signal will be more
backscattered by unpolluted sea water than by oil layer (Wismann, 1993; Bjerde et al., 1993).

25.4 EXAMPLES OF APPLICATION TO LITTORAL

MANAGEMENT
Some examples illustrate the role of remote sensing as an aid in administration and management of
coastal zones. Six cases are presented below: base mapping of coral environments, pre-selection of
 Applications 423

aquacuitural sites in tropical zones, thematic mapping of seaweeds, detection of hydrocarbon pollution,
characterisation of surface states of sea and monitoring temperature variations of the sea surface.

25.4.1 Mapping of coral environments

Acquisition of precise and up-to-date geographic information on reef zones, lagoons or, more generally,
shallow-water zones characterising intertropical regions is most often a difficult, time consuming and
expensive task, especially In the case of low islands or atolls. In fact, two situations are normally
encountered:
— For the terrestrial part, low altitudes, soft formations (friability of soil and intensive erosion),
difficulty in acquisition of aerial data over far-off islands and small number of reference points for
photogrammetry complicate setting up of a complete geodetic network.
— For the maritime part, shallow depths necessitate a dense grid of sounding profiles. Sea-floor
variations and isolated reef knolls are often randomly distributed. Possibilities of precise location are
limited and optical or radio-electric ranges small.
In view of the above, an intensive base mapping needs to be carried out In tropical coastal regions,
both for environmental management and navigation. To facilitate such investigations, the Naval
Hydrographic and Oceanographic Service of France mainly employs satellite remote-sensing data.
in French Polynesia for example, management of maritime regions, consisting of lagoons, atolls
and high islands, poses problems due to development of pearl-culture activities, conflicting among
themselves and with other activities (such as fishery, tourism). Hence, spatial remote-sensing data
have been used for preparing ‘spatial maps’ of these islands in regions for which conventional data
were not available or were fragmentary.
Based on the criteria of availability. Identical standards (ground resolution, radiometry, format,
etc.), diachronic acquisition and compatibility between graphic and semantic precision for the proposed
applications, SPOT satellite data have been used. The method developed comprises:
— rectification of the image acquired under a basic preprocessing level (1A or 1B) to level 2A
(UTM);
— association of a precise geographic grid (reference points obtained in situ);
— segmentation of the image into 3 major zones: marine region, bare soils, vegetation;
— estimation of bathymetric levels (method explained in Chap. 4) and calibration of the model
with the aid of reference points of known bathymetric levels;
— completion of zones drawn from the preceding stages of processing by addition of external
features (topographic data, roads, etc.) to obtain a spatial map (see CD 25.1: prototype spatial map of
Manihi island, original 1:50,000, Loubersac et al., 1994).
Besides direct cartographic importance, the digital form of remote-sensing data and their geocoding
allows restricting the number of cartographic reference points that enter In the geographic information
system for management.

25.4.2 A quaculture m anagem ent (raising tropical shrim ps)

Shrimp culture is highly concentrated in littoral regions of developing countries, in zones that are not
always readily accessible and have incomplete or obsolete maps. This activity Is mostly characterised
by construction of large ponds on land (a few thousand square metres to more than 20 ha of single
area). Shrimp raising Is today gradually emerging as a destroyer of natural environment to the extent
of self-endangering by destruction of biotopes that are necessary for itself.
In fact, the most favourable zones for construction of shrimp-culture basins are ‘salitrals’, large
areas of flat terrain without vegetation, in the proximity of mangrove environments (Fig. 25.5). Initially,
farms are set up in ‘salitrals’ within mangroves where juveniles are raised to nourish them and where
the water is enriched and subsequently sent into the basins. These ‘salitrals’ are readily manageable
424 Processing of Remote Sensing Data

and impact of management on environment is very often minimal. Such sites are favourable for intensive
shrimp culture.
When the bottom pressure becomes more, these farms spill over the mangrove, gradually
destroying it, with risk of destroying precious biological equilibria, including those that completely or
partly furnish their needs of juveniles. Such an overflowing of farms is due to high bottom pressure, as
well as due to incorrect location and characterisation of favourable zones, i.e., ‘salitrals’, in the available
maps. The methods described earlier for tropical maritime swamps have been successfully applied to
high-resolution spatial images for various objectives: to locate favourable zones, quantify their areas,
identify their forms, furnish qualitative information on soil types and drainage patterns, determine land
cover in river basins on upstream side of sites and accessibility of zones through land or sea routes
and pre-select sites for pumping sea water necessary for shrimp culture.
These features, combined with external data, directly not available from remote sensing, such as
physico-chemical quality of water, productivity of zones, logistic constraints, economic constraints,
etc. enable preparation of pre-selection maps of sites in medium scale (mainly 1:50,000). Such maps
have been used for the first time in the inventory and management plans of aquacultural sites in New
Caledonia (Loubersac, 1987; Populus et al., 1990; CD 25.2).

25.4.3 Thematic mapping of seaweeds

Increasing industrial demand for algal material, and especially the fucal, has led to estimation of
exploitable stocks of seaweed, particularly in intertidal zones. Maps of species classification, estimates
of coverage and biomass quantity in situ and empirical models relating the vegetal cover measured
and biomass available have been developed from the results of investigations (see above; CD 25.3;
after Bajjouk, 1996).
Unlike brown algae whose extent and spectral characteristics are generally stable over time,
green algae exhibit very large variations in development and distribution, often associated with pollution
by excess nitrates. Monitoring and control of their variation are difficult by conventional methods. Use
of species-discrimination methods in remote-sensing data (Populus et al., 1994) offers perspectives
of application in monitoring eutrophication of coastal environment, especially as support to modelling
these phenomena.

25.4.4 Detection of hydrocarbon pollution in sea

Techniques and methods for remote detection of hydrocarbons dumped fraudulently or accidentally in
sea are mainly developed since the 1970s in the United States, Canada, Sweden and France. These
methods make use of various spectral bands (see above).
Thermal infrared bands (Fig. 25.6) have been and are mainly used in France by ‘Douanes’ and
‘Marine Marchande’ to detect fraudulent dumps of hydrocarbons (degassed) and to dissuade polluters.
A satellite equipped with radar enables detection of hydrocarbon slick any time. ERS-1 provided
shocking Images of massive accidental pollution such as the one resulting from the accident of the
‘Aegean Sea’ in the Spanish Corogne Bay In 1992. An example of radar detection of hydrocarbon
pollution offshore of Portugal is shown in Fig. 25.7. It must be noted that the present-day frequency of
acquisition of such information by satellites is not yet compatible with the operational constraints
encountered in the struggle against pollution.

25.4.5 Monitoring surface state of sea by radar imaging

A good knowledge of regime of swells due to wind (amplitude and direction) is necessary for their
forecasting (in aid of navigation and tracking of ships) and for studies of sediment transport from the
 Applications 425

Fig. 25.6: Example airborne detection (thermal IR) of a hydrocarbon dump (dark shaded) in sea. The ship
throwing the dump into the water and its wake are clear (‘hot’ zones) (permission IFREMER).

Fig. 25.7: ERS-1 radar Image of a hydrocarbon pollution site offshore of Portugal (black layer due to low backscatter
to sensor). The white point in the southern part of the layer is probably a ship (after Kerbaol and Chapron, pers.
comm., permission IFREMER).

coast. Analysis of spectra of swell derived from processing of SAR images permits characterisation of
modifications in wavelength and direction of swell when it approaches the coast.
An ERS-1 SAR image of surface sea state under the effect of a storm is shown in Fig. 25.8.
Superposition of the SAR image and the coastline is given on the right. In the bottom.left, a segment
of SAR image depicts swell with nearly west-east orientation. In the top left, a SAR image segment
indicates the Impact of swell on the shore of the Audierne bay and especially the cells of surf at the
coast. In white hue. Dynamic monitoring of such phenomenon is important since It directly determines
the coastal sediment transport and rate of coastal erosion.
426 Processing of Remote Sensing Data

Fig. 25.8: Detection of surf cells of coastal waves under the effect of a storm in South Bretagne. ERS-1 SAR
image of 13/9/1993 (after Kerbaol and Chapron; also see Forget et al., 1996).

25.4.6 Detection and monitoring of sea-surface temperature

variations in littoral environment
Oceanographic applications of detection of sea-surface temperature by remote sensing are numerous:
pelagic fishery (Klimley and Bulter, 1988; Petit et al., 1994), monitoring dynamics resulting from mixing
and geostatistical processing of data (Gohin and Langlois, 1993), detection of frontal structures (Le
Vourch et al., 1992).
An example of synthesised analysis of sea-surface temperature variations associated with
ecological disturbances, eutrophication and seaweed proliferation in the water mass of the Bretagne
coast is presented here. In this region, abnormal growth of green algae is observed generally from
May to July and the proliferation is directly related to the excess of nutrients of agricultural origin,
combined with an increase in solar radiation. Studies have shown the topography and morphology of
the littoral zones (growth on shallow floors), hydrodynamic activity (growth In protected zones), nature
 Applications 427

of sediments (proliferation on sandy zones) and temperature (proliferation in hot zones in spring) as
factors that explain development of seaweeds (Piriou et al., 1993).
The AVHRR (NOAA) data enables spatiotemporal characterisation of warming of surface water
by an interannual synthesis of Its stages. For each year and for each image available, deviations of
temperatures relative to winter reference temperatures have been computed for a 12 x 12 km grid and
the Interannual mean of deviations obtained. (CD 25.4 shows results of synthesis for three periods:
early May, end May and early July, heating of coastal zones of Bretagne in three periods and the mean
of observed thermal deviations relative to a winter reference; after Piriou et al., 1993; IFREMER/CEE/
Bretagne region).
This type of Information provides a better understanding of dynamics of heating of bays and
protected zones In areas where seaweed proliferation has started and helps In explaining variations in
commencement of the above phenomenon depending on the site (heating later in Lannion and Morlaix
bays in the west than in Saint-Brieuc bay in the east). Thus it forms an aid to predictive modelling
through better understanding of the phenomenon.

25.5 CONCLUSIONS
Since the beginning of the 1970s, especially with the launching of the first Earth observation satellite
(ERTS1 precursor to LANDSAT MSS missions), several space missions, viz., SEASAT in 1978, NIMBUS
7 of the same period, LANDSAT TM series since 1983, SPOT series since 1986, ERS series since
1992 and SEAWiFS in 1997, have furnished large amount of Information on oceans and coastal seas.
Scientific and technical advances emerging from such data have been briefly indicated and
Illustrated above. During the same period Important developments have also taken place in airborne
remote sensing, as evidenced by organisation of international seminars on the subject since 1994.
It should be noted that specific characteristics of littoral environments impose constraints on
spatial and temporal observations and measurements which remote-sensing techniques cannot
necessarily satisfy.
Future perspectives of development in coastal applications of aérospatial remote sensing revolve
around three axes:
— very high spatial resolution (about 1 m) with new-generation satellites,
— ^very high spectral resolution (programmable) as surface topography (CIDAR) which new airborne
missions may provide,
— development of sensors for water colour measurements (see Chap. 4),
— use of minisatellites ensuring a high repetition of acquisition through offset cycles of orbits.
On the other hand, Independent of the technological and methodological developments specific
to remote sensing, a definite future exists in mixing and combining the georeferenced data with
conventional geographic data and digital modelling. That is why, at the levels of conceptualisation and
development of techniques, methodologies and applications, investigations are oriented towards close
linkage and synergy between data, methods and tools of remote sensing, hydrodynamic digital
modelling, hydrosedimentary or biological studies of ecosystems and geographic information systems.
Such synergy Is essential for developing operational regional information systems for coastal zones,
to be used in decision making, in optimised choice and communication of environmental Information.

References
Aschbacher J, Tiangco P, Giri CP, Ofren RS, Paudyal DR, Ang YK. 1995. Comparison of different sensors and
analysis techniques for tropical mangrove forest mapping, IGARSS’95 Congress, 3: 2109-2111.
BajjoukT, Guillaumont B, Populus J, 1996. Application of airborne imaging spectrometry system data to intertidal
seaweed classification and mapping, Hydrobiologia, 326/327, 463/471.
428 Processing of Remote Sensing Data

Ben Moussa H. 1987. Contribution de la télédétection satellitale à la cartographie des végétaux marins: Archipel
de Molène, Thèse de doctorat d’Université Aix Marseille II, 122 p.
Bjerde KW, Solberg S, Solberg R. 1993. Oil spill detection in SAR imagery, IGARRS’93 Congress, p. 943-945.
Blasco F, Lavenue F, Chaudury MU, KerrY. 1983. Simulations SPOT au Bangladesh. Étude des mangroves des
Sunderbans, Rapp CNES/GDTA, 31 pp.
Budd JTC, Milton EJ, 1982. Remote sensing of salt marsh vegetation in the first four proposed Thematic Mapper
bands, Int. Jnl. Rem. Sens, 3 (2): 147-161.
Caillaud L, Guillaumont B, Manaud F. 1987. Essai de discrimination des modes d’utilisation des marais maritimes
par analyse multitemporelles d’images SPOT, Rapport ATP Télédétection Spatiale. IG/SRETIE/MERE/7161.
Cuq F, Courmelon F, Madec V. 1993. Planification côtière de Guinée Bissau, 3 vol. Cartes FIT, UICN/DGFC-MDRA,
Édité en français, anglais et portugais.
Fantasia JF, Ingrao HC. 1974. Development of an experimental airborne laser remote sensing system for detection
and classification of oil slicks. Proc 9^^ Int. Symp. Rem. Sens, of Env., Ann Arbor Ml, pp. 1711-1745.
Forget P, Rousseau S, Cauneau F, Chapron B, Kerbaol V, Cuq F, Bonnafoux G, Blerard C, Garello R, Grassin S,
Bonicel D, Hajji H. 1996. Expérimentation radar GLOBESAR en baie d’Audierne, Rapport technique LSEET,
Univ. Toulon. Ref. 94/CNES/0380.
Gierloff-Emden FIG. 1982.‘Remote sensing for coastal areas’, Symp IGARRS’82, München WA-8, pp. 11-18.
Gohin F, Langlois G, 1993. Using geostatistics to merge in situ measurements and remotely-sensed observations
of sea surface temperature, Int. Jour, of Rem. Sens., 14(1): 9-19.
Guillaumont B. 1991. ‘Utilisation de l’imagerie satellitaire pour les comparaisons spatiales et temporelles en zone
intertidale’. Estuaries and Coasts: spatial and temporal intercomparisons, ECSA 19 Symposium, Elliot and
Ducrotoy Eds, Olsen and Olsen, pp. 63-68.
Guillaumont B, Bajjouk T, Talée P. 1997. ‘Seaweeds and remote sensing: a critical review of sensors and data
processing’. In: Progress in Phycological Research, Vol. 12, Chapman and Round Eds. Biopress, pp. 213-
282.
Guillaumont B, Gentien P, Viollier M. 1988. ‘Mesures radiométriques haute résolution du microphytobenthos intertidal’,
Proc. 4th Int Coll. On spectral signatures of objects in rem. Sens., Aussois ESA SP 287, pp. 333-336.
Guillaumont B, Lavavasseur J. 1988. Variations saisonnièrence de la réflectance en fonction de la phénologie des
plantes des marais, 3e Conférence Internationale sur les Zones Humides, Rennes, sept. 1988.
Guinard NW. 1971. ‘The remote sensing of oil slicks’, Proc. 7th Int Symp. Rem. Sens, of Env. Ann Arbor Ml, pp.
1005-1026.
Hollinger JP, 1974. The determination of oil slicks thickness by means of multifrequency passive microwave technique,
Nav. Research Lab. CG-D-31-75. Washington DC.
Klemas V, Gross MF, Hardisky MA. 1987. Evaluation of SPOT data for Remote sensing of physical and biological
properties of estuaries and coastal zones, Symp. Int. Spot 1: utilisation des images, bilans, résultats, Paris Éd,
Cepadues, pp. 1035-1040.
Klemas V, Philpott WD, 1980. The use of satellites in environmental monitoring of coastal waters. Final rpt. Univ
Delaware NASA NSP-1433.
Klimley AP, Butler SB, 1988. Immigration and emigration of a pelagic fish assemblage to seamounts in the Gulf of
California related to water mass movements using satellite imagery. Mar Ecol. Prog. Ser, 49:11-20.
Le Goulc M. 1988. ‘Utilisation de SPOT en hydrographie, Symp. Int. Spot 1; utilisation des images, bilans, résultats,
Paris Éd. Cepadues, pp. 1063-1068.
Le Vourch J, Millot C, Castagne N, Le Borgne P, OIry J-P. 1992. Atlas des fronts thermiques en Mer Méditerranée
d’après l’image satellitaire. Mémoire de l’Institut océanographique de Monaco, n° 16, VI, 152 pp.
Levavasseur G. 1986. Plasticité de l’appareil pigmentaire des algues marines macrophytes. Régulation en fonction
de l’environnement. Thèse de Doctrat d’État, Paris VI, 210 pp.
Loubersac L. 1983. Coastal Zone Inventory by high resolution satellites, AIpbach Summer School, 27 July-5 August
• 1993, ESA SP 205, pp. 87-94.
Loubersac L. 1987.‘SPOT, un outil d’aide à la présélection de sites favorables à l’aquaculture Bilan et perspectives
du projet PEPS ‘ALIAS Calédonie’, Symp. Int. Spot 1; utilisation des images, bilans, résultats, Paris, Éd.
Cepadues, pp. 1041-1049.
Loubersac L, Andrefouet S, Chenon F, Morel Y, Varet H. 1994. Information géographique dérivée des données de
la télédétection spatiale de haute résolution sur les lagons des îles hautes et des atolls. Application aux
environnements des îles de la Polynésie française: état et perpectives. Mémoires de l’Institut océanographique
de Monaco, 18:75-84.
 Applications 429

Mougenot B. 1990. Caractéristiques spectrales de surfaces salées à chlorures et sulfates au Sénégal, 2®journées
de télédétection: Caractérisation et suivi des milieux terrestres en régions arides et tropicales, Colloques et
séminaires, ORSTOM, Bondy, pp. 49-70.
Munday JC, MacIntyre WG, Penney ME. 1971. ‘Oil slicks studies using photographic and multispectral scanner
data’, Proc. 7th Int Symp. Rem. Sens, of Env., Ann Arbor Ml, pp. 1027-1043.
Petit M, Dagorn L, Lena P, Slepoukha M, Ramos AG, Stretta JM. 1994.‘Oceanic landscape concept and operational
fisheries oceanography’. Mémoires de l’Institut Océanographique de Monaco, 18: Les nouvelles frontières de
la télédétection océanique, pp. 85-97.
Piriou J-Y. 1993. Cartographie, des zones sensibles à l’eutrophisation; cas des côtes bretonnes. Rapport IFREMER/
CEE/Région Bretagne, Contrat 6510-90, 2 vol.
Populus J, Guillaumont B, Ruiz O, Tallec R 1994. ‘Biomass assessment of green algae proliferations with high
resolution airborne images’, Proc. of the first International ETRIM Airborne Remote Sensing Conference and
Exhibition, Strasbourg, France, 11-15 Sept. 1994, 3:153-164.
Populus J, Hertz R. 1986. Cartographie des magroves de la côte sud-est du Brésil avec Landsat TM, Photo
Interprétation, 85-2 fasc. 4 pp. 31-38.
Steward S, Spellicy R, Polcyn F. 1970. Analysis of multispectral data of the Santa Barbara oil slick, Publ. 3340-4-F,
Willow Run Lab. Univ. Michigan, 57 pp.
Viollier M, BelsherT, Loubersac L. 1985.‘Signatures spectrales des objets du littoral, Proc 3rd Int. Coll, on spectral
Signatures of Objects in Rem. Sens., Les Arcs, France, ESA SP 247, pp. 253-257.
Wismann V. 1993. Radar signatures of mineral oil spills measured by an airborne multi-frequency radar and the
ERS1 SAR, IGARRS’93 Congress, 3: 940, 942.
Zacharias M, Niemann O, Borstad G. 1992. An assessment and classification of a multispectral handset for the
remote sensing of intertidal seaweeds, Canadian J. of Rem. Sens. 18 (4): 263-274.
 26
Applications of Thermal-infrared
and Microwave Data

26.1 THERMAL-INFRARED DATA

It was shown in Chapter 1— Physical basis— that instantaneous values of surface temperature can be
related to diurnal évapotranspiration. This relationship has been used for numerous applications,
described below, using satellite data: estimation of water consumption by maize, detection of dry and
irrigated zones and mapping évapotranspiration in southwestern France.

26.1.1 Detection of drought in France

Detection of drought necessitates continuous monitoring of surface temperatures at a regular time
interval. Due to problems of cloud coverage impeding acquisition of thermal-infrared images,
combinations of data of five days of NOAA-AVHRR images have been used for this study. This analysis
was conducted by CNES at Toulouse by selecting the maximum of vegetation index of AVHRR images
acquired during a period of 5 days from March to October (the most Important period for crops). These
data comprise 48 images per annum. This method eliminates cloudy or defective images. These
images were geometrically rectified (see Chap. 13), superposed and corrected for atmospheric effects
by the empirical split-window method (see Chap. 1) and for temporal drift between the satellites.
Twenty meteorological stations distributed throughout the territory were chosen for representing major
regional climatic variations. Each of these sites was characterised by a 9 km x 9 km sector (or 9 x 9
NOAA pixels) for which surface temperatures were extracted from NOAA images of the various years
under study and for which temporal climatic variations measured at the station were also available.
Analysis of cumulative deviations (TSf^QAA "" ^^meteo) March to October between various
stations revealed (Fig. 26.1):
— latitude effect: marked difference between north and south sites (MIrecourt, Rennes, Perpignan);
— land-cover effect: areas where bare soils are predominant show higher surface temperatures
than those of zones which are covered by dense vegetation (as in MIrecourt) or are marshy (Saint-
Laurent-de-la-Prée);
— effect of proximity to coasts (case of Quimper): generally higher wind, humidity or rain tend to
reduce deviations (T s - Ta).
For a given station, variations in the observed temperature deviations between various years are
indicative of hydrological deficiencies and enable detection of stress periods.
Marked differences are thus observed between the years (1989-1990) of highest drought and
the year (1988) of maximum rain in the Toulouse region. Flowever, these differences are less significant
for the Avignon site where a large number of parcels are regularly irrigated and less affected by
droughts (Fig. 26.2).
 Applications 431
October

Perpignan
Avignon
Saint Laurent de la Prée
Rennes
Quimper
Mi recourt

Fig. 26.1 : Variation of cumulative deviations { T s - T a ) from March to October for six stations
(after Courault et al., 1994, in Agronom ie).

^ ( T s - T,) rc ) 2 (7 -,- T,) (°C)

Fig. 26.2: Comparison of cumulative deviations (7 s - Ta) from March to October for Toulouse and Avignon, 1988
to 1990 (after Courault et al., 1994).

For estimating the évapotranspiration (ETR) of each site using the agrometeorological model
‘Magref, one can calibrate the simplified model (Chap. 1, eqn 16) to the scale of France for each year
of study and thus map évapotranspiration by means of NOAA Images (Courault et al., 1994).
Meteorological data necessary for the model Tairand Rn are interpolated by krigging.
Maps resulting from the sum of surface temperatures are used to prepare ETR maps. Maps of 7s
deviations between years are prepared and used for preliminary prognosis of zones most affected by
drought (CD 26.1). It can thus be seen that Southwest, Charente and Normandy are the regions that
seem to have been most affected by lack of water in 1989 and 1990. The computed values of ETR for
Charente confirm these observations and can be used to quantify the deficit:
ETR 88 = 503 mm, ETR 89 = 356 mm, ETR 90 = 317 mm.

This first example clearly illustrates the significance of thermal infrared data of NOAA-AVHRR In
obtaining an overall estimation of water content of crops over large areas. This approach is useful to
432 Processing of Remote Sensing Data

detect and confirm location of zones most affected by drought during the preceding years and is
complementary to the method of monitoring vegetation Indices. The advantage of surface temperatures
compared to vegetation indices Is that the former represent instantaneous surface hydrological
conditions whereas the latter react to stress undergone by crops only after several days, which often
makes their interpretation more difficult.
One of the major problems lies in validation of these estimates. At this spatial level, only local
validation is possible in sites for which ground measurements are available (possible measurements
over parcels). At the regional level, validation is most difficult; measurements by airborne sensors may
be envisaged but they are cumbersome to employ and often limited in time. It is to be noted that such
experiments of validation were attempted in the projects HAPEX-MOBILHY (Andre et al., 1988) and
HAPEX-Sahel (Goutorbe et al., 1993).

26.1.2 Estimation of exchanges between soil, vegetation and

atmosphere
Various approaches exist for using surface temperature in models of évapotranspiration estimations.
The principle is as follows. In most cases these ‘deterministic’ models simulate, by means of more or
less complex equations, surface temperatures observed by the satellite. Such simulations give certain
parameters that are difficult to measure or estimate but important for computation of fluxes such as
stomatic resistance of vegetation covers or initial reserves of water in soil, etc. These methods consist
of comparing Ts in these transfer models by Inversion techniques. A preliminary simulation Is conducted
with a manipulation of parameters fixed a priori. The surface temperature estimated is then compared
with that observed by satellite. Attempt is made to minimise deviations between these two temperatures
by modifying input parameters of the model for each new simulation.
Some models operate at diurnal scale and necessitate only pinpoint data for calibration of these
parameters. Such Is the case of a model of Penman-Monteith type surface, combined with a simplified
description of atmospheric boundary layer. This model has been used, for example, to map
évapotranspiration in south-western France with the aid of NOAA-AVHRR images, the unknown
parameter being resistance of the cover. The resultant map indicates spatial variations of ETR.
Evapotranspiration values observed for the forest of Landes are of the order of 3 mm on average and
for vineyards and orchards 1.5 mm. Values obtained for crops such as wheat and maize vary between
2 and 2.6 mm per day for the date under consideration (Lagouarde and Brunet, 1991).
Other models, on the other hand, require several dates for continuously estimating various flows
during the cycle of a crop. In the latter case, deviation between observed and simulated temperatures
of various surfaces is minimised for the entire period by considering all dates for which remote-sensing
data are available.

26.1.3 Mapping rainfall distribution in Sahel

Two methods applicable to the Sahel region in Africa are described here. The first Is based on analysis
of cloud temperatures and the other of surface temperatures.
In the Sahel region rain is often associated with thick cloud formations with a cold top. Analysis of
METEOSAT images (6 images per day) gives a map of frequency of occurrence of such clouds with a
threshold fixed at -40°C. Association between the number of these occurrences and rain observed in
some stations is fairly strong, with a high correlation coefficient (Fig. 26.3a). Applying this relationship
to the entire image, a cumulative-rain map is proposed for the agricultural season of Western Africa
(Fig. 26.3b).
 Applications 433

Fig. 26.3: Mapping the rain (measured at a few stations): a) occurrence of clouds with a cold top detected in
Meteosat IRT images; b) application in Western Africa (after Lahuec et al., 1986).

Another method employed in Senegal gave a direct relationship between surface temperatures
measured with the IRT band of METEOSAT and rainfall measured at some ground meteorological
stations over several years (Assad et al., 1986). Regions affected by rain show lower surface
temperatures due to évapotranspiration whereas zones with lowest rainfall retain high Ts. A linear
relationship was established between these two variables (Fig. 26.4a), which enabled mapping
cumulative rainfall during the agricultural period (CD 26.2). A comparison of isohyets drawn manually
from ground rain-gauge network is satisfactory for dry and wet seasons (Fig. 26.4b).

26.1.4 Characterisation of frost zones

Thermal Images acquired at night are important for mapping frost zones. An example Is given for the
Rhone valley wherein NOAA-7 data acquired at 3 h 00 TU on 15 and 23 March 1982 were used to
detect zones sensitive to frosting (CD 26.3). Spring frost Is particularly feared since it leads to large
damage to crops. The first image acquired in low-wind condition on 15 March shows scattered zones
of variable degrees of cooling. The second Image of 23 March corresponding to high-wind (mistral)
conditions shows much more homogeneous zones, indicating the influence of wind on night cooling
(hazard of frost Is less for strong winds). The average vertical thermal gradient is in fact as much steep
as the wind is low and the sky Is clear, a condition corresponding to the so-called radiation frost (or
white frost). When wind blows, air in the lower layers of atmosphere Is stirred, inversion of thermal
profile is less pronounced and the net radiation is less negative. A low wind hence suffices to limit the
cooling to the vicinity of soil. This principle is sometimes used to fight against frost by mechanical
stirring of the lower layers of air.

26.1.5 Analysis of topoclimates from surface temperatures

In montane regions climatic observations are rare often due to sparse network of meteorological
stations. Thermal-Infrared remote sensing can furnish information on topoclimate of these regions
434 Processing of Remote Sensing Data

Z Rainfall (mm)

METEOSAT METEOROLOGICAL NETWORK

Fig. 26.4: a) Relationship between cumulative surface temperature (METEOSAT) and rainfall of some stations in
Senegal for the years 1984 and 1985; b) comparison of isohyets with ground rain-gauge data (after Assad et al.,
1986).

with high spatial resolution, in the north-west Massif Central, forest cover is dense on slopes and the
high évapotranspiration gives rise to a surface temperature close to that of air. Relationships between
surface temperature and altitude show gradients of about 0.7°C/100 m, of the same order of magnitude
as those of air (Fig. 26.5). When density of vegetation cover is less, effects due to slope and orientation
are more marked and the surface temperature then strongly depends on the radiation received. Thermal
inversion phenomena may be observed: the cold air accumulated by gravity at the bottom of valleys
 Applications 435

r,ro

Fig. 26.5: Relationship between surface temperature (HCMM thermal images) of 17 July 1978 over Mount Forez
and altitude derived from digital elevation models (after Lagouarde et al., 1983).

during night has no time to be warmed up in the morning under low wind conditions and temperatures
increasing with altitude are observed. The thermal images acquired over Mount Ventoux illustrate
these effects of topography on surface temperature (CD 26.4). Temperature increases with altitude up
to about 100 m and then decreases for higher altitudes.

26.1.6 Use of surface temperature and vegetation index

Numerous authors have combined data acquired in various wavelength domains to obtain more
information on surface characteristics. Such studies include relationships between normalised
vegetation index (NDVI = NIR ~ R / R + NIR; Chap. 4) and surface temperature since they facilitate
evaluation of variations in the structure of plant cover (by NDVI) and monitoring hydrological state of
surfaces (with Ts). Most often a negative correlation Is observed between these two variables. When
a surface depicts a high vegetation Index, it generally indicates a surface in which vegetation is dense
and active and évapotranspiration is high, tending to reduce surface temperature. This relationship Is
used for various applications such as detection of zones of forest fires or estimation of surface moisture,
etc.
A theoretical representation of the relationship between degree of coverage (which can be
expressed as a function of NDVI) and temperature difference (Ts - Ta) is shown In Fig. 26.6. In a given
region, points are generally distributed within a trapezium whose four corners correspond to the following
extreme situations: 1) vegetation totally covering the ground and well watered; 2) vegetation totally
covering the ground and dry; 3) bare soil saturated with water; 4) dry bare soil. A stress index, known
as water deficit index (WDi) has been developed from this scheme: WDI = AC I AB which aids in
detecting, for the same quantity of in-situ vegetation, differences in hydrological state along the position
of the point inside the trapezium (see Fig. 26.6).

26.1.7 Conclusion
The few examples described above show the diverse range of applications of thermal-infrared data:
evaporation mapping, detection of frost or stress zones, estimation of rainfall in Sahel, detection of
436 Processing oí Remote Sensing Data

Vegetation well
watered

Fig. 26.6: Theoretical diagram defined by the relationship between (T s - 7a) and degree of ground coverage
(after Moran et al., 1994).

forest fire hazards, etc. Role of remote sensing lies in furnishing objective criteria for comparing zones
with one another or for temporal monitoring to obtain information on spatial variability, which is often
missed in conventional measurements by meteorological networks (as in the case of montane or
desert zones). While meteorological satellites such as NOAA or METEOSAT provide such temporal
monitoring due to their high revisiting capability, their spatial resolution is low and methods of transfer
from regional to local scale are yet to be developed, awaiting launching of future sensors with higher
resolution.

26.2 MICROWAVE DATA: USE OF RADAR IMAGES

A radar unit derived from a military system was used in France for civil purposes only in April 1973.
The data were furnished in analog form (films). This was the period when the first scatterometers were
developed (group of F.T. Ulaby at the University of Kansas, USA).
Applications of microwave remote sensing have been particularly developed in the recent years
since the launching of the European Remote Sensing Satellite (ERS-1) on 16 July 1991, followed in
1992 by JERS (Japanese Remote Sensing Satellite) and in 1995 by ERS-2 and Canadian RADARSAT
of high geometric resolution. Microwave remote sensing has been mainly used in moist and cloudy
zones in intertropical or temperate regions to compensate for the impossibility of acquiring data in the
visible and near infrared bands.
It should be noted that physical phenomena associated with this spectral region are not yet well
understood and much research needs to be carried out based on ground, airborne and satellite
experiments (see Chap. 1). For example, the data acquired by SEASAT in 1978 proved to be a valuable
source of information, its short life span (4 months) notwithstanding. Processing and interpretation of
data acquired by radar, differing from those of the visible and Infrared bands, necessitate knowledge
of characteristics pertaining to this particular technology.
 Applications 437

26.2.1 Characteristics of radars

■ Spectral domain of imaging radars
The spectral domain of imaging radars, as shown in Table 1.4 (Chap. 1), lies between 30 and 2 cm or
1 and 15 GHz. This frequency range is determined by technological constraints:
— For frequencies lower than 1 GHz, technology for fabrication of radar systems with spatial
resolution of less than 20 m, i.e., compatible with current applications in remote sensing is not developed;
— Beyond 15 GHz, costs of fabrication are too high or required materials do no exist.
For example, atmospheric attenuation is very high In the millimetre band (7.5 to 1 mm or 40 to
300 GHz) due to absorption by oxygen and water vapour. Only objects of dimension close to that of
the wavelength can be detected and this feature hinders applications to earth resources. Moreover,
the limit of 40 GHz corresponds to a change in technology that enhances costs of equipment.
Within these constraints, experiments conducted by various groups have resulted in a consensus
on a large range of applications of these spectral bands. Band L (1 to 2 GHz) is important in
oceanography and useful in topographic and geological studies. Band C (4 to 8 GHz) Is optimal for
studying soil moisture and band K (10 to 15 GHz) for vegetation.
That Is why satellites ERS-1 and ERS-2 are equipped with a radar system with synthetic antenna
(Synthetic Aperture Radar, SAR) for band C (5.3 GHz) in the ‘wave’ mode and in the image mode. The
RADARSAT system also uses band C but with the possibility of a wide choice of beam widths and
incidence angles, ensuring swaths of 35 to 500 km and resolutions from 10 to 100 m (Table 26.1). The
European airborne survey EMAC 94/95, preparatory to the European program ENVISAT, used a
polarising SAR with band L (1.25 GHz) and C (5.3 GHz) for investigating snow and Ice in Scandinavia
and surface moisture of soils in Belgium. JERS-1 is equipped with a band L (1.3 GHz) radar.

H Sensitivity limits of radar systems

Systematic analysis of backscattered signal from various types of surfaces has been developed by
ground measurements. The backscattered signals, reduced to coefficient Oq (in dB), are calibrated
absolutely and hence comparable with one another. Distribution of Oq values of various ground objects
is schematically shown In Fig. 26.7.

-50 -30 -10 20 (dB)

H— — h-
Cairn lake Sea surface

Bare soil

Crops
Vegetation
Built-up surfaces Built-up surfaces
(horizontal) (vertical)

Fig. 26.7: Values of Oq for various surfaces.

It is evident that sensitivity limits of radar need to be precisely defined depending on the application
envisaged.

H Side looking radar (Real aperture lateral radar)

This system corresponds to older radar technology (Fig. 26.8) and continues to be used for airborne
missions. It comprises a transmitter-receiver antenna placed outside the platform and oriented laterally.
438 Processing of Remote Sensing Data

Table 26.1: Examples of satellites with microwave sensors

ERS JERS RADARSAT

Date of launching -1:25/07/1991 11/02/1992 04/11/1995

-2:20/04/1995
Altitude 785 km 570 km 800 km
Synthetic Aperture Radar Image mode
Spectral band C L C
Frequency 5.3 GHz 1275 MHz 5.3 GHz
Polarisation V-V H-H H-H
Specific characteristics Possibility to study waves Revisiting capability Various angles of
and wind speeds 44 days view, 24-day cycle,
revisiting at 1.5 day at
45° N possible
Angle of incidence on the 23° 35° 1 0 ° to 60°
ground swath
Spatial resolution 25-30 m 1 8 -18m 1 0 to 1 0 0 m

depending on angle
of view
Swath 1 0 0 krn 75 km 35 to 500 km
Other onboard sensors ERS-1 OPS (Optical Sensor):
" radar altimeter 3 bands in visible and
" ATSR scanner (4 bands NIR, 4 bands in infrared
centred on 1.6, 3.7,10.8
and 1 2 pm)
ERS-2
Same as above plus GOME
(Global Ozone Monitoring
Equipment)

The real aperture of the antenna beam (angle 0) determines the segment of ground, perpendicular to
the direction of platform’s motion, viewed almost instantaneously (a few milliseconds). Forward motion
of the aircraft determines the length of the scene recorded.
Resolution of the system is better in the transverse direction (due to the frequencies used) than in
the longitudinal (limited by the width of antenna).
In fact, two ground objects separated by a distance Ad such that Ad< c x l2 (where c is the
velocity of light, x the duration of Impulses emitted by radar) will be confused as one, since the
backscattered waves overlap each other. This distance is the limit of optical resolution (slant range
resolution) while transverse ground resolution (r^) is equal to ox /2 cos0, where 0 Is the depression
angle of the radar (angle of radar beam with the horizontal). Longitudinal resolution (r¡) of radar is
equal to the width of Its aperture beam to its contact with ground. Since the aperture angle is
approximately equal to the ratio of the wavelength used to the size of antenna, q ^ ( X / L ) x d (where X
Is the wavelength, L antenna length and of the inclined distance between the sensor and the object)
and longitudinal resolution can be improved by increasing the antenna size. However, limitations are
imposed by the bearing capacity of the platform and costs. Acquisition of images Is done In two
modes:
 Applications 439

Fig. 26.8: Principle of operation of side looking radar.

H: Platform altitude; \/: velocity of platform’s motion; 0 : aperture angle of antenna beam; tf. instantaneous
response of elementary surface; R: resolution of elementary surface.

— The analog mode, older, is based on conversion of the backscattered signal into a more or less
intense light spot on a cathode-ray oscilloscope. A sensitive black-and-white film placed In front of the
screen on a support rotates with a velocity corresponding to that of the platform motion and registers
the light sppts.
— ^The digital mode is used since 1978. Digital recording of backscattered signal enables calibration
of data. It is thus possible to compare information acquired on different dates.
Data acquired by real aperture lateral radar have advantage since they necessitate only simple
processing tools and show good spectral quality (resolution of 1 dB). On the other hand, they have the
disadvantage of poor spatial resolution: the transverse resolution depends on the nadir distance (which
determines the incident angle and its complement the depression angle), whereas the longitudinal
resolution varies depending on the wavelength and antenna size.

H Synthetic Aperture Radar

For Improving the spatial resolution and overcoming the constraints of antenna size, Doppler effect is
used. The successive positions of a real antenna, related to the forward motion of the platform, are
considered elementary sources, giving rise to a ‘synthetic’ antenna. When the platform arrives at
(co-ordinates -L/2, H), the ground point S (co-ordinates 0, 0) commences to receive the first pulses
(Fig. 26.9).
When the platform Is at after traversing a distance A^ A^ = L, point S ceases to be illuminated.
The principal lobe of the antenna of longitudinal resolution illuminates the ground at an angle 0 such
that 0 = /c An integration of the signals Is obtained by applying a coherent processing to the
bipolar In-phase and quadrature views recorded for the successive positions A^ to A^ of the platform.
440 Processing of Remote Sensing Data

Fig. 26.9: Principle of SAR (after Paquet, 1997).

Each impulse of the signal consists of a train of coherent waves. The backscattered waves are
received by an antenna over several seconds of flight of the platform, recorded and integrated. An
object occurring in the path of the antenna beam reflects part of the waves towards the latter. Depending
on the movement of the platform, the object-antenna distance will be equal or not to an integer number
of wavelengths. The object will be ‘seen’ only when this distance is equal to an integer number of
wavelengths. The object-antenna distance, initially large, passes through a minimum when the object
is nearest to the antenna (following the motion of the platform) and increases as the platform moves
away. The antenna receives the reflected waves, combines them with a reference train of coherent
waves and makes the two wave trains interfere with each other. When the reflected signal coincides
with the reference signal, the interference leads to high signal amplitude and the point is recorded as
bright. In the inverse situation, the signal amplitude Is small and the point recorded as dull. A ground
point hence corresponds to a sequence of bright and dark segments, of various wavelengths, constituting
a one-dimensional interference figure, known as radar hologram. Each ground object is represented
by several points in a hologram. When the hologram is Illuminated by a coherent light source, each
bright segment constitutes an independent source of coherent light. A single point exists behind the
hologram at which the light waves corresponding to the initial object are combined to form an image of
the object.
Focused radars, which analyse a large number of reflections of the same object, are differentiated
from non-focused ones that record a limited number of reflections.
The limit of longitudinal resolution r^of a focused is given by the relation = U2, where L is the
wavelength of the real antenna. The resolution is no longer dependent on the wavelength or altitude.
For non-focused radar, r/ = V2 -Jxd is independent of antenna size but dependent on the wavelength
and altitude.

26.2.2 Image quality

Images obtained by real-aperture or synthetic-aperture radars differ entirely from photos or images of
other spectral bands. Topographic features have major effect on the distribution and intensity of the
 Applications 441

backscattered signal. Moreover, very large variations in grey levels are observed between surfaces of
high and low backscatter coefficients. Speckle effect is produced due to surface irregularities of
dimensions similar to the Incident wavelength, even if the surface is homogeneous. This forms a noise
which increases with the amplitude of the backscattered signal and which can be attenuated by filtering.
Five specific features of radar images can be Identified:
— Distortions or Inversions of the position of objects may be produced depending on their relative
proximity, their size and system resolution. As can be seen from Fig. 26.10, object D situated in front
of object ABCE on the ground is observed within ABCE \n the image. In montane regions, relief is
sharp and peaks appear bent towards the sensor (Fig. 26.11).
— Radiometric values are affected by variations in incidence caused by changes In topography.
The intensity of response of segment BA is greater than that of segment AB, independent of composition
of these segments (Fig. 26.11).

Record on film (plan view)

Fig. 26.10: Distortions and inversions in radar images.

Receiver A lateral line of information

Fig. 26.11: Variations due to topography.

442 Processing of Remote Sensing Data

— Lengths are not preserved: segment BA' appears much shorter than segment AA' although
their ground lengths are identical.
— ‘Shadow’ effects (segment BC on ground, segment A 'C in image) represent absence of
backscatter during certain time. They appear as dark zones in the record (Fig. 26.11) since even
radiation scattered by atmosphere is not there as in the case of visible band images. In virgin
areas, the presence and disposition of such dark zones have been used to infer criteria of relative
topography.
— ‘Corner’ effect corresponds to a very large backscatter signal for some objects. Such is the
case of some buildings whose face irradiated by radar beam produces a specular double reflection.
The backscatter signal then Is greater than that of adjacent objects, which may be either diffuse— bare
soils or plant cover— or specular— calm water, paved surfaces (Fig. 26.12).

X Urban buildings
Vegetation \ calm w a te r,
D D D D
cover O D ^D D

DIFFUSE CORNER
Variable Oo SPECULAR Very high Oq
Oo= 0
Specular double reflexion

Fig. 26.12: Corner effect in radar images.

26.2.3 Applications of radar data

■ Mapping and topography
Some regions of the Earth cannot be photographed and hence cannot be mapped by conventional
techniques due to intense cloud coverage. This is one of the reasons for small proportion (42%) of
regions mapped in scales equal to or less than 1:50,000. Systematic airborne radar surveys carried
out during the years 1970-1980 in Brazil and Gabon have led to application of the technique in new
areas such as the entire Amazon basin but its cost Is prohibitive and preparation of mosaics compatible
with geometric requirements of mapping is difficult. The large amounts of satellite radar data now
available with systems ERS-1, ERS-2, JERS and RADARSAT (Table 26.1) respond to the cost and
technical constraints, while ensuring rapidity and simultaneous acquisition of large areas, although
topographic applications (production of digital elevation models, DEM) have been developed only
recently. Three main techniques are used for topographic mapping by radar data (Polidoh, 1995 and
1997).
 Applications 443

□ Radargrammetry
Radargrammetry is an adaptation of photogrammetry to radar geometry. It comprises the same two
stages: construction of a stereoscopic model with the use of reference points, if possible, and
stereoscopic analysis consisting of identification of pairs of homologous pixels in the two images.
Radargrammetry differs from photogrammetry in geometric equations and mode of stereoscopic fusion
that are specific to it. One of the first operational programs for analytical restitution was proposed in
1984.

□ Interferometry
The objective of interferometry is analysis of phase differences between radar signals received from a
given zone or from two positions of the same antenna (case of ERS-1 data, before the launching of
ERS-2, acquired with a repetition cycle of 3 or 35 days) or by two different antennas. Such an analysis
was conducted In the tandem ERS-1/ERS-2 mission with a very short time interval (less than 1 h)
since the same instrument was present on both satellites. In favourable conditions of acquisition, this
phase difference gives rise to an interferogram whose interference rings are very sensitive to ground
elevation variations. Since the phase of the signal is known only in the 2ti mode, it must be ‘unfolded’,
I.e., converted to absolute phase, often a difficult task. Applications of Interferometry will be seen in
the sections on geology and hazard monitoring.

□ Radarclinometry
Radarclinometry is based on the fact that any change in slope in the direction perpendicular to platform’s
trajectory Induces (except in specific symmetries) a local variation In backscatter.The reciprocal is not
true, however, since backscatter varies according to properties related to the nature of irradiated
surface. Nonetheless, this relationship can be inversed in the case of very homogeneous media such
as deserts, ice sheets or dense moist tropical forests. This inversion is done by radarclinometry by
converting the intensity (corrected as much as possible for parasitic phenomena that modify it from
the backscatter coefficient) into incidence angle (i.e., angle between incident wave and the local normal
to the irradiated surface). Assuming the local form of terrain, since the relation is not unique, the
incidence angle can be converted into slope and slope into altitude by integration. Considering the
numerous assumptions required, the results obtained by this method show that it is more applicable to
morphological description of terrain than precise computation of altitudes.
Digital elevation models (DEM) obtained from radar images are still few and hence performance
of radar mapping is difficult to evaluate since the accuracy of DEM is considerably affected by Influences
of sensor, landscape or technique used. Simulation of radar images is used to verify various techniques
of three-dimensional restitution.

H Geology
Possibilities of radar applications in geology, in terms of accuracy of images and diversity of landscapes,
were demonstrated by SIR-A and SIR-B systems onboard the space shuttles in 1981 and 1984,
respectively. Various satellite data now available permit envisaging many operational applications.
The advantages of radar applications in geology are mainly related to the effects of surface
roughness or relief on backscatter. Through them, geologists identify lithological units that are exposed
or occur under a thin sand cover. Structural features corresponding to folds, fractures or veins can also
be marked. Distinction between flexures and faults could be made In the ERS-1 data acquired over
Levant fault in the Middle-East (Chorowicz et al., 1995). In desert climate, this marking can be done
directly while In temperate or inter-tropical climate, the effects of morphology on vegetation are used.
These data are incorporated in base maps (geological Inventories) or used for detailed investigations
such as mineral exploration, improving water resources, development of Infrastructure (siting railways
or roadways). Interpreted lithological characteristics are often complementary to those of satellite
444 Processing of Remote Sensing Data

images of the visible and near infrared bands. Thus some rocks show a smooth surface in the visible
band due to ‘desert patina’ but surfaces of various degrees of roughness in microwave data; this
feature enables their identification. That is why combined analysis of satellite data of various sensors
is important and operationally feasible.
However, many parameters that influence the backscatter signal hinder operational use of radar
data in geology. For example, suppression of relief, mentioned in the preceding section, may result in
erroneous estimation of dips, false reversal of layers or irregular drawing of faults. Moreover, although
the dielectric constant of many rocks Is known in mineralogy, use of radar images based on petrologic
models Is not yet developed (Deroin and Scanvic, 1995). In fact, due to the influence of geometric and
dielectric parameters on backscatter mentioned earlier, laboratory models cannot be extended to field
conditions.
In conclusion, for geological applications, radar cannot replace optical remote sensing but is
complementary, since information it restores is less detailed but close to the original. An excellent field
of application of microwave data is structural geology in the scales of 1:100,000 to 1:250,000.

■ Oceanography
Operational uses of radar images are most developed In oceanographic studies (sea ice, surface
state of sea, petroleum slicks, etc.) and, in particular, in coastal oceanography for the contact zone
between ocean and emerging lands. This zone has great ecological significance and constitutes venue
for varied and conflicting stakes (see Chap. 25). In general, coastal zone boundaries coincide with
those of the continental shelf (Forget and Cauneau, 1995). Depths rarely exceed 250 m before steep
gradients are indicated at the level of continental slope and reach values between 2000 and 4000 m.
Hence 7.5% of the surface of oceans and seas corresponds to coastal zone. Entire seas form its part:
the English Channel, North Sea, Baltic Sea, China Sea, Yellow Sea, the Sonde platform, and the vast
gulfs and bays.
It must be recalled that the littoral belt is characterised by a strong interaction between the
atmosphere (wind), hydrosphere (water and its movements: waves, currents, etc.) and solid constituents
of the continental lithosphere (Forget and Cauneau, 1995). The littoral zone depicts numerous and
varied forms: beaches, dune systems, cliffs, rocky coasts, salt marshes, lagoons, deltas, etc. These
environments undergo large variationsthat justify continuous monitoring and assessment, for example,
for surveillance of beach erosion or identification of shoreline. Proximity to coast and shallow depths of
open seas. In combination with the intensity of tides and associated currents, distinguish them from
high seas.
Remote-sensing data in the visible band are very useful for littoral inventory and surveillance;
however, radar data also show great significance due to their all-time and day-and-night capability,
particularly for intertropical or Arctic/Antarctic zones. In middle latitudes, their advantage lies mainly in
determination of parameters not possible with optical systems or In monitoring exceptional events
such as floods, storms, black tides, etc. In fact, delineating coastline from the visible-band data is
easier, less noisy and of finer resolution than from microwave data. Contrarily, observation of sea
roller (crest length, amplitude variations following crests, ‘group’ effects) and its movements and wave
surges, which enable delineation of certain bathymetric boundaries, as well as monitoring black tides
constitute important applications of microwave remote sensing.
Microwave data are best suited for monitoring hydrocarbon pollution, although other types of
images (visible, thermal) also assume certain importance. Systems of automated detection of slicks
are under development, but wind may severely hamper detection. Identification of physical (thickness,
age) or chemical characteristics of slicks seem possible from multiband (visible to microwave) data.
However, this application is not yet In operational stage since multifrequency and multipolarisation
radar systems are required. Time of access to data (less than 24 h) also constitutes a problem to be
solved to make the technique operational. Theoretical researches are undertaken to determine
mechanisms of attenuation of roughness by hydrocarbon layers in the presence of wind and waves.
 Applications 445

On the other hand, principles of detection of bathymetry by radar data are known but preliminary
attempts of modelling using controlled airborne experiments have so far resulted in an underestimation
of the phenomena and quantitative application in near future is not foreseen. The SAR seems to
constitute a valuable tool for description of internal waves and their dynamics, although ground data
are necessary for inferring certain characteristic parameters such as wave amplitudes. Microwave
data are capable of detecting surface expression of internal oceanic dynamics (internal fronts in
continental slopes and shoals). Similarly, in favourable conditions, coastal currents can be studied
using effects of roughness, form of slicks and refraction of sea roller.
Lastly, detection of ships and halieutic applications are under study. Resolution is the main
limiting factor. In fact, most fishing boats are not accessible to ERS imagery whose transverse resolution
is 26 m and longitudinal resolution 30 m. Detection of wakes, whose remanence is identified but
not explained, assumes calm sea conditions, not always satisfied. Detection of fish banks has been
studied by airborne radar but operational usage of satellite data is not feasible for reasons of spatial
resolution on the one hand and due to the sporadic nature of surface manifestation of fish banks, on
the other. Due to their specific properties, microwave data can be useful in fishing by their capacity to
locate thermal fronts and contrasts and, hence, oceanic regions where this resource is likely to be
present.
Study and surveillance of sea Ice constitute a domain of operational applications, since any-time
and day-and-night capabilities of microwave remote sensing are valued by conditions of cloud coverage
and duration of polar nights. In these regions, density and movement of ice are important markers In
Identification of general circulation in oceans. These data constitute useful information for
sedimentologists in studying transport of solid particles, for biologists in exploitation of biological
resources and for climatologists in investigation of depression zones. Further, monitoring sea ice
variations gives information on long-term climatic events (global warming, greenhouse effect). Lastly,
monitoring of ice has also economic importance. The need to know maritime current conditions,
especially for access to Baltic ports, imposes a strict surveillance of ice masses. Similarly, proper
functioning of offshore drilling platforms requires permanent checking of the state of waves and proximity
of ice masses. Satellite radar Images, operational and repetitive, represent a significant Information
support for these applications.

M Hydrology and snow gauging

Knowledge of Alpine type of snow or ice surfaces is very important for hydrological applications. In
fact, snow surfaces change in area from 47 million km^ to 4 million km^ from winter to summer (Dedieu,
1995). It is hence necessary to have seasonal maps of snow and estimate the corresponding water
equivalent. These applications are not yet in operational stage. In fact, mapping of snow Is possible
when it is wet on the surface but distinction between dry snow and non-snowy zones is difficult, unlike
in the case of passive microwave systems. The equilibrium line of glaciers is visible in SAR images,
facilitating reconstruction of annual mass balances; on the other hand, direct estimation of equivalent
water of snow cover is not proven, although polarimetry seems to be a promising tool.
Current researches are oriented towards use of multispectral and multipolarisation systems.
Specialists presently consider that a weekly repetition of observations is required with a resolution of
30 m during the period of snow melting. X-band (8 to 12 GHz) seems preferable over C-band due to its
higher sensitivity to surface roughness and a high incidence angle (45°) is essential in montane
regions.
Ice movements (30 cm d“ “^ to 2 m d” "*) have been monitored by interferometry using ERS data.
However, detection of melting ice vis-à-vis snow and temporal offsets of dry snow due to wind effects
are sources of errors.
The most advanced countries in research on such themes, for which scientific problems and
economic activities are closely related, are Finland, Switzerland, Austria and Canada.
446 Processing of Remote Sensing Data

H Hazard monitoring
Revisit capability of satellites together with any-time and day-and-night nature of radar are the basic
advantages for hazard monitoring. This is particularly established for monitoring floods in regions (or
periods) of intense cloud cover, for which visible and infrared data cannot be acquired. For example,
ERS data were used for monitoring floods of Oder river in Poland and Germany in 1997, and RADARSAT
data for assessing the magnitude of floods In Manitoba in May 1997.
Like the visible and infrared data, microwave images also enable surveillance of volcanic activity.
They are irreplaceable when volcanic activity occurs in high latitudes, as in Alaska and Iceland, under
a glacier in particularly unfavourable conditions of observation.
Interferometry, specific to microwaves, finds novel applications in monitoring earthquakes, fault
movements and landslides.
Cost of data and their processing and limitations in operational use of remote-sensing data for
some applications are less important in hazard studies in view of human lives and socio-economic
consequences Involved. It Is also true that this is a technology of rich countries and many hazards are
produced in poor countries that do not have the means of paying for such information.

■ Agriculture and forests

Application of microwave remote sensing for agriculture and forestry is least exploited and has
progressed little in the recent years, although radar imagery has great potential in this domain. The
main reasons are absence of multi-frequency and multi-polarisation facilities and unfavourable incidence
angles for the presently available data. Three major types of applications can be identified.

□ Monitoring and quantification of hydrological processes in soils

Various levels of investigation of aqueous state of soils are ‘entry’ (precipitation), transfer (on surface
by runoff, at depth by infiltration) and ‘exit’ (évapotranspiration, recharge of aquifers, river discharge,
useful reserves, etc.). At each level, various parameters need to be considered; climate, morphology,
soils, geology, vegetation cover, etc. Relationships of these parameters with water movement are
analysed and progressively modelled.
Radar contributes to evaluation of surface characteristics and analysis of relationships between
backscatter signal and the phenomena under study. In-situ experiments have been conducted to
select optimal instrumental conditions for determining surface moisture (frequency around 5 GHz,
incidence angle between 7° and 17°) and to identify the percentage capacity of the field from the first
5 cm of soil as being the expression of water content influencing the dielectric constant of soil and
hence the backscatter signal (Fig. 26.13).
Satellite radar images provide information over vast areas and hence reveal regional variations of
surface moisture. However, causes of variation of backscatter signal are such that data interpretation
most often remains qualitative. Use of diachronic data for creating colour composites, for example.
Indicates contrasts between dry and moist soils, bare or covered by various types of vegetation. In the
Sahel region, band C and incidence angle of 23° of SAR and ERS were found favourable for mapping
temporal variations of soil moisture provided that preliminary knowledge about the state of soil and
vegetation is available. In a totally different scale and other climatic conditions, scatterometer data of
ERS have been used for monitoring melting of permafrost In the Boreal zone.

□ Monitoring of crops and yields

Considering the operational nature of crop and yield monitoring by data acquired in the visible and
infrared bands, radar data are significant only for their any-time nature and high frequency of acquisition.
In this case, they provide complementary information, filling gaps In crop inventories for very cloudy
climatic periods. However, while differentiation of certain crops by microwave data has been
 Applications 447

Calibrated
backscatter
5- coefficient

0- 0 •

-5 -.
o:r
o 1980 — field capacity 23.5%
• 1981 — field capacity 22.2%

Soil moisture
(in % field capacity)

25 50 75 100 125

Fig. 26.13: Relationship between backscatter signal and soil moisture (after King, 1979).

demonstrated (Fig. 26.14), confusions exist between species (in particular in the case of cereals)
whereas geometric resolution of radar data is much lower than that of the visible and infrared data.
Lastly, the cost of data and their processing constitutes a serious limitation in their operational
usage for this type of applications.

□ Inventory and estimation of forest populations

It is mainly the structures and ages of population, rather than forest species, that are differentiated
from radar data. Identification of forest species and groups is much more difficult than with satellite
data of the visible and near infrared bands, particularly when they represent heterogeneous populations.
That is why microwave data are essentially used for surveillance of deforestation, especially in tropical
humid forests where cloud conditions hinder acquisition of data in other spectral bands. Even in this
case, interactions of waves with objects are not fully understood and data is interpreted qualitatively in
images earlier filtered for contrast enhancement and speckle reduction.

26.2.4 Conclusion
Many applications of microwave data still remain unexploited since studies are necessary to understand
the influence of various characteristics of objects on backscatter signal and the signal needs to be
processed before Its interpretation. Moreover, the number of satellite microwave systems remain limited,
their recent enhancements notwithstanding, and they offer restricted choice of spectral bands,
polarisations, incidence angles, etc. Lastly, for microwave remote sensing to become operational,
service societies have to develop. In fact, organisations directly concerned with applications just
mentioned above are interested In thematic information and not in remote sensing data.
All this assumes efforts In research, teaching and formation as well as Increased participation of
thematic specialists in specification of radar systems and products.
448 Processing of Remote Sensing Data

y (d B )
East-West axis
• 10

■ 10 -

\ XX

-20

Y (dB) North-South axis

j Bare
-2 0 -1 5 -10
X Bare rough soil • Corn 20-35 cm
smooth soil

(a) April 1980

V(dB)
East-West axis

-10

• •
• •

-1 5
Y (dB) North-South axis
-16.
-17 -15 -10
o Beet root 45 cm • Corn-Maize
80-100 cm, 45 cm

(b)June 1980

Fig. 26.14: Backscatter signal of various crops on different dates (Flight VIGIE band X, 1980; King, 1979).
 Applications 449

References
André J-C, Goutorbe J-P, Perrier A, Becker A, Bessemoulin P, Bougeault P, Brunet Y, Brutsaert W, Carlson T,
Cuenca R, Gash J, Gelpe J, Hildebrand P, Lagouarde J-P, Llyod C, Mahrt L, Mascari P, Mazaudier P, Noilhan
J, Ottle C, Payen M, Phulpin T, Stull J, SchmuggetT, Taconet O, Tarrieu C, Thepenier R, Valencogne C, Vidal-
Madjar A, Weilla A. 1988. HAPEX-MOBILHY: First results from the Special Observing Period, Ann. Geophys.,
6:477-492.
Assad E, Freteaud J-P, Kerr Y, Lagouarde J-P, Seguin B. 1986. Utilisation de la thermographie infrarouge dans
Testimation de l’évaporation à l’échelle régionale. Application au Sénégal, Agron. Tropicale, 40 (4): 279-285.
Chorowicz J, Kofi B, Chalah C, Chotin P, Collet B, Poli JT, Rudant J-P, Sykioti S, Vargas G. 1995. Possibilités et
limites de l’interprétation géologique des images (SAR) ERSI, Bulletin SFPT, 138: 82-95.
Courault D, Clastre P, Guinot J-P, Seguin B. 1994. Analyse des sécheresses de 1988 à 1990 en France à partir de
l’analyse combinée de données satellitaires NOAA-AVHRR et d”un modèle agrométéorologique. Agronomie,
14:41-56.
Dedieu J-R 1995. Application du radar spatial à l’étude des neiges et des glaces. Bulletin SFPT, 138:80.
Deroin J-P, Scanvic J-Y. 1995. Apport de l’imagerie radar à la cartographie géologique: exemples et réflexions,
Bulletin SFPT, 138: 96-109.
Forget P, Cauneau F. 1995. L’imagerie radar en milieu marin côtier et littoral: état de l’art et perspectives. Bulletin
SFPT, 138:73-79.
Goutorbe J-P, Label T, Tinga A, Brouwer J, Dolman AJ, Engman ET, Gash JGC, Hoepffner M, Kabat P, et al. 1993.
Hapex-Sahel: a large scale study of land-atmosphere interactions in the semi-arid tropics. Annales Geophysicae,
12:53-64.
King Ch. 1979. Contribution à l’utilisation des micro-ondes dans l’étude des sols. Thèse INA-PG, 122 pp.
Lagouarde J-P, Valery P, Belluomo P, Soulier M-A. 1983. Cartographie des topoclimats forestiers. Mise au point
d’une méthodologie d’analyse de l’effet du relief sur les thermographies: application aux données HCMM sur
le nord-est du Massif Central, Agronomie, 3 (10): 1011-1018.
Lagouarde J-P, Brunet Y. 1991. ‘Suivi de l’évapotranspiration réelle journalière à partir des données NOAA-AVHRR
lors de la campagne HAPEX-MOBILHY’, 5® coll int. Mesures physiques et signatures en télédétection,
Courchevel, ESP SP 319:569-572.
Moran MS, Jackson RM 1991. Assessing the spatial distribution of evapotranspirtion using remotely sensed inputs,
J. of Environmental Quality, 20 (4): 725-737.
Polidori L. 1995. Apport de la simulation d’images à la validation des techniques de cartographie radar. Bulletin
SFPT, 138:16-25.
Polidori L. 1997. Cartographie radar, Gordon and Breach Science Publishers, Canada, 287 pp.
 Glossary

b: Spectra! bands of SPOT are designated as: b1 (or XS1 ) for green, b2 (or XS2) for red, b3 (XS3) for
infrared and b4 for reflective infrared. bO is used for the first band (blue) of Vegetation instrument.
Band: Part of an image comprising a group of pixels pertaining to a single spectral band or derived
from a computation based on a single series of spectral bands.
Base: Distance covered between two successive aerial or satellite photos or images.
Bias: It measures the deviation between expectation of a series of measurements and the nominal
value, in other words, between the centre of gravity of the group and the expected value, i.e., the
reference value or nominal value. In Fig. 17.1, it is represented by the distance between the
centre (x) of the target (nominal value) and the centre (y) of the group of measured points. It can
be estimated by the intergroup distance vis-à-vis the reference value.
BIL (Band Interleaved by line): Storage format of spectral bands on satellite information base, line
by line.
BIP (Band Interleaved by Pixel): Storage format of spectral bands on satellite Information base, pixel
by pixel.
BSQ (Band Sequential): Storage format of spectral bands on satellite information base, scene by
scene.
CAP: Centre of Archival of Preprocessing of SPOT image, where radiometric and geometric correc­
tions are applied to images.
CCD sensors: Charge-transferring devices (Charge-Coupled Devices) or receivers. Several thou­
sand units are assembled In a single integrated circuit. The energy arriving at each unit creates a
proportional electric charge. The analogue value of charge is transmitted to a sampler which
converts it into a digital number.
CCT (Computer CompatibleTape): Magnetic tape comprising 9 tracks of 1600 bpi and 240 feet long.
It can store 331 Mbits. A LANDSAT image consists of about 300 Mbits and a SPOT image, 220
Mbits. In the beginning of the tape, input device (ID) is recorded, which includes the reference
number of the band, sensor, date of Image acquisition, mission, trajectory and range, preproc­
essing executed and adopted format.
Chorology: Chorology (from Greek khoros, meaning country or district, and logos, meaning logic or
science) is the study of relationships existing between characteristics of semantic (thematic)
units (internal factors) and their distribution in three-dimensional landscape (external factors).
Chroma: It is one of the three variables of the Munsell code used as reference for colour coding. The
higher the chroma, the smaller the grey content of a colour.
Compact: Said of a map unit whose perimeter is small compared to its area. The opposite Is referred
to as digitate.
452 Processing of Remote Sensing Data

Complex units: Several complex units exist:

Juxtaposition: Unit for which no chorological laws are defined.
Association: Unit for which defined chorological laws cannot be expressed graphically.
Sequence: Unit for which defined chorological laws are dominated by a preponderant factor such as
slope, topography, lithology or duration.
Combination: Unit for which a certain number of laws that characterise chorological laws are defined:
organisation of soils, soil types, size, reciprocal positions, etc.
Contrast: Ratio of the typological (mathematical) distance of two sites to the geographic distance
between the same sites. This value characterises the boundary state between the two sites.
DEM (Digital Elevation Model): Digital model of elevations in a region.
Digital characteristic: The set of digital numbers of each pixel for each band existing in an image.
Normally, for better perception, these values are connected by straight-line segments. It does not
represent a spectral characteristic since the various bands are not calibrated relative to each
other.
Digital number: Radiance (luminescence) value of a pixel expressed as an Integral number, often
over a scale of 256 values.
D igital-num ber spectrum : Set of digital numbers for each band, or for several bands, which corre­
spond to a geographically defined object In an image.
Exactitude: Narrowness of accord between a measurement or estimate of a parameter and its nomi­
nal value (IGN, 1997). It includes the two concepts of bias {b) and precision (p) by the equation:

It is often indicated by mean square error. The nominal value, by definition, serves as the reference. To
be exact, a measured value must be precise as well as without bias.
Field of investigation: Area targeted in an investigation or study.
Field of view: Largest area that can be sensed from a single point of observation.
Geographic Inform ation System (GIS): It is the information system that facilitates collection and
organisation, management, analysis and integration, preparation and presentation of geographic
data derived from various sources, ultimately contributing to spatial management.
GPS (Global P ositioning System): A system of 24 satellites orbiting around the Earth, which trans­
mit very accurate data about their position and distance relative to a receiver on the Earth. The
receiver, usually held in hand, sees the satellites that are above the horizon if the sky is clear and
no obstacles exist around (thick forest cover, walls, metal objects, people). Depending on the
number of satellites received (3 to 12), the GPS receiver indicates its geographic position in two-
or three-dimensions. A precision of about 1 cm can be achieved.
Graphic analysis: The part of spatial analysis devoted to shapes of map units, their position, neigh­
bourhood, dispersion, etc., as well as their representation: colour contrasts, drawing shapes, etc.
Group: A set of combined pixels derived from a classification.
Hue: It is one of the three variables of the Munsell code used as reference for colour coding. It Indi­
cates the respective quantities of red, blue, yellow, etc., of which the colour is composed.
IFOV (Instantaneous Field of View): Area from which a sensor receives a single signal whose value
is formed by summation of component values of the area.
Instantaneous field of view on ground: IFOV
 Glossary 453

Laser (Light Amplification by Stimulated Emitted Radiation): Amplification of light by stimulation

of emitted radiation, through resonance of stationary electromagnetic waves between two paral­
lel mirrors.
Level of analysis: Dataset that defines, for a map or a map unit, the type and number of variables
used as well as precision of measurements and observations.
Level of perception: It corresponds to the tools and types of measurements or estimates employed
for detecting a given level of organisation.
Lidar (Light Detection And Ranging): Detection with the help of light. This system operates similar
to radar and records reflection, from the Earth’s surface, of a wave emitted by a source.
LUT (Look Up Table): Colour table.
Map unit: Group of geographically distributed map zones having the same semantic content.
Map zone: Graphical unit having a closed boundary and a localised semantic content assumed ho­
mogeneous at a definite probability level.
Map: Orthogonal projection, on a plane, of a group of map zones representing organised groups of
objects defined by their boundaries (containei), shapes, contrasts and semantic contents.
Group of containers defined as an extension and representing sets of organised groups of objects that
are characterised by a comprehensively defined group of contents.
It is the result of a classification of objects; It shows spatial distribution and organisation of the objects
and their conceptualisation.
Mapping: It comprises the scientific studies and operations, art work and techniques employed on
results of direct observations or investigation of data for the purpose of preparing and establish­
ing maps, plan views and other modes expression, as well as their usage (International Carto­
graphic Association united with UNESCO, Paris, 1966).
It includes techniques that enable representation of objects after preparation of maps. In soil mapping,
it mainly consists of two-dimensional representation of soil cover which is a three-dimensional
system, incorporating the factor of time duration.
It simultaneously encompasses operations of spatial analysis (map production) and graphic repre­
sentation (cartography sensu stricto). Map production Is the process of understanding the organi­
sation of objects in a landscape, based on development of models according to chorological laws.
Mixel: A heterogeneous pixel which comprises several objects and whose spectral characteristic is
determined by the spectral composition of all the objects geographically contained in the resolu­
tion element.
It Is known or assumed that the composition Is derived from various, spectrally different, objects such
as vegetation, bare soils, water, clouds, shadows, etc. A very high digital number is observed on
images when passing from one semantic unit to another or over boundaries of units differing in
spectral characteristics.
Nadir: Point vertically below the optical centre of a camera.
Nucleus: Group of pixels used a priori to characterise an object for which a classification is to be
designed. This term is similar to ‘training zone’.
Pachy:Term In the Soil Reference Manual {Referentiel Pedologique) used to indicate an abnormally
thick soil.
Path: Ground track vertically above which a satellite passes. It is also the projection of the satellite
orbit on the Earth surface. For SPOT, each path is represented by two K-marks Indicated on
either side of the satellite trajectory.
454 Processing of Remote Sensing Data

Pattern: A group of pixels or objects, different but spatially organised, which repeat in a particular
manner. A pattern is hence a ‘homogeneous unit formed by units that are heterogeneous in
nature’. This constitutes a basic structural element.
Pixel: Term derived from contraction of picture element.
Precision: Narrowness of accord between a measurement or estimate and the mathematical expec­
tation of this measurement or estimate (David and Fasquel, 1997). It measures fluctuations of a
series of measurements around its mathematical expectation; It is given by the standard devia­
tion from the mean of the series of measurements. In Fig. 17.1, it is shown as a double arrow
which gives an estimate of the ‘diameter’ of the group of measured points. It can be determined
from the intragroup distance.
Pushbroom sensor: Group of sensors that acquire information on a row in a single sweep. The
sensors are pushed by the platform and hence the name.
Quarter rule: Any unit represented on a map must have at least 14 cm^ or, for elongated zones, more
than 2 mm in width and 1.5 cm in length. This old rule corresponds to a legibility limit.
For topographic maps with hatches, this rule Indicates that the hatches situated on the steepest
gradient are spaced at an interval of one-fourth of their length.
Radar (RAdio Detection And Ranging): Microwave sensor which generates an image of the ground
by means of temporal scanning.
Radiance or luminance (energy): Intensity emitted by unit apparent area along a direction q for a
point source of area dA, through a solid angle w.
Raster: When a data plane is considered, a grid placed above it delineates cells or pixels. Each cell is
made to correspond to a digital dataset, which is said to be in raster format. Unlike vector storage,
no real boundaries exist for each grid. Raster form is the most common for images.
Reflectance: Ratio of the energy In a wavelength band reflected by an object to the energy received
from the Sun by this object and for the same wavelength band.
Remote sensing: Data and techniques used for determining the physical and biological characteris­
tics of objects using measurements made from a distance, without physical contact of the latter
(JO of 11 December 1980).
Row: Along the trajectory of a satellite, images are divided over a certain north-south distance; the
marker of their centres constitutes a row (or line), indicated by letter J for SPOT satellites.
Scale: Ratio of a distance on the map (represented) to the same distance on the ground (representa­
tive).
Semantic analysis: Any part of spatial analysis which gives attributes for characterising pixels or
zones: digital number values, tree heights, soil colours, species present in a plant canopy, etc.
Soil landscape: Combination of soil horizons and landscape elements, viz., vegetation, effects of
human activity, geomorphology, hydrology, substratum or parent rock, whose spatial organisation
In its entirety determines a soil-cover or a part of it. A soil landscape often comprises several soil-
landscape units.
Its subgroups are: soil-landscape units and soil-landscape elements.
Soil-landscape unit Each unit corresponds to a soil system, often based on geomorphology, and can
be represented by a map unit (It can be divided Into soil-landscape elements).
Soil-landscape element Each corresponds to one or several typological units, spatially connected,
having a simple spatial organisation (a sequence, for example).
 Glossary 455

Spatial (geometric) resolution: The smallest area for which a single value is obtained for the variable
studied. The spatial resolution of a map can be determined by the ratio of the spatial field of the
map to the number of sites studied.
Spatial analysis: Spatial analysis consists of studying spatial relationships between various objects
distributed in a plane (or volume). Objects can be studied in terms of their abundance, estimated
by area or percentage of surface. It is also Important to evaluate the nature of neighbourhood
between various types of objects.
Spatial field: Largest area that can be analysed including all the sites studied.
Spectral band: Wavelength interval defined by two threshold values, the beginning and the end (for
example, 510-590). A spectral band measures only one energy (reflectance, radiance or digital
number) value, which is equal to the double spatial and radiometric integral between the two
limits, contained in a pixel.
Spectral characteristic: The set of radiance or reflectance values of an object for a set of spectral
bands.
Stripping: Parallel strips visible in LANDSAT images due to differences in sensitivity of photodetectors,
Structure: An aggregate formed by objects and their relationships. For thematic analysis in image
processing, it represents a system of organisation that depicts relationships between objects
studied (pixels, zones, etc.) but not their characteristics derived simply from their attributes or
digital numbers. OASIS, VOISIN, convolution filters, etc. are considered, from this point of view,
structural methods.
In soil mapping, it represents description, analysis and evaluation of textures produced by agglomera­
tion of constituents of varied grain-sizes and nature.
Surface state of soil: Composition and organisation of soil surface at a given moment. It takes into
consideration slaking crusts, salt efflorescence, stony nature, cultivation works, cracks and other
surface features, as well as coverage of soil by algae, moss and other vegetation.
TORS (Tracking and Data Relay System): System that enables the earth observation satellites
LANDSAT-4 or 5 to send images received from other satellites to ground stations.
Texture: Pattern formed by objects independent of their inter-relationships.
For mathematical analysis in Image processing, the term is almost equivalent to the term structural
used In thematic analysis.
In soil mapping, it represents arrangement of constituent ‘grains’, defined by their grain-size and na­
ture, considered independent of each other.
Thematic group: Population of pixels which are a posteriori, I.e., after a classification, combined Into
a single group and carry the same name as the nucleus used to represent them during classifica­
tion.
Tomography: Spatial representation of the material situated at a given depth from the surface of soil
cover.
Training zone: see Nucleus.
Value: Parameter of a colour, especially used in Munsell code. A ‘value’ is made up of the same
quantity of energy in blue, green and red spectral bands.
Wavelength: It is expressed in micrometers or nanometers.
Whiskbroom scanner: A sensor which acquires data through a rotating or oscillating mirror inclined
at 45 to the vertical and situated perpendicular to the direction of motion of the platform.
 General References
Specialised Scientific Journals on Remote Sensing
Major research papers are published in various journals exclusively devoted to remote sensing. Some of these are
listed below. Some of these publications can be accessed from Internet currently or in near future. With the possibilities
of information transfer through electronic networks, utility of publishing on paper becomes questionable in somuch
as it is often essential to give images on several illustrations of remote sensing applications.
Advances in Space Research

Bulletin de la Société Française de Photogram m étrie et Télédétection, 2, avenue, Pasteur, 94160 Saint-Mandé,
France.
Bulletin d ’inform ation de l ’Institut géographique national, 6-8, avenue Biaise Pascal, Cité Descartes, Champs-sur-
Marne, 77455 Marne-la-Vallée Cedex 2.
Canadian Jou rna l o f R em ote S ensing/Journal candien de télédétection, CASI, 222, rue Somerset Ouest, suite
601, Ottawa, Canada K2P 0J1.
EARsel, Advances Rem ote Sensing, Robin Vaughan (Editor in Chief), APEME, University of Dundee, Dundee,
DD1 4HN, Scotland, Grande-Bretagne.
Fotointerpretacja w geograffi, Katowice, Pologne.
IEEE Transactions on G eoscience an d Rem ote Sensing, IEEE Centre, P.O. Box 4122, Hong Kong.

In te rn ation alJourn al o f R em ote Sensing, University of Dundee, Dundee, DD1 4HN, Scotland, Grande-Bretagne.

M icrowave Rem ote S ensing A ctive an d Passive, Addison-Wesley Cy.

Photogrammétrie E ngineering & Rem ote Sensing, ASPRS, 5410 Grosvenor Lane, suite 210, Bethesda, MD 20814-
2160, États-Unis.
Photo-interprétation, Éditions ESKA, 27, rue Dunois, 75013 Paris, France.

Remote Sensing o f Environm ent, Elsevier Science Publisher, P.O. Box 211,1000 AE Amsterdam, Pays-Bas.
Remote Sensing Reviews, Harwood Academic Publishers, 1 Bedford Street, London WC2E 9PP, Grande-Bretagne.

Scientific Journals Comprising Papers on Remote Sensing

The following periodicals on various subjects also Include scientific articles on remote sensing.
Agronomie
Agronomy Journal
Agronomie tropicale
Annales Geophysicae
Annales de géophysique de France
458 Processing of Remote Sensing Data

Annales des Mines

Applied Optics
Bulletin de la Société française de minéralogie et cristallographie
Bulletin de la Société de préhistoire Nord
Comptes-rendus de l’Académie d’ agriculture de France
Comptes-rendus de l’Académie des sciences
Economie Geology
Étude et gestion des sols
Geoscience and Remote Sensing
Journal Optic Society of America
Journal of Atmospheric and Oceanic Technology
Journal of Environmental Quality
Limnology and Oceanography
Mémoires de l’Institut océanographique de Monaco
Progress in Phycological Research
Science du sol
Soil Science Society of America Journal
Sols
Water Resources Research

Ph.D.Theses
Ph.D. theses produced every year from the French and foreign universities constitute primary sources of information
on remote sensing. In acknowledgement of the cooperation of our collaborating scientists and former students, the
recent theses on remote sensing, recently awarded at the INA-PG and other institutes, are listed below.
Courault D., Étude de la dégradation des états de surface du so l p a r télédétection. A n a lyse s spectrales, spatiales
et diachroniques, Thèse INA-PG, 1989,239 p.

EscadafaI R., Caractérisation de la surface des sols arides p a r observation de terrain e t p a r télédétection. Aplicatlon;
example de la région deTataouine (Tunisie), Thèse de pédologie. Université Paris VI, ORSTOM, Paris, 1989,
317 p.
Gilliot J.-M., Traitem ent e t in te rp ré ta tio n d ’im ages sate llita ire s S po t: A p p lica tio n à l ’an a lyse des voies de
com m unication, Thèse de doctorat, Université ParisV, 1994,197p.
King Ch., Contribution à l ’utilisation des m icro-ondes dans l ’étude des sols. Thèse INA-PG, 1979,122 p.
Orth D., Typologies et caractérisation des prairies perm anentes des m arais du Cotentin, en vue de le u r cartographie,
p a r télédétection satellitaire, p o u r une aide à le u r gestion. Thèse INA-PG, 1996,149 p. plus annexes.

Robbez-Masson J.-M., Reconnaissance et déiimitation de motifs d ’organisation spatiale. Application à la cartographie

des pédopaysages. Thèse de doctorat, ENSA Montpellier, 1994,161 p.

Rogala J.-R, A pproche num érique de l ’espace agricole. Thèse docteur-ingénieur INA-PG, 1982.
Yongchalermchai C., Étude d ’objets complexes, sol/plante, à différents niveaux d ’organisation: de la parcelle au
paysage. Thèse de l’INA-PG, 1993, Sols, n°19, 232 p.
 General References 459

Proceedings
Several colloquia, seminars, summer schools etc. are organised, at the rate of almost one even per
week, throughout the world. This is a testimony to the active research in remote sensing. Proceedings
of some of these events, organised in France by the ONES, CNRS, French Association of
Photogrammetry and Remote Sensing or International Society of Photogrammetry and Remote
Sensing, are cited below:
Actes de télédétection IRT, La-Londe-les-Maures
BRGM, Manuels et Méthodes
Colloque International. Signatures spectrales d’objets en télédétection.
CNES, École d’été
CNIG, Groupe de travail
CNRS, École d’été de physique spatiale: Principes physiques et mathématiques de la télédétection.
IFEN, Groupes de travail
IGARSS Digest,Washington.
INRA, Les colloques
International Symposium of Remote Sensing of Environment
iRD-ORSTOM, Journées de télédétection. Études et Thèses

Other Publications
CNES Magazine (trimestriel)
Collection de l’École normale supérieure de jeunes filles
Earth Observation Quateriy, ESA
Éditions Cépadues
Spot magazine (semestriel)
Télédétection, journal du réseau de télédétection de I’AUPERLF-UREF
Télédétection satellitaire, Éditions Paradigme.

Books
General books or on specific topics of remote sensing are given below.

On General aspects
Bonn F. and Rochon G., Précis de télédétection. Vol. 1: Principes et méthodes. Vol. 2: Applications
thématiques. Presses de l’Université du Québec/AUPELF, 1992, 485 p.
Caloz R. and Collet C., Précis de télédétection. Vol. 3: Système d’information géographique et traitement
numérique, PUQ, à paraître.
Chevallier R., La photographie aérienne, Armand Colin, 1971,227 p.
460 Processing of Remote Sensing Data

Girard C.-M. and Girard M.-C. Applications de la télédétection à rétude de la biosphère, Masson,
1975, 186 p.
Girard M.-C. and Girard C.-M., Cours de photo-interprétation, INA-PG, 1970, 208 p.
Girard M.-C., and Girad C.-M., Télédétection appliquée. Zones tempérées et intertropicales, Masson,
1989, 260 p.
Lillesand T.M. and Kiefer R.W., Remote Sensing and image Interpretation, 3rd edition, John Wiley &
Sons, 1994, 750 p.
Monget J.-M., Cours de télédétection, CTAMN, Octobre 1994, Sophia-Antipolis.
Smith J.T.Jr, Manual of Color Aerial Photography, Am. Soc. of Photogrammetry, 1968.
Wilmet J., Télédétection aérospatiale. Méthodologie et applications, SIDES, 1996, 300 p.

On Specific Topics
Cocquerez J.-P. and Philipp S., Analyse d’images: filtrage et segmentation, Masson, 1995, 457 p.
Coster M. and Chermant J.-L., Précis d’analyse d’images. Presses du CNRS, 1989, 560 p.
Gonzalez R. and Woods R., Digital Image Processing, Addison Wesley, 1992, 716 p.
Guyot G., Climatologie de l’environnement De la plante aux écosystèmes, 1997.
Jensen J.R., Introductory digital image processing. A remote sensing perspective, Prentice-Hall, 1986,
379 p.
Kunt M., Traitement numérique des signaux, Dunod, 1981,402 p.
Lliboutry L., Sciences géométriques et télédétection, Masson, 1992, 289 p.
Mather P.M., Computer processing of remotely-sensed images. An introduction: 2nd Edition, Wiley,
1999, 292 p.
Mulders M.A., Remote sensing in soil science, Elsevier, 1987, 379 p.
Paquet G., Détection électromagnétique: fondements théoriques et applications radar. Masson, 1997,
320 p.
Photographie aérienne et urbanisme. Centre de recherche d’urbanisme, Paris, 1969.
Polidori L., Cartographie radar. Gordon and Breach Science Publishers, Canada, 1997, 287 p.
Pratt W., Digital image processing, 2nd edition, Wiley, 1991,698 p.
Rosenfeld, Kak, Digital picture processing, Academie press, 1982.
Rousselet M., Graphisme 3D, ETSF, 1985, 223 p.
Serra J., Image analysis and mathematical morphology. Academie press, 1982.
Worboys M.F., GIS, a computing perspective, Taylor & Francis, 1995, 376 p.
Wyszecki G., and Stiles W.S., Color Science: concept and methods, quantitative date and formulae,
Wiley, 1982, 950 p.
 Useful Internet Sites
Nowadays enormous amount of information is becoming available or INTERNET. This is particularly
very useful in the case of remote sensing data which is based on numerous images. A list of important
sites on remote sensing is given here. These sites are regularly updated and hence provide recent
information. This is particularly so in the case of development of satellite sensor and platforms.
For any details regarding the book, CD-Rom or the latest version of the OASIS program, contact:
[email protected]. inra. fr
Girard ©lacan.grignon. inra.fr

Information Available in French

http://www.spotlmage.fr
SPOT Image: for any information site in french and english
5, rue des Satellites, BP 4359, 31030 Toulouse Cedex 4
Tél.:05 62 19 40 40
http://www.spotimage.fr/accueil/siriuswelcome.htm
http://Lacan.grignon.inra.fr/resources/resources.htm.
For Vade-mecum books and courses on Remote Sensing télédétection et Cours de télédétection.
http://sol.ensam.inra.fr/silat
For enquiries about the specialised “Mastere Systems” on local information for area management:
French data on remote sensing and GIS (Bac + 6 level)
http://www.ign.fr/sfpt
Société française de photogrammétrie et télédétection.
http://www.cnig.fr
Conseil national de l’information géographique.
136, bis, rue de Grenelle, 75700 Paris 07SP
http://www.lgn.fr/GP/photaer/exemples.html
Institut géographique national: commande de photographies aériennes,
http://
État des cartes de sols pour la France.
http://vi vlane. roazhon.inra.fr/snas/index. html
Analyse de terre pour la France.
http://www.gdta.fr/
Groupement pour le développement de la télédétection spatiale.
462 Processing of Remote Sensing Data

http://www.ccrs.nrcan.gc.ca
Centre canadien de télédétection. Quelques sites en télédétection y sont donnés.
http://www.aupelf-uref.org
Le réseau de télédétection de rAUPELF-UREF.

Information Available in English

http://www.auslig.gov.au/acres/prod_ser/
SPOT-images receiving Centre at Hobart and Alice Springs.
http://www.ivv.nasa.gov
NASA Information Centre which regularly issues LANDSAT-TM images.
http://hdsn.eoc.nasda.go.jp/guide/homepage.html
Japanese National Remote Sensing Centre.
http://www.noaa.gov
NOAA/Natlonal Geophysical Data Center, 325 Broadway, E/GC4, Dept 993
Boulder, CO 80303, USA.
http://www.spaceimaging.com/
Spaceimaging society, istributor of IKONOS Images.
http://www.ceo.org/
Centre for Earth Observation-lspra: European Remote Sensing Program (Chap. 17).
http://www.esa.lnt/
All information concerning the European Space Agency.
http://www.digltalglobe.com/
Earthwatch Society, distributor of QuickBird satellite (4m). Available archives available on http://
archivedigitalglobe.com/
http://www.earsel.org
European Association of Remote Sensing Laboratries.
http://seawifs.gsfc.nasa.gov/SEAWIFS/LICENSE/checklist.html.
Orblange Society, distributor of Seawifs data.
http://www.vgt.vito.be
SPOT 4-Vegetatlon program of CNES
http://www.terra_story_barc.asp/terraserver.microsoft.com
Microsoft Global Service of High resolution imagery.
http://www.inpe.br/english/lndex.htm
INPE National Space Centre of Brazil, Distributor of data.
http://eospso.gsfc.nasa.gov/
NASA EOS program.
http://edcwww.cr.usgs.gov/earthshots/slow/tableofcontents
USGS collection of images and case histories.
http://edcwww.cr.usgs.gov/
EROS Data Center, World distributor of data.
http://makalu.jpl.nasa.gov/
AVIRIS Microwave data server at JPL/NASA.
 Useful Internet Sites 463

http://usgs.gov/products/satellite/tm.html
USGS.information about LANDSATTM data
http://southport.jpl.nasa.gov/
Archives of synthetic aperture radar JPL7NASA.
http://radarsat. space.gc.ca/
RADARSAT society in Canada.
http://rsi.ca
RADARSAT International data centre for Information on radars and applications.
http://ewse.ceo.org
Shows existing organisations and material (CD-ROM etc.)

Data P rocessing Softw are

http://lacan.grignon.inra.fr/resources/resources/htm
TeraVue (Remote sensing).
http://www.erdas.com
Erdas (Remote sensing).
http://www.ermapper.com
ER/Mapper (Remote sensing).
http://www.esri.com
ESRI (CIS).
http://www.clarklabs.org
Idrisi (CIS).
http://www.maplnfo.com
Mapinfo (GIS).
http://www.khoral.com
Khoros (Image processing).
http://www.noeslsvision.com
Noesis (Image processing).
 Index
Aircraft 34, 38, 255, 256, 258, 260, 262, 264, 270,
272, 274, 438
Abiotic components 283 Albedo 7,9,19, 88,135
Absorptance 73 Algae 8 8 , 378, 386, 414, 418, 424, 426, 454
Absorption 10, 13, 14, 15, 29, 36, 57, 73, 74, 79, 80, Alignment 26, 100, 155, 249, 316, 320, 322
83, 90, 173, 378, 384, 386, 388, 406, 408, 418, Alluvial cone 104
436 Alluvial plains 100
Absorption bands 10, 73, 79, 83, 386, 406 Alluvial zones 155
Accommodation distance 265 Altimeter 50,241,416,438
Accords 334, 336, 339, 340, 341,342 Altitude 14, 23, 32, 34, 38, 40, 41,44, 46, 48, 50, 52,
Achromatic axis 60, 61,63 111, 241, 242, 243, 246, 249, 250, 252, 255,
Achromatic point 62 256, 260, 262, 264, 266,268,270, 272, 274, 276,
Acquisition 15,16, 31,34, 36. 41,42, 43, 44, 50, 52, 278, 302,304, 318,346,364,366,422,434,438,
72,114, 122, 125, 129, 142,156, 174,188,189, 440, 442
191,221,233, 234, 243,246,249, 252, 255, 256, Altitude positioning 302
262, 286,293, 294, 296,300,302, 304,316,318, Alunite 388, 406, 408
332,340,348, 352, 360,368,373, 376, 378,380, Amplitude 13, 24, 208, 223, 224, 226, 228, 229, 406,
386,392,396,410,416,422,424, 426,430,438, 408, 422, 424, 440, 444
442, 446, 451 Anaglyph 268, 276
Acquisition conditions 293, 294, 302 Analytical method 97, 226
Active remote sensing 21,422 Angle of convergence 264
Active system 20, 23, 24, 60, 61,66 Angle of incidence 7, 23, 26, 27, 29, 30, 50, 78,156,
Actual values 138 249, 438
Additive 57, 59, 60, 61,68, 226 Angle of inclination 40, 80, 258
Additive colours 57, 60 Angle of observation 314
Adequacy of data 300 Animation 243
Aerial photography 6 6 , 255, 258, 260, 262, 276, 278, Apollo 48, 52, 256
284, 318,340, 356 Approximation 18, 118, 178, 199, 200, 201, 202,
Aerial phytomass 80, 84, 85, 8 6 , 286, 344, 346, 348, 203, 204, 205, 206, 224, 231,234, 235, 241,244,
350, 352, 354, 355, 356 245, 249, 250, 252, 290, 304
Aerodynamic temperature 16,17,19 Aquaculture 414,420,422
Affectation errors 303 Aries 410, 411,414, 420, 438, 442, 444, 452, 454
Agglomeration 101, 153, 155, 318, 332, 454 Arithmetic combination 130,132,133,134,136,138,
Aggregation 284, 290, 332 141, 160
Agricultural activity 84, 378, 380 ARTEMIS 44
Agricultural calendar 286, 294 Artificial intelligence 154,358
Agricultural plots 99, 155, 156, 157, 158, 172, 174, Ascendant hierarchic classification 142,143,144,146,
314, 320, 326, 332, 362 148,150,153,154,156,158,159,160,161,163,
Agriculture 18, 36, 44, 73, 8 6 , 108, 278, 332, 371, 165, 166, 175, 178, 188, 362
372, 376, 394, 396, 446 Ascendant methods 117
Agrolandscape 313, 314, 320, 322, 326, 328 Ascendant approach 291
Agrolandscape elements 314 Association 38, 114, 155, 168, 174, 192, 194, 284,
Agrolandscape units 314, 320, 322, 326, 328 304, 314, 332, 390, 402, 404, 406, 408, 422,432,
Agrometeorological model 372, 374, 430 452
AHC 150, 152, 156, 159, 160, 165, 166, 352, 362, Atmosphere 9 , 1 0 , 1 1 , 12, 13, 14, 16,19,20,21,38,
364 41,50, 87, 90,114, 314, 378, 386,432, 442, 444
466 Processing of Remote Sensing Data

Atmospheric 4. 9, 10, 11, 12, 13, 14, 15, 18, 19, 21, Biomass 304
36, 44, 67, 6 8 , 79, 84, 8 6 , 87, 90, 122, 128, 255, Biotic components 283, 284
294, 416, 430, 432, 436 Bit 6 8 , 69
Atmospheric absorption 10,14, 36 Black 59, 60, 61, 65, 6 6 , 6 8 . 69, 121, 123, 125, 126,
Atmospheric effects 12, 15, 36, 8 6 , 87, 430 127,129,130, 134,139,153,155,160,175,181,
Atmospheric perturbations 9, 87 182, 256, 258, 268,276, 278
Atmospheric radiance 12, 13, 14, 18, 19 Black body 3
Atmospheric radiation 4, 12, 13, 14 Black tides 444
Atmospheric scattering 11,12 Black-and-white infrared 8 8
Atmospheric windows 4, 10 Blue 10,11,48, 52, 57, 58,59, 60, 61, 62, 63, 64,66,
Atolls 422 67, 6 8 , 69, 70, 71, 73, 75, 127, 128, 129, 130,
Automated classification 350, 360, 373 134,137,139,140,160,166,200, 203,209,255,
Automatic mapping 332, 336, 342 268, 318, 378, 382, 384, 388, 390, 392, 398,408,
Averaging filter 216, 220 418, 451,452, 454
AVHRR 2 , 13, 15, 44, 46, 48, 8 6 , 290, 374, 376, 405, Blue-violet 418
410, 426,430, 432 Boundary 9, 10, 58, 6 6 , 8 8 , 1 1 0 , 142, 154,158, 163,
AVIRIS 38, 79 166,169,174,193,197. 215, 216, 219, 221.228,
Azimuth 6 , 41, 156 230,231,252, 286,288,298,302, 316,360,362,
410,432, 452
B Boundary mixel 154, 204, 288, 316
Brightness 13, 62, 63, 104, 121,127, 129, 135,151,
B1 42, 87, 89,117,118,121,122,123, 130,131, 136, 154,155,158, 160, 316, 378,382, 384, 386, 388,
137,138,161,168,169,170,181,236, 336,451 420
B2 42,44, 87, 89, 117, 118, 123, 130, 131,136, 137, Brightness index 135,420
138,142,161,163,168,169,171,172,174, 236, Brightness temperature 13
300, 360, 451 Buildings 129, 130, 249, 318, 356, 442
B3 42, 44, 237, 288, 360, 117, 118, 122, 123, 124, Burning 353, 409
125,130,131,138,139,142,161,164,169,170, Byte 52, 210
171, 174, 181,451
band C 437, 446
band K 437
band L 437 Cband 63, 127, 136, 138, 153
bandTM 6 X Calcite 388
band X 448 Calcium 82, 83, 84
Backscatter 5, 22, 23, 24, 25, 26, 27, 28, 29, 30, 50, Calculation of distance 288
422, 424, 436, 438, 440, 442, 444, 446, 448 Calibration 23, 34, 90, 128, 142, 249, 308, 422, 432,
Backscattered 23, 25, 28, 30, 422, 436, 438, 440 438
Band combinations 128 Calm water 442
Bare soil 9,13, 15, 18, 26, 27, 28, 29, 44, 48, 70, 75, Cameras 32, 52, 255, 256, 258, 260, 264, 272
79, 81, 104, 111, 117, 129, 130, 134, 135, 139, Canny filter 226
145,148,149, 153,154,155,157,161,165,166, Canopy 16, 17, 18, 19, 29, 75, 79, 80, 85, 8 6 , 167,
167,168,169,170,172,174,179,182,184,187, 172, 174, 283, 286, 294, 306, 356, 358, 418,
294,296, 316, 318, 322, 328,332, 334, 348, 350, 454
356,364, 366, 378,384,392,394,414,420,422, Capillary waves 30, 422
430, 434, 442, 452 Carbonates 383, 384, 390
Base 255, 256, 262, 264, 266, 268, 274, 276 Carbon dioxide 10
Basic plan 114 Cardinal 290, 306
Bayes decision rule 176 Cardinal units 306
Bathymetry 8 8 , 416, 444 Carotene 73, 74
Bench 106, 108 Cartesian space 229, 235
Bias 296, 300, 304, 306, 358, 366, 451,452 Cartographic generalisation 252
Bicubic interpolation 237 Cartographic method 127
Big Bird 52 CASi 38, 121,418,420
Bilinear interpolation 236, 237 CCD 36, 38, 44, 46, 52, 222, 256, 451
Binary segmentation 125 CCD sensors 44, 451
Biological efficiency 85 CCT 451
 Index 467

Cellulose 79 Co-supervision processing 191

Changes 16,19, 78, 79, 80,111,136,176,195,197, Coarse components 388
204, 205, 206, 226, 238, 256, 260, 274, 283, 284, Coastal 30, 44, 8 8 , 89, 90, 332, 414, 416, 418, 422,
294, 296, 330, 342, 348, 356, 359, 373, 374, 388, 424, 426, 444
390, 405,414, 432,440 Coastal currents 444
Changes in vegetation 356 Coastal zones 414, 416, 422, 426
Chlorophyll 72, 73 Coastline 414,424,444
Chlorosis 294 Coasts 414,430,444
Chorological 96, 103, 105, 106, 112, 115, 133, 158, Code 52, 57, 63, 6 8 , 69, 71, 99, 101, 108, 118, 1 2 1 ,
159,172,173,175,188,189,191,192,284,286, 126, 127, 129, 130, 134, 148, 149, 201, 210,
313, 318, 326, 394, 396, 398, 452 211,326, 362,374, 378, 388, 394,398,451.452,
Chorological law 96, 103, 105, 106, 112, 115, 133, 454
158,159,172,175, 284,286,313, 318,326,394, Coefficient of absorption 90
396, 398, 452 Coherence 24, 34, 187, 283, 394
Chorological model 188, 191 Coherent monochromatic wave 25
Chorology 191,451 Colour 5, 11, 44, 52, 57, 58, 59, 60, 61, 62, 63, 64.
Chroma 5, 25, 36, 42, 44, 46, 50, 52, 58, 59, 60, 61, 65, 6 6 , 67, 6 8 , 69, 70, 71,75, 76, 80, 82, 83, 8 8 ,
62, 63, 65,66, 6 8 , 109, 121,126, 127, 135, 138, 90, 94, 97, 103, 109, 110, 113, 114, 121, 122,
139,198, 219, 256, 258,268,294, 313, 360, 378, 123,126,127,128,129,130,131,133,134,136,
416, 451 140,141,142, 148, 151,152,153, 158, 159,160,
Circular 40, 41,42, 44, 46, 100, 108, 304, 316, 318 165,166,167, 168,170,172,174, 175,178,181,
Civilian time 9 182,184,189,192,193,199, 201,209, 219, 223,
Classes 69, 70, 104, 115, 116, 121, 125, 126, 130, 255, 256, 258, 264, 278, 318, 320, 322, 328, 330,
133,134,136,141,142,143,158,167,168,169, 332, 334, 336, 350, 356, 360, 362, 366, 368, 378,
170.171.172.175.177.183.187.194.198.199, 380, 382, 386, 388, 390, 392, 398, 402, 406,408,
200, 203, 204, 205, 224,284, 286, 288, 290, 298, 418,426, 446, 451,452,454
302, 304, 306, 308, 318,320, 328, 330, 332, 334, Colour code 63, 69
336, 339, 340, 342, 356, 362, 364, 366, 368, 390, Colour combinations 70,110, 127
418 Colour composites 57,114,128,129,130,140,141,
Classification 69, 94, 97, 107, 110, 113, 114, 115, 256, 328, 330, 332, 360, 362, 446
116,117,118,119,120,121,125,127,133,135, Colour infrared 5, 52, 67, 6 8 , 71, 8 8 , 97, 159, 258,
136,137,140, 141,142,143,144,145,146,147, 278
148,150,153,154,156,157,158, 159,160,161, Colour perception 57, 69, 71
162,163,165,166,167,168,172,173,175,176, Colour printing 71
177,178,180,181,182,183,184,186,187,188, Colour representations 69
189.190.191.192.193.194.197.198.199, 200, Colour sensation 57, 58, 59, 63
201,202, 203, 204, 205,206,216, 229, 286, 288, Colour table 69,452
290,294,296, 298,302,304,306, 308, 318,320, Combination 58, 59, 70, 84, 85, 1 1 0 , 113, 114, 115,
326,328,332, 334, 336,339,340, 342,348, 350, 121,127,128,130,132,133,134,136,138,139,
352,356,358, 360,362,364,366, 368,373, 398, 141,142,154,155,160,161,173,174,193,194,
402, 404, 420, 424, 452, 454 214, 233, 235, 240, 241,242, 246, 252, 255, 286,
Classification probabilities 181 288, 290, 294, 308, 314, 316, 326, 348, 350, 356,
Clay 172, 184, 276, 352, 360, 382, 388, 394, 396, 366, 368, 394, 398, 402, 430, 444, 452, 454
406, 408 Commission error 340, 342, 368
Clayey 172, 184, 360, 388, 394, 396, 408 Communication lines 98
Climatic efficiency 85 Compactness 155, 194, 199
Climatic variations 294, 430 Complementary colours 58, 61,69
Climax 344, 346 Complex map units 194
Clinometer 266 Complex units 336, 452
Closure 98, 99, 218, 219, 222, 228, 231,276 Composition vector 194,195,196,198,204,205,298,
Closure of edges 228, 231 314
Cloud 30, 141, 158, 161, 171, 174, 180, 182, 184, Compression 244
249, 264, 392, 430, 432, 442, 444, 446 Computer interpretation 110, 296
Cloud cover 141,249, 430, 442. 444, 446 Computer monitor 57, 69, 109, 111, 127, 189, 198,
Cloud coverage 430, 442, 444 246, 256, 328, 334, 350, 374
Cloudy 358, 430, 436, 446 Concealed points 242, 243
468 Processing of Remote Sensing Data

Concept of form 217 Crystals 91

Conditional probabilities 177, 181 Cubic convolution 290
Cone cells 57 Cubic representation 60
Confusion 302 Currents 30, 8 8 , 129, 414, 416, 418, 444
Confidence interval 304 Curvatures 302
Coniferous 172, 302, 318, 320, 324, 326, 332, 334, Cyan 57, 60, 61,66, 67, 70, 128, 129, 200, 318
336, 338,339, 340, 341,342 CZCS 44, 90, 408

D
Connectivity 212,326
Constructions 101, 109, 130, 249, 414
Continuous anamorphosis 124
Contrast 58, 70, 8 8 , 94, 97,102, 109, 110, 115, 121, Date 38, 48, 50, 52, 81, 82, 97, 107, 110, 112, 114,
124,126,127, 128,129,138,139,142,154,176, 117,122,127,136,141,154,156,174,188,229,
192, 216, 224, 249, 251,264,304, 316, 318, 322, 249, 260, 264, 276, 278, 283, 293, 294, 296, 298,
326, 328, 332, 384, 398,402,405, 408, 410,444, 300, 316, 318, 320, 322, 326,328, 332,336, 338,
446, 452 340, 348, 352, 356, 359, 360, 362, 366, 368, 374,
Control points 244, 245, 246, 249,250, 251,252,304, 378, 390, 392, 394, 398,408,414, 416, 418,422,
306, 332, 336, 338, 339, 340, 342, 364, 368 432, 438, 448, 451
Controlled mosaic 272, 274 Date of acquisition 52,114,174, 249, 293, 294, 296,
Convolution 212, 213, 214, 216, 220, 226, 237, 238, 316, 332, 340, 348, 360, 368
239, 290, 454 Date of photography 260
Convolution integral 226 Datum plane 249
Convolution mask 214, 238, 239 Decision making 110, 119, 401,404, 426
Convolution method 237, 238 Deformation 36, 234, 235, 243, 244, 245, 246, 248,
Convolution product 212, 213, 214, 216 249, 250
Copper 406 Degraded 90,291,352,382
Coral reefs 414 Delaunay triangulation 244, 245
CORINE land cover 145, 150, 198, 296, 304, 308, DEM 29, 44, 52, 106, 111, 115, 120, 164, 165, 166,
316, 330, 332, 334, 336, 338, 341,342, 366 170,241,242, 246, 248, 250,252, 255, 276,278,
Corn 60, 81,82, 108, 176, 330, 360, 434, 442 293, 316, 318, 322, 328, 360, 362, 368, 376, 390,
Corner effect 442 405, 410, 424, 442, 446, 452
Cornice 108 Dendrogram 116,144,145, 146,147,148, 149,157,
Correlation 29, 81,130,131,135,137,141,169,172, 362
175,191,276, 292, 296, 306, 316, 332, 340, 355, Density 5,15,16,19, 61,66,75, 81,97, 98,101,104,
356, 360, 374, 382, 390, 392, 404, 405, 418, 142,176,304, 316, 318,326,352, 404,405,408,
420, 432, 434 418, 422, 434, 444
Correlation coefficient 131, 137, 141,296, 355, 360, Deposit 401,402, 404, 405, 406, 408, 410, 412, 418
432 Depression 23, 27, 100, 256, 360, 410, 438, 444
Cosmos 52 Depression angle 23, 27, 438
Coverage 27, 34,42, 52, 79, 80, 81,82, 86,132,154, Depth 13, 19, 30, 31, 83, 8 8 , 89, 90, 128, 155, 171,
167,169,171, 174,183, 215,235, 255, 256, 258, 172, 284, 318, 386, 401,408,418, 422,444,446,
260, 264, 272, 274, 278, 290,316, 322, 326, 328, 454
336, 348, 350, 355, 374,376,404, 405, 410, 424, Descendant approach 287
430, 434, 436, 442, 444, 454 Descendant methods 116
Cracks 84, 378, 454 Description 16, 31, 84, 99, 107, 108, 109, 111, 113,
Criteria of Choice 293 115,161,210, 268, 284, 296, 302, 320, 322, 326,
Crop identification 294 328, 378, 396, 432, 442, 444, 454
Crop inventory 294, 371 Desert varnish 386, 411
Crops 12, 46, 70, 81, 85, 125, 134, 141, 148, 149, Detection 20, 32, 23, 50, 52, 6 8 , 70, 75, 106, 113,
152,154,156,157,161,166,168,170,171,172, 128,130,150, 208, 222, 223, 224, 226, 227, 228,
174,179,180, 182,183,184, 201,278, 286, 290, 230, 231,232, 266, 294, 302, 318, 326, 227, 386,
294, 302, 304, 316, 318,320,322, 324, 325, 326, 388, 390, 398, 401,404,405,408, 412,416,422,
332, 334, 336, 344, 346, 348, 350, 352, 356, 360, 424, 426, 430, 434, 444, 452, 454
362, 364, 366, 368, 372, 373,374, 376, 382, 396, Detector 32, 34, 36, 48, 216, 232, 454
410, 412, 430, 432, 446, 448 Development 17, 26, 31, 36, 50, 52, 65, 67, 80, 8 6 ,
Crusts 84, 8 6 , 378, 380, 390, 394, 454 90, 101, 115, 154, 174, 252, 255, 256, 262,
 Index 469
278, 286, 293, 294, 308, 330, 355, 358, 360, 372, Distortion 36,110,119,234, 235, 243, 244, 245,246,
406,408,412, 414,420,422,424, 426,442,444, 248, 250, 256, 272, 274, 316, 440
452 Distribution 10, 26, 71,72, 84, 96, 9 7 ,100, 101,103,
Diachronic 70,88,136,137,141,187,189,191,234, 117,121,126,133,134,136,139,145,146,147,
246, 288, 290, 294, 314, 318, 350, 356, 358, 362, 156,159,169, 175,176,177,183,191,192,194,
366, 368, 398, 405, 420, 422, 446 214, 216, 217, 221,250, 251,252, 288, 304, 306,
Diachronic analysis 141,234, 420 308, 314, 316, 318, 326, 346,348, 350, 359, 394,
Diachronic classification 368 396, 402, 424, 432, 440, 451,452, 436
Diachronic data 8 8 , 290, 294, 350, 358, 366, 446 Diversity 153,161,302, 318, 346,359,362,418,420,
Diachronic study 141, 191,362, 356 442
Diachronic variations 288 Dolerite 408,410
Diaspore 406 Doppler 22, 438
Dicotyledons 80, 352 Doppler effect 22, 438
Dielectric 21,26, 28, 29, 30, 31,444, 446 DORIS 44
Dielectric constant 21,28, 29, 30, 444, 446 Drainage networks 102, 1 1 1
Diffuse 7, 27, 72, 89, 322, 442 Drought phenomena 8 6 , 430
Diffuse attenuation 89 Dry 13, 18, 28, 29, 30, 75, 79, 80, 81, 100, 129, 130,
Digital cameras 256 174,179,180,181,182,183,184, 295, 318, 345,
Digital characteristic 128, 150, 151, 152, 153, 157, 346, 349, 352, 353, 355, 364, 366, 386, 387, 388,
158, 159, 172, 173, 176, 178, 184, 452 390, 391,392, 394, 409,421,430, 433, 435,436,
Digital classification 114, 296, 298 445, 446
Digital elevation model 43, 241, 249, 276, 320, 328, Dry matter 80. 85, 8 6 , 345, 346, 347, 353, 355, 356,
342, 364, 394, 398, 434, 442, 452 364
Digital filtering 208, 113, 214, 233 Dry soil 19,30,386,392,408
Digital image 110,115,208, 209, 212,215, 223,224, Dry vegetation 75, 130, 394
235, 256, 276 Drying of soils 294
Digital interpretation 150 Dynamic range 123, 124, 134, 135, 136, 138, 139,
Digital model 20, 132, 426, 452 296, 362
Digital number 70,121,122,123,124,125,126,128, Dynamic-range enhancement 128

E
130,131,132,133,134,136,138,140,141,142,
143,146,148,150,151,152,153,154,157,158,
163,166,167,168,169,171,172,173,176,178,
179,191,283, 286, 288,290,308,328,348,362, Earth Observation Centre 308
364,366,378, 380, 390,392,398,408,451,452, Earthquakes 446
454 Ecosystem 283, 344, 418, 420, 426
Digital-number spectrum 452 Edge detection 208, 222, 224, 226, 227, 228, 230,
Digitate 192, 193, 360, 451 231,232
Digitisation 102, 128, 194, 210, 248, 256 Edge detector 216
Digitising table 246, 248 Efficiency 73, 85, 8 6 , 210, 224, 352, 371
Dips 444 Efficiency of interception 73, 85, 352
Direct transformation 235 Efflorescence 83, 378, 380, 382, 390, 420, 454
Diseases 75, 79, 294, 374 Elevation 415
Dispersion 70, 155, 354, 356, 416, 452 Embankments 99
Display 6 8 , 110, 111, 116, 118, 120, 1 2 1 , 123, 126, Emission 3, 7, 13, 14, 20, 21,24, 69
127,132,134,140,148,153,158,189,199, 237, Emissivity, 16
243, 246, 251,330, 360 Emulsions 65, 6 6 , 67, 256, 258, 276, 278
Distance 6,9,22,23,24. 25.38,63,67, 86,115,117, Energy 3, 4, 5, 6 , 12, 18, 19, 20, 21,22, 23, 25, 32,
118,119,120,126,144,147,156,159,161,162, 34, 36, 40, 65, 6 6 , 6 8 , 89. 128, 231, 388, 408,
163,167,193,194,195,196,197,198,199,201, 414, 451,454
202,203,204, 205,206,209,212,219,227,243, Energy absorbing efficiency 85
245,246,250, 256,262,264,266, 268,270,272, Environmental factors 314, 348
274,283,286, 288, 300,302,314, 316,322,326, ENVISAT/MERIS 90
362, 398, 405, 408, 438, 440, 451,452, 454 EOSAT 14, 40, 44, 46, 48, 410, 432, 434, 436
Distances 24,115,117,118,119,120,126,144,147, Episyenite 402, 404, 410
159,163,194,198,199,202,203. 204,205,206, Equal population function 124
264, 266, 268, 274, 288, 302, 405 Equivalent water thickness 80, 362
470 Processing of Remote Sensing Data

Erectophyll 80, 353 Expanded 227

Eroded 218,227,401 Expansion 142,218,219
EROS 48 Exponential function 124
Erosion 155,184, 218, 219, 328, 336, 346, 348, 360, External factors 83, 451
368, 394, 396, 398, 402, 416, 422, 424, 444 Extrapolation 17, 115, 244, 250
Error 12,14,16,18, 30, 36, 8 6 , 87, 90,111,119,146, Eye 52, 57, 58, 60, 6 8 , 69, 71,94, 97, 109, 121,124,
175,176,178,183,193,234,244, 246, 248,249, 126,127,128,129,153,154,155, 264, 266, 268,
251,252,272, 286, 288,294,296, 298, 300, 302, 316, 318,371,376, 398
304,306,308, 326, 332,334,336, 338, 339, 340, Eye-to-photograph distance 266

F
341,342, 362, 364, 368, 444, 452
Error matrix 302, 304, 306, 308, 334, 336, 338, 340,
341, 364,342,362
Error of affectation 176 Factorial analysis 390
ERS 3,5,7, 9, 1 0 , 1 1 , 13, 14, 15, 16, 17, 19,20,21, Factorial maps 398
22, 23, 24, 25, 26, 27, 29,30, 31,32, 34, 36, 38, Fading 25
40, 41,42, 44, 48, 50, 52,54, 57, 60. 61,63, 64, Fallow lands 75, 100, 320, 332, 334, 336, 344, 356,
65, 6 6 , 67, 6 8 , 69, 70, 72,74, 75, 76, 78, 79, 80, 374, 396
81,82, 84, 85, 8 6 , 87, 8 8 ,89, 90, 91,94, 97, 98, Farming systems 286
99, 100, 101, 102, 103, 104, 105, 106,107, 108, Faults 404, 405, 410, 442, 444
109,110, 111, 113,114,115,117, 118,120,121, Feldspars 406
122,123,124, 125,126,128,129, 130,131,132, Fences 276
133,134,135, 136,137,138,139,140,141,142, Field 14,16. 17,18,21,23,25,31,32,34,36,44,46,
143,145,146, 148,149,150,151,152,153,154, 48, 52, 58, 70, 72, 75, 77, 79, 80, 81,82, 8 6 , 94,
155,157,158,159,160,161,163,165,166,167, 97, 99, 100, 109, 110, 114, 115, 122, 126, 128,
168,171,172,173,174,176,177,178,179,184, 134,145,150,153,154,161,165,166,167,175,
187,188,189,191,194,198,199, 200, 201,204, 178,186,187, 188,189,191,196,197, 208, 222,
208, 209, 210, 211,214, 215,216, 217, 220, 221, 224, 229, 230, 243, 246, 249, 252, 256, 258, 266,
222, 224, 226, 227, 228, 233, 234, 235, 236, 240, 272, 276, 278, 283, 284, 288,290, 293, 294, 296,
241,242, 243, 244, 246, 248, 249, 250, 252, 255, 298, 304, 314, 316,318,326,328, 332,356, 360,
256, 258, 260, 262, 264, 266, 268, 272, 274, 276, 362, 368, 374, 376, 378, 382,386, 388, 390, 392,
278, 283, 284, 286, 288, 290, 293, 294, 296, 298, 394, 396, 402, 404, 408, 420,444, 446,452,454
300,302,304, 306,308,313,314, 316,318,320, Field of Investigation 187,196, 452
322, 325.326,328, 330,332,334, 336,340,344, Field of study 1 0 0 , 153, 187, 196
346,348,350, 352,354,356,359, 360,362, 364, Field of view 23,32, 34, 36,44,46,48, 52, 58,72, 80,
366,368,372, 374,376,378,380, 384, 386,388, 8 6 , 97, 99, 110, 153, 243, 246, 249, 256, 258,

390,392,394, 396, 398,401,402, 404,405,408, 272, 278, 294, 314, 316, 328, 394, 452
410,412,414, 416,418,420,422, 424,426,432, Film 34
434, 436, 438, 440, 442,444,446, 451,452,454 Filter 52, 61,66, 67, 6 8 , 97, 113,142, 208, 214, 215,
ERS- 1 15, 29, 30, 40, 50, 294, 376, 424, 426, 436, 216, 217, 219, 220, 221,222, 224, 226, 227, 228,
438, 442 233, 238, 255, 256, 286, 290, 318, 328, 360, 362,
ERS-1/ERS-2 442 440, 446, 454
ERS-2 30, 50, 436, 438, 442 Filtered 67, 238, 360, 446
ERTS 36,48,111,115,188,189, 225, 386, 426, 442, Filtering 6 6 , 67, 97, 113, 142, 208, 214, 215, 216,
451 217, 219,220, 221,222,224,227, 228,233, 318,
Euclidean distance 117,144 328, 360, 440
Eutrophication 424, 426 Finite element modelling 244
Evaluation 36, 153, 177, 178, 186, 187, 188, 192, Firn 91
231,306, 308, 328, 336,348,352, 355, 388,396, Fish banks 444
434, 446, 454 Fisheries 426
Evapotranspi ration 18,19,20,86,408,430,432,434, Fixer 65
446 Flexures 442
EWSE 308 Flight height 256, 258, 260, 262, 266
Exactitude 300, 452 Flight lines 262, 264
Exitance 3, 6 Flight plan 256, 260, 264, 274
Exogenous data 142, 298 Flood 36, 98, 249, 293, 294, 359, 360, 364, 366, 368,
Exotech 32, 384 398, 444, 446
 Index 471

Flowers 75, 318, 350 Gaussian 119, 121, 122, 136, 177, 178, 226, 227,
Fluorescence 416,422 296, 298, 334
Focal distance 256, 266, 274 Gaussian parametrisation 177
Foliar index 16, 17 GEMINI 48, 52, 256
Forest 52, 6 8 , 96, 97, 99, 100, 104, 108, 111, 114, General classification 145, 150, 159
125,131,134,136,139,142,145, 149,151,152, Generalisation 191, 197, 206, 219, 220, 252, 284,
155.156.157.160.161.165.166.167.168.169, 290, 293, 336, 352, 358, 359, 368
170,171,172,174,179,180,182,184,195, 200, Generalised superposition 220
201,202, 203, 204, 205, 249, 262, 283, 286, 290, Geochemical 402,404,408,412
294, 302, 304, 313, 314, 316, 318, 320, 322, Geochemistry 401,402, 404, 412
324, 325, 326, 328, 332, 334, 336, 338, 339, 340, Geographic 36, 72, 102, 106, 110, 112, 113, 114,
341,342, 344, 346, 348, 360,362, 364, 366, 368, 116,117,122,123,124,126,133,139,141,142,
394, 396, 398, 420, 432, 434,436, 442,446,452 143,144,145,146,147,154,155, 157,158,161,
Forest cover 104, 157, 286, 290, 318, 434, 452 163,167,168,171,172,175,177,178,182,188,
Forest crops 316 189.191.192.193.194.195.199, 227, 234, 241,
Forest fire 434, 436 244, 246, 248, 252, 256, 262, 264, 276, 278, 284,
Form 3, 9, 12, 17, 19, 22, 23, 24, 25, 26, 27, 29, 30, 286, 288, 294, 300, 302, 304, 313, 314, 322, 326,
31,32, 34, 36, 38, 41,50, 52, 57, 62, 71,72, 75, 328, 330, 332, 334, 336, 342, 344, 346, 348, 350,
76, 78, 82, 8 6 , 87, 8 8 , 89, 90, 94, 96, 97, 99, 356, 358, 362, 364, 370,374, 376, 378, 392, 396,
100,101,102, 103,104,106,107,108,109,110, 398, 402, 404, 405, 414, 422, 426, 452
111,112,113, 114,115,118,120,121,122,123, Geographic approach 141,146,178
124,125,126,127,128,129,130,131,132,133, Geographic database 227, 246, 252
134,136,137,138,139,140,141,142,143,144, Geographic distance 126, 195, 322, 398, 452
145,146,147,148,150,154,155, 156,157,158, Geographic field 294, 326
161,167,169,172,173,175,177, 182,183,184, Geographic information 36, 102, 110, 113, 142, 146,
186,187,188,189,191,192,193,194,196,198, 154,189,191, 194,199, 234, 246, 276, 302, 328,
199, 208, 209, 210, 212, 214, 217, 218, 221,227, 334, 358, 370, 374, 376, 396, 398, 422, 426,452
228, 230, 231,234, 235,236, 237, 240, 241,243, Geographic information system 36, 102, 110, 142,
244, 245, 246, 248, 249,250, 252, 255, 256, 258, 154.189.191.194.199, 234, 246, 276, 328, 334,
262, 264, 266, 268, 270,272,274, 276, 278, 284, 358, 370, 374, 376, 396, 422, 426, 452
286, 288, 290, 293, 294,296,298, 300, 302, 304, Geographic location 167, 342, 374
306, 308, 313, 314, 316,318,320, 322, 324, 325, Geographic mask 142, 362
326, 328, 330, 332, 334, 336,342, 344, 346, 348, Geographic masking 142, 362
350, 352, 355, 356, 358, 359,360, 362, 364, 366, Geographic method 114,122,124,126
368, 370, 372, 374, 376,378,380, 386, 390, 392, Geographic model 145
394, 396, 398, 401,402,404,405, 406,408,410, Geographic restitution 334, 336, 356
412, 414,416, 418, 420,422,424, 426,432,434, Geologic mapping 405,410
436, 438, 440, 442, 444, 446, 451,452, 454 Geology 43, 97, 103, 105, 314, 396, 398, 401,402,
Fracture 402, 404, 405, 408, 410, 442 418, 442, 444, 446
Fractured 404, 408 Geometric correction 36, 110, 142, 234, 246, 294,
Fragmentation 322 302, 332, 352, 451
Frequency 4, 20, 21,22, 24, 25, 28, 30, 44, 50, 123, Geometric deformation 234, 243
130,134,149,150,153,154,156,157,158,160, Geometric distortion 36, 235, 243, 246
163.165.169, Geometric interpolation
176,178,179,180,186,194, 208,235, 237
214, 221,246, 302, 308,355,416, 424,432,436, Geometric precision 300
438, 444, 446 Geometric quality 243
Frost 30, 294, 432, 434, 446 Geometric rectification 36
Fuzzy groups 198, 288, 290, 358 Geometric resolution 23, 34, 36, 50, 129, 256, 260,
Fuzzy logic 154, 193, 302, 308 288, 290, 293, 294, 296, 358, 368, 436, 446
Fuzzy relationship 338, 339, 340, 342 Geometric transformation 113, 234, 235, 237, 240,

G
241,243, 246
Geometrically rectified 298, 430
Geomorphological form 322
Gain 2 2 , 23, 8 6 , 100, 124, 128, 133, 136, 139, 175, Geomorphologicai units 304, 316
184, 255, 268, 278,346,348,350,386,424,432 Geomorphology 98, 103, 108, 266, 284, 286, 320,
Gamut 6 8 , 255 368, 394, 454
472 Processing of Remote Sensing Data

Geophysics 401,402 209, 268, 318, 332, 336, 338, 339, 341,342, 346,
Georeferencing 246 350, 352, 378, 390, 392, 398,418, 424, 426,444,
Geostationary 40, 41,42, 44, 46, 48 451,454
Geostatistics 404 Grey level 4, 58, 60, 61, 63, 65, 94, 104, 109, 124,
Gibbsite 406 125,126,127, 134,139,147,196, 208, 209, 210,
GiS 22, 23, 25, 36, 105, 108, 110, 111, 115, 246, 212, 214, 216, 217, 219, 220, 221,223, 224, 227,
248, 252, 278, 320, 328,336,364, 374, 394, 398, 288, 440, 451
401, 414, 424, 438, 442, 444, 452 Grid 61, 193, 209, 212, 244, 245, 246, 248, 262, 274,
Glass 52, 6 6 , 69, 255, 260, 268 300, 316, 326, 362, 372, 374, 422, 426, 454
Global change 348 Gridding 209, 244, 245
Global precision 304 Ground control points 244, 336
GMS 40, 44, 374, 376 Ground data 298, 304, 340, 352, 357, 358, 395, 399,
GOES 22, 40, 44 412,445
Goethite 406 Ground resolution 23, 44, 410, 416, 422, 438
GOME 50, 438 Ground track 43, 453
Gossan 401,402, 406 Ground-truth 304, 306, 307, 308, 327, 328, 331,337,
GPS 38, 52, 111,256, 260, 262, 452 352, 367, 368, 369, 372, 373, 374, 394
Gradient 28, 216, 221,224, 225, 226, 227, 228, 229, Groundwater table 105
231,418, 420, 432, 434, 444, 454 Group 69,78,99, 107, 114, 115, 116, 117, 118,119,
Grading 116,117 120,133,141, 144,145,146,147, 148,149,150,
Grain 28, 29, 34, 65, 84, 91, 94, 371, 388, 402, 418, 151,152,153, 154,155,156,157, 158,159,160,
454 161,162,163,164,165,166,172,173,174,175,
Grain size 28, 29, 84, 91,388, 418 183,185,188,189, 203, 204, 209, 217,218,219,
Granite 402, 406, 408, 410 232, 249, 284, 286, 289, 290, 300, 303, 306, 308,
Graphic 32, 36, 38, 6 6 , 67, 71,72, 89, 97, 101, 102, 316, 334, 349, 359,360, 390,394, 398,400,436,
103,104,105,106,109,110,112,113,114,116, 444, 451,452, 453, 454, 455
117,122,123,124,126,127,133,139,141,142, Group relationships 217
143,144,145,146,147,154,155, 157,158,161, Gypsum 386, 388, 390, 394, 406, 408
163,167,168,171,172,175,177,178,182,188, Gyroscope 260

H
189,191,192, 193,194,195,199, 227, 228, 234,
235, 241,243, 244, 245,246,248, 250, 252, 255,
256, 260,262, 264, 266, 268, 270, 272, 274, 276,
278, 283, 284, 286, 288, 294,296, 300, 302, 304, Habitat 101, 102, 103, 108, 276, 320, 356, 359, 366
313, 314, 318, 320, 322,326,328, 330, 332, 334, Haematite 401,406
336, 342, 344, 346, 348, 350,352, 356, 358,362, Hardwood 13, 96, 132, 200
364, 370, 374, 376, 378, 382,392, 394, 396, 398, Harrowing 27
402, 404, 405, 414, 420, 422, 426, 436,440, 442, HCMM 410,434
444, 452, 454 Hedges 99
Grass 13, 16, 72, 74, 75, 77, 80, 81, 100, 101, 108, Herbaceous formations 344, 346,348,350, 352,356,
125.135.141, 142,157,160,165,166,168,170, 358
171,174,179,180,181,183,184,187,194,195, Heterogeneity 153,161,172,174,178,194,195,196,
200, 201,202, 203, 204, 205,284, 290, 294, 296, 197,198,199, 288. 290, 302,316,318,336,366
302.306.316, 320,324,325,328, 332,334, 336, High-pass filters 216
344, 346, 348, 350, 352,354, 355, 356, 358, 360, Hillock 320, 324
362, 364, 366, 368, 396, 418 Histogram 121, 122, 123, 127, 130, 131, 132, 133,
Grassland 13, 16, 72, 74, 75, 77, 81, 100, 101, 108, 134,135,136,137,139,140,142,144,146,156,
125.135.141, 142,157,160,165, 166,168,170, 157,158,159,163,165,166,167, 168,169,170,
171,174,179, 180,181,183,184, 187,194,195, 172,173,175,176,177,179,181,182,194,195,
200, 201,202, 203, 204,284,290, 294, 296, 302, 198, 201,202, 203, 204,205,206, 296,304, 326,
306.316, 320, 324, 325,332, 334, 336, 344, 346, 362, 392
348, 350, 352, 354, 355,356, 358, 360, 362, 364, Homomorphic filtering 219, 220
366, 368, 396 Homothetic ratio 258
Green 9,12,42,57, 58,59, 60, 61,62, 63, 6 6 , 67, 6 8 , Homothety 238, 241,243
69, 70, 74, 75, 78, 79, 81, 8 6 , 122, 123, 125, Horizons 82, 83, 105, 378, 394, 454
126,127,128,129,130,134,136,139,153,159, Hot spot 8 6 , 87
160,166,174,179,180,181,183,184,200,201, Hough transformation 228, 230
 Index 473
HRG 52 131,134,135,136,137,138,139,142,156,157,
HRV 32. 36, 43, 50, 52, 90, 189, 316, 418 158,159,171,172,173,178,179, 256, 258,260,
Hue 63, 64, 70, 86, 87, 104,121,127, 129, 139, 187, 278,288, 294, 296, 302,304,318, 334, 348, 350,
318, 378, 388, 394, 406, 424, 432, 452 356, 362, 364, 368, 378,384,386, 388, 390,392,
HViR 51 398,405,406, 410, 418,420,422, 424,430,432,
Hydraulic potential 388 434, 436, 438, 444, 446, 451
Hydrocarbon 30, 88, 414, 416, 420, 422, 424, 444 Infrared colour composite 129, 334, 362
Hydrolandscapes 314 Initialisation 189
Hydrology 102, 103, 108, 314, 394, 396, 444, 454 INSAT 40,44
Hydrothermal 401,402, 405, 406, 408 Instantaneous field of view 34, 36, 452
Hyperstereoscopic 276 Interception efficiency 85
Interferometry 442, 444, 446
Intergroup distance 118,119,120,144,147,300,451
Internal factors 83,451
Ice 9, 12, 23, 27, 30, 31, 36, 38, 44, 48, 50, 52. 61, Internal structure 76, 77, 78, 79, 80
63, 69, 70, 72, 73, 78, 79, 80, 82, 84. 85, 8 6 , 87, International Commission on Illumination 58
90, 91, 100, 104, 108, 114, 116, 118, 119, 120, interocular distance 266, 268, 270
125,126,127,128,131,132,133,134,135,138, Interpolation 15, 234, 235, 236, 237, 238, 239, 244,
140,145,147,148,149,150,155,161,162,163, 252, 286, 374
169,170,172,174,175,177,178,181,182,183, Interpolation methods 234, 235, 236, 286
186,187,188.191,197,228,234, 246,249,251, Interpretation 30, 36, 46, 55, 57, 69, 70, 71,72, 84,
252,255,260, 266, 268,270,278, 288, 290,293, 8 6 , 91, 93. 94, 97, 98, 101, 102, 103, 104, 105,

294,296,298, 302,304,306,308, 314,322,326, 106,107,108,109,110, 111, 114,115,116,122,

328,330, 332, 334,336,344,348, 352, 355,356, 126,127,128,129,131,132,134,135,136,139,
368,372,373,374,376,378,380,390,401,402, 142,143,146,148,149,150,153,154,155,156,
406,408,410,414,422,426,432,436,442,444, 158,161,163,165,166,167,169,174,175,178,
446, 451 191,196,201,222, 234,251,252, 255,256,266,
ICI 58,59,61,62,64 268,274,276, 278,284,292,294, 296,300,306,
Identification 75, 84, 8 6 , 94,114,149,139,161,167, 308,313,314, 316, 318,320,322, 325,326,328,
175,179,182,188.189,197,199, 231,250,251, 330.332.342, 348,350,356,360, 362,366,368,
252,286,294, 296,298,302,304, 306,318,328, 372,373,374, 378, 392,394,396, 398,401,402,
332,336,340, 342,344,346,348,350,355,362, 405, 408, 412, 432, 436, 410, 446
364, 368, 392, 401,406, 442, 444, 446 Interpretation model 72,108,114,146,161,178,191,
Identification error 252 348
IFOV 34,452 Interpretation of classification 147,148,154
Image processing 68,110,113,114, 115,116,117, Interpretation of dendrogram 149
118,121,123,127,132,134,136,144,147,154, Interpreter 104, 108, 109, 115, 125, 142, 144, 186,
186,187,189,192,193,194,209, 213,222,244, 313,316,322, 325,326,328,330,332,373,408,
246, 328, 382, 454 412
Image zone 153, 154,155, 178, 181, 194, 221,306 Intertidal 332,414,418,422,424
Imaging radars 23, 436 Intertropical zones 345
Impulse response 2 1 2 , 213, 226 Intragroup distance 118,119,120,144,147,286,300,
Incident angle 316, 438 454
Inclination 23, 40, 41,48, 80, 8 6 , 258, 264, 272, 294 Intrapixel 119,144
Independent samples 304, 306 Intraplot 172,286
Index 14, 15, 16, 17, 80, 85, 8 6 , 87, 1 2 1 , 134, 135, Inverse transformation 235, 236
138,141,188,199,217,290,304,336,352,354, Inversions 440
355, 374, 380, 392, 418, 420, 430, 434 IR/R 87, 134, 135, 352, 354
Indicator 106,154,183,223,244,300,328,348,356, IRC 34, 36, 38, 40, 41. 42, 44, 46, 63, 98, 99, 100,
374, 390, 405, 406 108,128,133,153,165,170,178,192,218,237,
Indicator species 348 245,255,256, 258,260,262,264,270,272,274,
INFEO 308 300,304,316,318, 350,360,362.404,410,416,
Infrared 4,5,9,10,11 ,1 2 ,1 3 ,1 4 ,1 6 ,1 8 , 20,21,31, 438, 444, 451
42, 43, 44, 46, 48, 50, 52, 65, 6 6 , 67, 6 8 , 69, 70, Iron 3, 11,28. 30, 40, 44, 50, 52, 6 8 , 81, 84, 87, 8 8 ,
71,73, 76, 78, 79, 80, 81,82, 83, 84, 85, 8 6 , 87, 97, 99, 100, 101, 104, 105, 106, 107, 187, 296,
88,90,91,97,102,122,123,126,128,129,130, 314.320.322.326.328.330.342, 346, 348,352,
474 Processing of Remote Sensing Data

356,359,366, 368,374,378,384, 386,388,390, Landscape 36, 43, 96, 97, 99, 107, 108, 1 1 1 , 133,
392,394,396, 406,410,414,418,422,424,426, 153,161,174,189,191,197, 266, 276, 284, 286!
444 304, 313, 314, 316, 318, 320,322, 326, 328, 346!
Iron pans 386, 408 360, 394, 396, 398,414,420,422, 442,451,452,
IROS-AVHRR 405 454
Irrigated 332, 336, 338, 339, 340, 342, 430 Landscape units 36, 108, 174, 286, 304, 314, 320,
Irrigation 100, 101,316 322, 326, 328, 394, 420, 422, 454
Iteration 104,116,117,162,163,164,165,166,182, Landslides 446
184, 186, 199, 200, 218, 230, 231,314, 326 Laplacian 142,226,286,362
IZBC 366 Laser 32, 44, 50, 256, 416, 422, 452
Lateral swing 44
Lead 30,31,36,52, 58,79, 80, 84, 86,104,111,113,
116, 117,128, 154,161,165,167, 172,176,182,
JERS 44, 46, 50, 436, 438, 442 193,197,209, 234,235,243,248, 252,286,290,
JERS-1 436 294,296,298, 304,316,344,350, 371,374,382,
Joint hierarchic model 286 396, 401,402, 404, 406, 410, 416, 432, 440
Juxtaposition 194, 336, 452 Leaves 16, 17, 26, 72, 73, 74, 75, 76, 77, 78, 79. 80,

K
8 6 , 157, 286,316, 318, 394

Legal time 9
Legend 70, 83, 97, 109, 134, 145, 146, 160, 165,
Kband 252 166,189,191,320, 330,352,366,368,394,396
K-means 162 Lens stereoscope 268
Kaolinite 406 Leucogranite 402,406,410
Kappa 292, 304, 306, 308, 334, 339, 340, 364 Level of analysis 284, 322, 326, 452
Kepler 38,40 Level of exploitation 345
Khl^ test 307 Level of observation 286
KS estimator 306 Level of precision 297
Krigging 431 Lidar 32, 452
Lignin 79
Limestone 397
Lineage 139
L band 5, 10, 13, 15, 18, 34, 36, 38, 42, 43, 44, 46, Lineaments 404, 405, 406
48, 50, 52, 65, 6 6 , 67, 74, 75, 79, 81,83, 89, 90, Linear cluster 170,378,390,392
121,124,127,128,129,130,131,136,138,139, Linear features 125,138,178
141,150,219, 284, 288,290,294, 296,298,302, Linear function 124
378,384,392,406,408,416,418,424,436,438, Linear position 301
440, 446, 451,452, 454 Lithologic 396,410,412,442
Laboratory 25, 72, 73, 74, 75, 76, 79, 82, 191, 284, Litter 344
296, 378, 380, 386, 388, 444 Littoral environments 28,418,426
Lagoons 332, 364, 422, 444 Local contrasts 322
Lambert’s law 7, 27 Local deformation model 250
Lamprophyre 404,410 Local heterogeneity 288
Land Cover 17, 96, 101, 104, 105, 108, 109, 129, Local histogram 194, 195, 198
139,141,142,145,150,155,198, 274,286,288, Local models 244
290,294, 296, 302, 304,308,316, 320, 324,326, Location 122,126,167,178,197,226,227, 249,250,
328,330,332,334,339,341,342, 348,356,358, 268,292,304, 308, 316,328,342, 373,374,401,
360, 362, 364, 366,368, 394,396, 398,414,424 405, 422, 424, 432
Land use 100, 154, 174, 274, 342, 414, 416, 420 Location error 251,252
Land zone 414 Logical generalisation 284
LANDSAT 14, 32, 34, 36, 41, 44, 46, 48, 89, 128, Logical mask 142, 143
129,174,189, 252,286,294,306, 316,328,330, Logicaimasking 142
356,364,366, 368,371,390,404,408,410,426, Longitudinal resolution 23, 438, 440, 444
451,454 Look Up Table 69, 123, 126, 134, 452
LANDSAT MSS 34, 46, 286, 294, 408, 410, 426 Low oblique photographs 258
LANDSAT TM 46,128,129,174, 252, 286, 294, 306, Low-pass filters 216, 221,222
316, 328, 330, 356, 364, 368, 426 LOWTRAN 14, 122
 Index 475

LUT 3 , 1 2 , 15, 18, 19, 21,23, 25, 30, 34, 36, 38, 41, Maximum altitude 255, 262
42, 43, 44, 46, 48, 50, 52, 65, 70, 8 8 , 90, 91,94, Maximum resolution 286
97, 99, 106, 109, 111, 114, 118, 119, 122, 124, Maximum-likelihood 116, 117, 165, 176, 177, 180,
126,128,129,131,136,141,150,163,188,189, 184, 186, 286, 296, 306, 334
193, 212, 213, 214, 216, 220, 224, 226, 227, 237, Maximum-likelihood classification 116,117,165,176,
238, 239, 248, 249, 252,256, 260, 262, 276, 278, 177, 180, 184, 186
284, 286, 288, 290, 293, 294, 296, 302, 304, 306, Mean filter 221
308, 314, 316, 318, 326,332, 344, 348, 352, 358, Mean of errors 300
366, 368, 372, 376, 390, 394,401,402,404, 406, Mean pixel 118,144
408.410, 412, 414,416,418,420, 422, 424,426, Measure of similarity 288
434, 436, 438, 440, 442, 444, 446, 452, 454 Measurements 6 , 12, 14, 15, 17, 18, 19, 23, 25, 27,
29, 30, 32, 48, 58, 72, 73, 74, 75, 78, 79, 80, 81,
M 82, 83, 90, 97.114,122,126,128,150,161,178,
188,191,246, 255, 256,288,290, 300, 304, 348,
Magenta 60, 61,66, 67, 70, 129, 134, 398 350, 352, 354, 355, 356, 374, 378, 386, 388, 390,
Mahalanobis 117 398,408,410, 412, 414,420,426, 432,436,451,
Maize 13, 16, 74, 81,82, 85, 171,430, 432, 448 452, 454
Major thematic zones 145,160 Median 119. 122, 216, 217, 221,227, 290
Mangrove 388, 414, 420, 422, 424 Median filter 221
Manhattan distance 118, 126, 196, 198 MERIS 50, 63, 90
Manual masks 142 Mesophyll 78, 79
Map 36, 48, 50, 97, 103, 104, 105, 106, 107, 109, Metalliferous provinces 402
110, 111, 114, 131,137,142,145, 150,160,178, Metallogeny 401,402,410
187,188,189,191,192,193,194, 198,199, 200, Metallotect 402, 404, 405, 412
206, 227, 231,234, 242,243,244, 246, 248, 249, Metamerism 63
252, 256, 258, 262, 264, 266, 268, 270, 272, 274, Metamorphic 402
276, 278, 283, 284, 286, 288, 290, 293. 294, 296, Meteorological 14,16,30,46, 73,352,258, 260,371,
298, 302, 304, 306, 308,314,316, 318, 320, 322, 372, 374, 376, 378, 380,405,410,430,432,436
326, 328, 330, 332, 334,336, 338, 339, 340, 341, METEOSAT 14, 40, 44, 46, 48, 410, 432, 434, 436
342, 348, 350, 352, 356, 359,360, 364, 366, 368, Methane 10
373, 374, 376, 378, 388,390,392, 394,396, 398, Method 14,15,18, 21,30, 61,73, 8 8 , 89, 90, 97,102,
405.408.410, 412,414,420,422,424,430,432, 107,111,113, 114,115,116,117,119,120,122,
4 3 4 , 442, 444, 446, 451,452, 454 123,124,125,126,127,133,134,137,140,141,
Map boundaries 302 142,144,145,146,150,156,159,161,162,163,
Map unit 103,104,107,109,110,114,189,192,193, 165,166,167,170,175,176,177,184,186,187,
194,199, 200, 206, 278, 284, 286, 288,326, 328, 188,189,193,194,195,198,199, 200, 205, 206,
330, 336, 394, 451,452, 454 208, 215,216, 219, 221,222,224, 226, 227, 228,
Map zone 150, 189, 320, 322, 326, 328, 342, 360, 229, 230, 231,233, 234, 235, 236, 237, 238, 242,
392, 394, 452 244, 245, 246, 249, 250, 252, 255, 268, 276, 278,
Mapping 50,105,107, 1 1 1 , 193, 258, 278, 284, 286, 283, 284, 286, 288, 290, 292, 293, 296, 298, 302,
288, 294, 296, 306, 316, 328, 330, 332, 334, 336, 304, 306, 308, 313, 320,322, 325, 326, 328, 330,
340, 342. 350, 352, 364,366,368, 378, 394, 398, 332, 334, 336, 348, 350, 356, 358, 360, 364, 366,
405,408, 410, 412,420,422,424, 430,432,434, 372, 373, 374, 376, 394,398,402, 404, 405,412,
442, 444, 446, 452, 454 418, 420, 422, 424,426, 430, 432, 436,442,454
Markovian methods 229 Method of synthesis 97
Mask 27, 29, 74, 75, 79, 1 2 1 , 125, 126, 127, 136, Method of validation 114, 189
141,142,143,145,148,154,158,159,160,161, Metric 7, 8 , 13, 19, 23, 26, 30, 34, 36, 38, 43, 44, 48,
165,175,179,182,184,188,192, 214, 218, 219, 50, 52, 54, 58, 75, 78, 89, 90, 110, 111, 113,
238, 239, 242, 243, 286,249, 296, 298, 348, 356, 114,116,117, 118,119,120,122, 123,125,126,
360, 362, 364, 366, 368, 404, 408, 410, 422 129,133, 136, 141, 142,144,145, 146,150,154,
Masking 75, 1 2 1 , 125, 126, 127, 141, 142, 143, 145, 155,159,161, 162,163,164,166, 167,170,171,
159,160,161,165,175,184,188, 242, 243, 286, 172,175,177, 178, 179,188,191,192, 209, 212,
296, 298, 348, 362 215, 216, 217, 218, 228, 229, 230, 231,234, 235,
Mathematical filtering 318 236, 237, 240, 241,243,246, 248, 255, 256, 258,
Mathematical morphological operators 286 260, 268, 274, 278, 283, 284, 288, 290, 293, 294,
Maturation 78, 80 296, 298, 300, 302, 308,316, 318, 326, 332, 336,
476 Processing of Remote Sensing Data

348, 352, 358, 360, 366, 368, 380, 390, 405,408, Modelling 20, 71, 125, 187, 191,234, 235, 240, 241,
414, 416, 418, 420, 422,430,436, 440, 442,444, 243, 244, 245, 322, 424, 426, 444
446, 451,454 Modes of exploitation 348, 350, 354, 360, 368
Metric camera 52, 255, 256, 260, 278 MODIS 48, 50
Micas 406,410 MODTRAN 14,15
Micrometric screw gauge 274 Modulation 24, 25
Microphytobenthos 418 Moisture 9, 12, 19, 28, 29, 30, 50, 82, 83, 84, 350,
Microtalweg 172 378, 386, 388, 390, 392, 408,418, 434,436,446
Microwave 1 0 , 12, 20, 21, 22, 25, 30, 31, 50, 296, Monochromatic 25, 59, 63, 126
316.376, 378, 386, 390,398,405,414,416,418. Monocotyledons 80
420, 422, 430, 436, 438, 444, 446, 454 Montane regions 304, 432, 440, 444
Middle infrared 12,43,48,73, 79,80, 81,83, 88,129, Montmorillonite 406, 408
137,139, 294, 350, 356, 362, 368, 386, 388, 390 Morphology 30,43, 98,103,104,108,113, 217, 218,
Mie scattering 1 1 219, 220, 266, 274, 284, 286, 320, 324, 326,
Mineral 29,72, 8 8 , 316, 318,332, 336,338,341,342, 368, 394, 396, 398, 426, 442, 446, 454
364, 390, 401,402, 404,405,406, 408,410,412, MOS-1 36
414, 418,420, 442, 444 Mosaic 204, 246, 252, 272, 274, 286, 294, 328, 332,
Mineral deficiencies 75 336, 360, 364,418, 442
Mineralogical 29 Mosaicking 294
Minimum clearance 262 Mountain pastures 70,318
Minimum resolution 284, 332 Moving window 197, 198, 199, 214, 215, 290, 336
Minimum sorting distance 327 MSS 32, 34, 46, 48, 83, 87, 286, 290, 294, 408, 410,
Mining geology 401 426
Mirror stereoscope 268 MSU-SK 50
Mixel 109, 116, 117, 118, 154, 157, 171, 172, 176, Mud 29,414,418,420
193, 204, 286, 288, 290, 316, 318, 368,418,452 Multifactorial analysis 390
Mixels 116, 154, 157, 171, 176, 193, 204, 286, 288, Multiquadric functions 245, 246
290,316, 318,368,418 Multiscale 252, 404
Mixels of combination 154 Multispectral 5, 34, 42, 44, 48, 50, 52,109,121,127,
Mode 12, 14, 15, 16, 18, 19, 20, 26, 30, 31,38, 43, 145,170,171,175,193,197,219,256,284,332,
44, 50, 54, 71, 72, 73, 87, 8 8 , 90, 91, 98, 99, 352, 376, 401,406, 416, 418, 422, 444
100.101.104.106.107.108.110, 111, 114,116, Multispectral bands 5,127
119.122.125.128.131.132.142.145.146.150, Multispectral cameras 256
161,165,167,168,169,170,172,175,178,187, Multispectral segmentation 170,171, 175
188,191,194,197,199, 210,221,223, 224,226, Multithreshold 125
228, 229,230, 234, 235,236,237, 240,241,243, Munsell 62, 63, 64, 71,83, 149, 378, 380, 382, 384,
244, 245,246, 248, 249,250,251,252,256, 264, 386, 388, 451,452, 454
266, 268, 274, 276, 278, 286, 288, 290, 294, 304,
313, 314, 316, 318, 320, 322,326, 328, 332, 342,
344, 348, 350, 354, 360,362,364, 366, 368, 372,
N
373.374.376, 380,390,392,394, 396, 398,401, Nadir 36, 52, 91,258, 410, 438, 452
402, 404, 408, 412,414,422,424, 426,430,432, National representation 372, 374
434, 436, 438, 442, 444, 446, 452 Nature of classes 302
Model 12, 14, 15, 16, 18, 19, 20, 26, 30, 31,43, 50, NDVI 86,87, 134, 135, 182, 183,290,354,355,418,
54, 71, 72, 73, 87, 8 8 , 90, 91, 100, 104, 107, 434
108.110, Near and middle infrared 350
111, 114,122,125,128,131,132,142,
145.146.150, Near infrared170,172,175,
161,165,167,168, 42, 46, 48, 50, 70, 73, 76, 79, 80, 81,
178,187,188, 191,194,197,199, 221,223, 224, 82, 8 6 , 87, 90, 91,102, 128, 129, 130,131,134,
226, 228, 229, 230, 234,235,236, 237, 240, 241, 137,138,139,142,171,172,173, 302,348,350,
243, 244, 245, 246,248,249,250, 251,252,264, 378, 384, 386, 388, 390, 392, 406, 436, 444,
266, 268, 274, 276, 286,288, 290, 313, 314, 316, 446
318, 320, 322, 326, 328, 332, 342, 348, 364, 366, Nearest neighbour interpolation 236
372, 373, 374, 376, 380, 390, 392, 394, 398,401, Negative 59, 65, 70, 87, 110, 125, 128, 131, 134,
402,404,408, 412, 422,424,426, 430, 432,434, 276, 374, 390, 432, 434
442, 444, 446, 452 Neighbourhood 113, 127, 141, 155, 172, 193, 194,
Model heterogeneity 199 195,196,197,198,199,202,205, 206,212,214,
 Index 477
215, 216, 218, 219, 221,226, 230, 237, 286, 296, Organisation 36, 70, 96, 97,128,133,153,158,159,
314, 316, 342, 360, 398, 452, 454 160,161,172,174,175,189,194,198,199, 205,
Neighbourhood curve 196,197 246, 249, 255, 262, 283, 284, 286, 290, 294, 300,
Neighbourhood function 196 304, 308, 313, 314, 316, 320,322, 326, 328, 332,
Neural networks 154, 231,233, 288, 290 336, 364, 374, 376, 378, 394, 396, 426, 446,452,
New band 133, 136, 139, 140, 141,362 454
Night 30, 44, 408, 410, 422, 432, 434, 444, 446 Orientation 6 , 26, 79, 80, 130, 223, 231, 270, 276,
NIMBUS 41,44,90,426 290,316,410,424,434
NOAA 12, 13, 15, 16, 40, 44, 46, 48, 8 6 , 290, 374, Orthophotography 246
376, 426, 430, 432, 436 Orthophotos 276
NOAA- 6 44 Oscillating mirror 34, 46, 454
NOAA-7 46, 432 Overall accuracy 306, 308, 340
NOAA-9 16 Oversampling 286
NOAA-AVHRR 46, 8 6 , 374, 376, 430, 432 Oxides 386,406,410
Nodal axis 40 Oxygen 10,11,40,436
Noise 90, 137, 138, 139, 140, 193, 208, 216, 2 2 0 , Ozone 1 0 , 13, 44, 50, 438
221,222, 223, 224, 226, 227, 228, 233, 296, 440
Noise reduction 208, 220, 221, 2 2 2
Nominal value 300,451,452
Non-controlledmosaic 273 panchromatic 5, 36, 42, 44, 46, 50, 52, 65, 6 6 , 109,
Normalised vegetation index 8 6 , 418, 420, 434 121,127,135, 138,139,198, 219, 258, 294, 360,
Nuclei 115, 116, 133, 144, 159, 162, 163, 164, 165, 416
177,178,179, 180,181,182,183, 184,185,186, Panoramic camera 256, 278
188,198, 199, 200, 201,202, 203, 204, 205, 206, PAR 85
298, 327, 360 Parallax 248, 249, 264, 274, 276
NZEFFI 366 Parallelepiped 99, 158, 161, 167,168, 173, 175,178,
188,193,366
o Parallelepiped classification 158,161,167,168,173,
175, 178
OASIS 113, 115, 116, 127, 161, 194, 197, 198, 199, Parametric methods 228, 234, 235
200, 201,296, 328, 336, 360, 454 Parametric space 229, 230
Objective 18, 24, 34, 46, 48, 52, 61, 105, 114, 116, Parasites 75
124,133,144,145,146,148,150, 154,159,160, Passive 20, 21,25, 32, 50, 416, 422, 444
161,162,163, 165,170,174,175, 177,184,186, Passive remote sensing 21,422
187,188,189,191,192,194,198, 234, 255, 256, Passive systems 25
258, 262, 270, 272, 286, 290, 293, 296, 298, 302, PASTEC 44
308, 314, 320, 328, 330,332,348, 352, 356, 360, PASTEL 44
362, 368, 374, 396, 398, 401,414, 424, 436, 442 Pastoral values 355
Oblique photos 258 Patchiness 155
Oceanography 436, 444 Pattern 25, 26, 96, 98, 99, 100, 101, 189, 193, 194,
Octet 210,211,227 195,196,197, 198, 199, 201,202, 283, 290, 313,
Omissions 306, 332, 341,342, 368 314, 316, 322, 326, 328, 332, 336, 338, 341,342,
Open field 99 364, 394, 405, 424, 454
Opening 218, 219, 227, 228, 286, 334, 405 Pb-Zn 402, 404
OPS 51,382,438 PCA see principal component analysis
Optical axis 256, 258, 270, 274 Peat bogs 332, 364, 366
Orbit 16, 38, 40, 41,42, 43, 44, 46, 52, 48, 50, 234, Penetration 31,89,296,418
243, 404, 426, 452 Percentage of dry matter 344, 346
Orbital drift 16 Perception 57, 69, 71, 94, 110, 124, 194, 227, 264,
Orbital plane 40, 41,42 266, 283, 284, 313, 328, 386, 394, 405, 452
Orchards 13, 96, 108, 278, 320, 324, 326, 396, 432 Perception level 194,283,284
Ordinal classification 290 Performance table 177, 182, 183, 184, 186, 188
Ordinal units 306 Perimeter 42, 114, 154, 189, 193, 199, 328, 451
Organic matter 72, 81, 82, 84, 8 6 , 8 8 , 90, 378, 382, Periodicity 40, 48, 188, 414
384, 386, 388, 390, 418 Permafrost 446
478 Processing of Remote Sensing Data

Permanent grassland 72, 74, 75, 77, 81, 141, 142, Plagiophyll 80, 353
174, 290, 294, 302, 306, 344,346, 348, 350, 352, Planar positioning 302
354, 355, 356, 358, 360, 362, 364, 366, 368, 396 Planck 3,4, 13, 20
Permittivity 29 Planophyll 80, 85
Perspective 178, 241, 242, 262, 264, 266, 268, 276, Plant communities 318, 348,352, 362,366,368,420
313, 314, 318, 376, 424, 426 Platform 23, 32, 34, 36, 38, 41,234, 235, 243, 255,
Petroleum slicks 444 264, 266, 270, 272, 293,390,416, 418,436,438,
Phacelia tanacetifolia 75 440, 442, 444, 454
Phase 21, 103, 104, 107, 114, 119, 123, 126, 133, Plots 99, 100, 101, 104, 108, 109, 124, 125, 155,
134,144.187,188, 221,231,298, 300, 302, 325, 156,157,158,160,165,166,168,169,172,173,
332, 382, 404, 418, 438, 442 174,184,195, 286,314,316,318, 320,322,326,
Phase differences 442 328, 332,334, 336,356,360,362, 366,368,374,
Phenological 72, 80, 81, 294, 318, 348, 350, 356, 382
366, 368,414, 420 Ploughing 28, 383, 393
Phosphates 410 POAM 44
Photo-acquisition 255 Pointing error 248, 251
Photo-interpretation 97, 255, 356 Points of vision 314
Photo-interpreter 330, 332 Polarisation 21,22, 25, 30, 50, 268, 438, 444, 446
Photogrammetric 43, 255 Pollutant 30,88, 91,414,416
Photogrammetry 255, 268, 276, 278, 422, 442 Pollution 8 8 , 90, 348, 414, 418, 420, 422, 424, 444
Photography 5, 52, 65, 6 6 , 6 8 , 246, 255, 258, 260, Polygonal function 125
262, 276, 278, 284, 318, 340, 356, 416 Polynomial methods 237, 244, 245, 252
Physical characteristics 320 Polynomial model 235, 240, 243, 246, 249, 252
Physicochemical characteristics 294 Polynomial modelling 240, 243
Physiognomy 306, 348 Poplar plantations 317
Physiological state 72, 79, 80, 294, 336, 348, 366, Porosity 30, 84, 378, 388
368 Post-supervision 191
Phytomass 80, 81, 84, 85, 8 6 , 286, 344, 346, 348, Postulate 118
350, 352, 354, 355, 356 Pre-processing 95, 114
Phytoplankton 88,418 Pre-supervision 191
Pigments 57, 58, 61,63, 71,73, 74, 76, 8 8 , 90, 127, Precession 41,42
350, 418 Precise thematic study 145,161
Piled-up pixels 155 Precision 15, 72, 97, 109, 173, 193, 210, 211, 214,
Piles 155 245, 246, 248, 251,270, 284,290, 292, 296, 300,
Pilosity 74 302,304,306,318,326,328,336,348,374,396,
Pitch 234, 260, 272, 402, 410 422, 452, 454
Pitchblende 402,410 Precision estimator 306
Pixel 17,18, 32, 34, 36, 38, 48, 52, 6 8 , 69, 70, 8 6 , 90, Precision level 72, 211, 302, 304, 396
94, 109, 110, 113, 114, 116, 117, 118,119, 120, Precision of attributes 302
121,123,124,125,126,127,128,130,133,136, Precision of linear position 300, 302
137,138,139,141,142,143,144,145,146,147, Precision of position 300
148,149,150,151,153,154,155,156,157,158, Precision of shape 302
159,160,161,162,163,164,165,166,167,169, Preliminary processing 121,143
170,171,172, 174,175,176,177, 178,179,180, Presentation scales 189
181,182,183,184,186,187,188,189,193,194, Primary colours 58, 59, 60, 6 8
195,196,198,199, 200, 201,202, 203, 204, 205, Principal component analysis 121,137,138,139,140,
206, 209, 210, 211,212, 214,215, 216, 221,224, 141, 186, 188, 318, 326, 352, 360, 372, 392,
225, 227, 235, 236, 237, 238,248, 249, 251,252, 408
283, 284, 286, 288, 290,292,294, 296, 298, 302, Principal components 127, 136, 138, 141, 316, 320,
304, 306, 308, 316, 318, 326, 328, 332, 334, 336, 322
341,350, 352, 356, 360, 362, 364, 366,368, 398, Printing 57, 61,67, 71, 126, 256, 276, 318
418, 430, 442, 451,452, 454 Probability 43, 103, 116, 117, 119, 159, 165, 167,
Pixel neighbourhood 193, 360 176,177,181,182,183,184,186, 226, 232,288,
Pixel-by-pixel classification 286, 302 334, 398, 452
Placing in conformity 246 Probability images 182
 Index 479
Processing 91,93 Radarclinometry 442
Processing 36, 43, 52, 54, 6 8 , 69, 70, 71,89, 90, 94, Radargrammetry 442
96,102,109, 110, 111, 113, 114, 115, 116, 117, RADARSAT 45, 50, 51,436, 437, 438, 442, 446, 463
118,121,123,124,126,127,132,133,134,136, Radial distortion 274, 275
141,142,143,144,145,147,148,154,174,175, Radiance 6 ,12, 13,14, 17,18, 20, 35, 36, 37, 84, 85,
186, 187, 188, 189, 191, 192, 193, 194, 199, 90, 114, 128, 136, 288, 289, 290, 294, 295, 303,
206,208, 209, 210, 212, 213, 214, 215, 216, 333, 381,452, 455, 456
217, 220, 221,222, 231,233, 234, 235, 244, 246, Radiation 3, 4, 6 , 7, 9, 10,11, 12,13, 14,17, 20, 21,
248,252,256, 276,283,286,293, 296,298,313, 25, 26, 27, 31,32, 34, 42, 44, 59, 72, 73, 76, 78,
316,328,334, 336,352,358,360, 368,382,394, 80, 85, 8 6 , 87, 8 8 , 89, 91, 314, 352, 386, 388,
398,412,416,422,424,426,436,438,446,451, 408, 422, 426, 432, 434, 442, 452
454 Radiative surface temperature 17
Processing system 69,187, 210, 212, 213, 246 Radio-sounding 12
Production 73,85,106,256, 268,346,352,371,376, Radiometer 5,16,17, 32, 34,46, 50,75, 82, 83,304,
398, 414, 418, 442, 452 352, 354, 355, 356, 384, 416, 418, 420
Projection 6,36, 40, 63, 64,106,110,132,140,146, Radiometric correction 141, 142, 246, 294, 366
234,241,242, 243,246,249,256, 258, 274,276, Radiometric interpolation 236
278, 316, 324, 332, 392, 414, 452 Radiometric mask 126,141,348
Pseudo-median 221 Radiometric masking 126,141,348
Pseudo-stereoscopy 270 Radiometric method 114,122,126
Pseudo-true colour 128 Radiometric model 125,145,150,167,170,172,175,
Punctuate position 300 178, 188
Pushbroom 34, 36, 454 Railroads 100
Pushbroom scanner 34, 36 Railway tracks 125, 316
Pushbroom sensor 454 Rain 11,15,23, 28,29, 30, 34, 57, 65,84, 91,94, 97,
Pyrophyllite 406 98,100,101, 102, 108, 110, 111, 115, 133,134,
141,142,154,184,187,230,231,243,264,268,
Q 272,274,294, 296,298,304,313, 314, 316,318,
322,330,348, 350, 352,354,355, 358, 359,366,
Qualitative evaluation 308, 336 368, 371,372, 374, 376,378,388, 394,396,402,
Qualitative nature 302 404, 412,416, 418,422,424,426,430,432,434,
Quality 6 8 , 90, 114, 117, 120, 126, 159, 165, 175, 436, 438, 440, 442, 452, 454
182,183,184,186,188,193,198,199,201,202, Ramp 224
203, 204, 205, 243, 244, 249, 250, 255, 278, Random errors 176, 302
281,293,294,300,302,308,336,346,348,355, Random sampling 304, 306
356, 359,360, 368,372,373,376,412,424,438, Rangelands 346,348,350, 356
440 Ranked qualitative nature 302
Quality of classification 117,159,182,184,186,201, Raster 110,209,248,454
203 Raster data 248
Quality of data 302 Rayleigh 11,20
Quality of interpretation 300 Real maximum interval 123
Quality of results 120,186, 205, 294, 368, 372, 373 Receiver 6,7,13,14,17, 22, 23, 25, 32, 34, 97,124,
Quantity of information 140, 284 255, 436, 451,452
Quarries 88,138, 334, 336, 388 Reconnaissance 105, 106,145, 160, 166, 258, 398
Quarter 109, 161, 189,454 Rectification 36, 244, 245, 246, 248, 249, 250, 251.
Quarter rule 189, 454 252, 274, 276, 302, 414, 422
Quartz 6 6 , 260, 390, 394, 402, 408 Rectification model 248, 249, 252
Quick-look 6 8 Rectified photographs 274

R
Red 4, 5, 6 ,7, 9,10,11,12,13,14,15,16.17,18,19,
20, 21,22. 23, 24, 25, 26, 28, 29, 30. 31,32, 34,
36, 38, 41,42, 43, 44, 46, 48, 50, 52, 54, 57, 58,
Radar 5, 20, 21, 22, 23, 24, 25, 29, 30, 32, 44, 50, 59, 60, 61,62, 63, 64, 65, 6 6 , 67, 6 8 , 69, 70, 71,
376,404,416,418,422,424,436,438,440,442, 72, 73, 75, 76, 77, 78, 79, 80, 81,82, 83, 84, 85,
4 4 4 , 446, 452, 454
8 6 , 87, 8 8 , 89, 90, 91,94, 97, 98, 100,101,102,

Radar equation 22 103,104,105,108,109,110,111,112,113,114,

Radar hologram 440 116,117,118,119,120,121,122,123,124,125,
480 Processing of Remote Sensing Data

126,127,128, 129,130,131,133, 134,135,136, 252, 255, 258, 266, 272, 274, 278, 283, 284,
137,138,139,140,141,142,143,145,146,147, 286, 290, 293, 294, 296, 300, 304, 306, 313, 314,
148,149,150,153,154,155,156,157,158,159, 318, 320, 322, 326, 328, 330,332, 334, 336, 342,
161,162,163,164,166,167,168,169,170,171, 346, 348, 350, 352, 356, 358, 359, 360, 364, 366,
172,173,174, 175,176,178,179, 181,182,183, 368, 382, 384, 386, 390, 392, 394, 396, 398,401,
184.186.187, 188,189,191,192,193,198,199, 402, 404, 405, 406, 408,410,412, 414,418,420,
201,202, 203, 204, 206, 208, 209, 212, 214,216, 422, 426,430, 432, 434,436,440, 442,444,446,
217, 219, 220, 221,222, 224, 226, 227, 229, 231, 452
235, 237, 238, 242, 243,244, 245, 246, 249, 250, Regrouping 146, 147, 150, 159, 284, 302, 328, 366
252, 256, 258, 260, 262, 264, 266, 268, 270, 272, Rejected 181, 182, 183, 186, 396
274, 276, 278, 284, 286, 288, 290, 294, 296, 298, Rejection 177, 181, 182, 184, 186, 302, 336
300, 302, 304, 306, 308,313, 316, 318, 320, 322, Rejection threshold 177, 181, 182, 186, 302
326, 328, 330, 332, 334, 336, 340, 344, 348, 350, Reliability 187,302,322
352, 354, 355, 356, 358,359,360, 362, 364, 366, Reliability of interpretation 322
368, 371,372, 373, 374, 376, 378, 380, 382, 384, Relief 111, 157, 248, 249, 253, 255, 256, 262, 265,
386, 388, 390, 392, 394, 396,398, 401,402,404, 266, 267, 269, 270, 271, 272, 273, 274, 275,
405, 406,408, 410,412,414,416, 418,420,422, 278, 315,316, 317, 318, 327, 396,441,443,444,
424,426,430, 432, 434,436,438, 440,442,444, 449
446, 451,452, 454 Remote sensing 3, 4, 5, 7, 9, 10, 12, 20, 21,22, 26,
Red-Green-Blue system 58, 62 30, 31,32, 36, 38, 40, 41,46, 52, 54, 57, 6 8 , 72,
Reference 6 , 19, 24, 40, 59, 6 8 , 70, 82, 8 6 , 89, 94, 73, 8 6 , 87, 89, 90, 97, 101, 102, 106, 113, 114,
103,110,113, 114,117,144,145, 159,161,167, 115,116,119, 142,154,161,175, 176,177,178,
174,176,177,187,188,189,191,198,199, 204, 186,189,192, 194, 233, 255, 260, 262, 278, 283,
205,231,232, 244,245,246,248, 249,250,251, 284, 286, 293, 294,300,306,308, 313,322,344,
252,264, 270, 272, 274,276,278, 293,298, 300, 346, 348,350, 352, 356,358,359, 360,366,368,
302,304, 306, 308, 314,316,322, 326,328,334, 371,372, 374, 376, 378, 394, 396, 398,401,402,
336,340,348, 352,358,360,364, 366,368,373, 404, 405,408, 410,412,414,416, 418,420,422,
374, 378, 384, 392, 396, 398,422, 426,440,442, 424, 426, 432, 436, 444, 446, 454
451,452 Repeatability 188
Reference classes 306 Repetition 46, 48, 52, 72, 244, 286, 304, 376, 416,
Reference data 188, 189, 191, 231, 244, 246, 248, 426, 442, 444
249, 252, 298, 304, 306, 366, 373 Resampling 290, 294, 302, 356
Reference plots 334, 336 Resampling of pixels 356
Reference points 114, 244, 246, 249, 252, 272, 276, Resolution 15, 18, 23, 25, 34, 36, 38, 41,43, 44, 46,
302, 304, 306, 340, 422, 442 48, 50, 52, 90, 94, 97, 99, 106, 109, 111, 114,
Reflectance 7, 8 , 12, 2 1 , 61, 63, 64, 65, 6 8 , 72, 73, 122,126,129,131,150,188,189,193,227,248,
74, 75, 76, 77, 78, 79, 80, 81,82, 83, 84, 85, 8 6 , 249, 252, 256, 260,262,278,284, 286, 288,290,
8 8 , 91, 102, 114, 122, 123, 128, 132, 134, 136, 293, 294, 296, 302, 304, 314,316, 318, 326, 332,
161,167,173, 191,288, 290, 294, 304, 316, 318, 358, 366, 368, 376, 390, 394,404, 408, 410,412,
348, 350, 352, 354, 378, 380, 382, 384, 386, 388, 414,416,418, 422,424, 426,434, 436,438,440,
390, 392, 398, 405, 406, 418, 420, 422, 454 4 4 4 , 446, 452, 454

Reflectance curve 61, 63, 64, 65, 75, 81, 83, 136, Resolution of pixels 287, 302
173, 378, 382, 384, 386, 418, 406, 420 Restoration 36, 61,268, 332
Reflected energy 128 Resultant image 202, 221,236
Reflective 5,31,50,73,79, 80,81,82, 83, 88,91, 129, Retina 57, 58, 264
139, 294, 356, 362,364, 368,386, 388,390,451 Revisit capability 42,43,46,48,50,52, 376,418, 446
Reflective infrared 364, 451 RGB 62, 63, 69, 209
Reflective middle infrared 73,79, 80, 81,83, 88,129, River 98, 101, 102, 108, 114, 125, 134, 135, 139,
139, 294, 356, 362, 368, 386, 388, 390 146,152,155, 156,157,158,160, 161,165,166,
Refractive index 78 174,179,180, 184,195,198,200, 201,202, 204,
Region 3, 4,10,15,16,17, 2 2 , 30, 36, 38, 42,43, 44, 205, 227, 270, 316, 320, 322, 324, 334, 336, 364,
46, 50, 8 6 , 8 8 , 98, 99, 100, 101, 102, 103, 106, 366, 368, 396, 424, 446
107,110,112,113,118,122,123, 138,141,142, River basins 424
145,148,150, 155,158,159,160,161,166,167, Riverine forests 134, 139, 152, 156, 157, 160, 161,
170.172.174.175.176.178.179.182.184.187, 165,166,179,180,184,185,317, 320, 321,324,
188,191,192,199,221,222,223, 235,244,250, 365, 366
 Index 481

RMS 301,302 191,192,193,194,211,214, 221,223,231,236,

Road 52, 6 8 , 79, 97, 98, 100, 101, 1 0 2 , 104, 108, 238, 243, 245, 246, 248, 249, 250, 251,252, 255,
109,113,114, 122, 125,139,146, 153,155,157, 266, 274, 284, 286, 288, 290, 296, 298, 302, 304,
166,168,171, 172,178,195, 227, 229, 246, 249, 308, 316, 318, 322, 326, 328, 330, 334, 336,342,
252, 316, 318, 332, 334, 336, 338, 339, 340, 341, 348, 350, 359, 360, 362, 366, 372, 376, 378, 390,
342, 362, 366, 368, 422, 442 394, 396, 398, 408, 414,422,424, 426, 430,444,
Rocks 318, 332, 394, 396, 402, 405, 406, 408, 410, 446, 451
412, 414, 418, 444 Satellite 11, 12, 13,14,16, 17, 23, 30, 31,32, 34, 36,
Roll 23, 30, 41,61,72, 142, 155, 234, 256, 260, 272, 38, 40, 41,42, 43, 44, 46, 48, 50, 52, 64, 6 8 , 69,
274, 306, 308, 314, 371,374, 376, 382, 386, 402, 71,75, 79, 84, 8 6 , 8 8 , 90, 97, 99, 100,101,106,
444 109,110,111,113,114,117,118,121,122,125,
Roman roads 1 00 126,127, 128, 130, 136,137,139, 142,145,157,
Root mean square 26, 300 167, 176,193, 194, 198, 206, 234, 235, 243, 245,
Rotating mirror 34, 46 246, 249, 255, 256, 258, 264, 266, 268, 272, 274,
Rotation 34, 41, 42, 44, 46, 63, 89, 239, 240, 241, 276, 278, 284, 286, 290, 294, 296, 304, 308, 313,
243, 246, 274, 286, 316, 386 314, 316, 318, 320, 322, 326, 328, 330, 332, 342,
Roughness 16, 17, 19, 21,23, 25, 26, 27, 28, 29, 30, 348, 350, 352, 356, 360, 364, 366, 368, 371,372,
83, 84, 8 6 , 8 8 , 304, 378, 379, 380, 381,382, 383, 376, 378, 384, 386, 388,394, 396, 398, 404,405,
388, 389, 391,393, 443, 444, 445 410, 416, 422, 424, 426,430,432, 436, 438,442,
Row 5, 10, 12, 20, 21,22, 25, 26, 27, 29, 30, 31,34, 444, 446, 451,452, 454
36, 44, 48, 50, 52, 6 8 , 69, 70, 79, 80, 83, 90, Satellite photos 106, 451
100,113,117,128,129,134,152,153,154,156, Satellite velocity 38, 40
157,160,166,179,183,201,202, 203, 204, 205, Savannah 348, 386
214, 228, 238, 288, 290, 294,296, 300, 306, 316, SAVI 86,87
318, 322, 336, 356, 366, 368, 374, 376, 378, 380, Scale 19, 25, 36, 94, 98, 99,109,111,145,189,197,
386, 390, 398, 405, 406,410,414, 416,418, 420, 206, 238, 242, 246, 248, 250, 252, 255, 256,
422,424, 426, 430,436,438,442, 444, 446,452, 258, 260, 264, 266, 268, 270, 272, 274, 276, 283,
454 284, 313, 316, 322, 328, 330, 332, 338, 340, 342,
Row effect 50, 80 344, 355, 364, 366, 368, 394, 396, 398, 404,405,
Rules 118, 119, 136, 140, 188, 191,322, 231,342 410, 414,416, 424, 430,432, 436, 442,444,446,
Runoff 446 452
Scale changes 197, 206, 238, 256, 283, 284
Scale factor 242
Scale transfer 396
Salinity 30, 390, 416, 418, 420 Scale variations 270, 272, 276
Salitrals 420, 422, 424 Scanner 5, 34, 36, 44, 48, 50, 8 6 , 90, 248, 256, 260,
Salt 30, 65, 83, 84, 8 8 , 2 2 2 , 332, 346, 359, 364, 378, 410, 416, 418, 438,454
382, 386, 388, 390, 414, 418, 420, 444, 454 Scanning system 44, 50
Salt efflorescence 83, 378, 382, 454 Scatterometers 23, 24, 25, 50, 436
Sample size 302 Schistose 402
Samples 74, 304, 306, 356, 378, 380, 382, 386, 390, Sea 9, 10, 13, 15, 25, 29, 30. 48, 50, 52, 54, 70, 75,
406, 408 79, 8 8 , 89, 90,114, 118, 128, 129,137, 142,144,
Sampling mode 304 154,176,194,198, 226, 229, 231,332, 262, 264,
Sampling plan 368 270, 278, 290, 294, 308,314, 318, 322, 342, 348,
Sampling procedure 304 352, 360, 364, 368, 371,372, 374, 376, 380, 390,
Sand 13, 29, 31,42, 57, 69, 8 8 , 97, 109, 138, 144, 398,401,414, 416, 418,420,422, 424, 426,432,
146,149,150,158, 201,202, 304, 316, 332, 336, 436, 444, 446
352, 378, 380, 386, 388, 390,394, 396,402,408, Sea roller 30, 444
410, 414, 418, 420, 422, 426, 442, 451 Sea rollers 30
Sandstone 402,410 SEASAT 30, 426, 436
SAR 3, 14, 15, 16, 18, 22, 29, 30, 36, 50, 57, 58, 67, Seashores 114
6 8 , 70, 71, 72, 89, 90, 97, 103, 104, 109, 110, Season 9, 70, 137, 142, 176, 290, 294, 314, 318,
111,113,114,115,116,118,119,120,121,123, 322, 348, 352, 360, 364, 371,380, 390, 398,420,
124,127,132,133,134,136,138,140,142,143, 432, 444
145,150,153, 160,161,162,163,164,165,166, SeaWiFS 90, 426
168,175,177,178,179,183,184,186,188,189, Secondary colours 59
482 Processing of Remote Sensing Data

Sediment 27, 8 8 , 401,417, 421,424, 425 221,226, 233, 288, 296, 382,386, 390,422,436,
Sedimentary formations 402,410 438, 440, 442, 444, 446, 448, 452
Sedimentary dynamics 416 Signature 72, 176
Segment 17, 98, 115, 125, 133, 136, 137, 141, 143, Significance of errors 308
144,145,146,147,148,150,153,155,160,162, Silicates 406,410
164,167,168,169,170,171,172, 175,182,187, Silts 388, 396
192, 222, 223, 228, 230,231,233, 241,242, 264, Simplification 284, 340
266, 270, 274, 286, 296,302, 313, 334, 360, 362, Simulation studies 308, 328
366, 373, 386, 392,422, 424, 438, 440,442, 452 Sinks 410
Segmentation 115,125,127,133,137,141,145,146, Size 17, 28, 29, 36, 75, 78, 84, 91, 94, 98, 99, 101,
147,150,160, 162,164,167,168, 169,170,171, 106,109,116, 130,145,154,174, 176,195,196,
172,175,187, 222, 223, 230, 231,233, 286, 313, 197,198,199, 200, 204,205,212, 215, 216, 221,
422 227, 228, 237, 239, 243, 249, 256, 284, 286, 288,
Segmentation method 233 290, 296, 298, 302, 304, 316, 318, 320, 322, 324,
Segmentation of histogram 169 325, 326,336, 340, 342, 356,360, 373, 388, 390,
Segmentation of image 125, 171,231 401,404,405, 408, 414,418,420,438,440,452,
Semantic 70, 96, 97, 103, 104, 108, 1 1 0 , 126, 132, 454
133,145,146,150,175,183,189, 191,283, 286, Skeletisation 219,228
300, 302, 304, 308, 322, 326, 328, 336,394, 398, Skewed 122
422, 451,452, 454 SKYLAB 48, 52
Semantic distance 126,398 Slaking 84, 8 6 , 378, 380, 382, 388, 454
Semantic errors 286, 308 Slaking crusts 84, 8 6 , 378, 380, 454
Semantic evaluation 336 Slant range resolution 438
Semantic information 132, 133, 146, 394, 398 Slope 86,89,96, 100, 103, 106, 108, 111, 125, 130,
Semantic precision 302, 422 141.142.155, 223, 224,249, 266, 274,278, 286,
Semi-controlled mosaic 274 314, 316, 318, 320, 322, 324,326, 328, 356, 360,
Senescence 78, 79,81 364, 366, 378, 384, 392, 396, 410, 434, 442
Senescent 74, 75, 78, 80, 348 Small agricultural zones 313, 328
Senescent vegetation 74, 78, 80 Smooth surface 27, 81,444
Sensitivity 20, 34, 48, 58, 6 6 , 67, 8 6 , 87, 90, 129, Smoothing 199,284
436, 444, 454 Snakes 230, 231,233
Sensor 5, 6,12,13,14,15, 16,17, 20, 21,31,32, 34, Snow 9, 13, 30, 31,70, 72, 91, 122, 294, 316, 332,
36, 38, 41,43, 44, 46, 48, 50, 52, 84, 90, 114, 392, 436, 444
416,128,136,188,191,193,221,222, 234, 235, Snow gauging 444
243, 252, 255, 256, 258, 260, 262, 264, 266, 272, Software 13, 71, 109, 110, 113, 115, 116, 118, 120,
278, 286, 288, 293,372, 376, 380, 386,390,416, 121,122,123, 133,136,142,161,167,176,178,
418,424,426, 432,436,438,440, 442,444, 451, 188,198, 243, 246, 248, 250, 251,276, 286, 290,
452, 454 313,316,328
Separability 177, 179, 180, 182, 183, 184, 186 Soil vii, viii, 13, 15, 16, 18, 19, 26, 27, 28, 29, 30, 48,
Separability table 177, 179, 182, 183, 186 63, 64, 69, 70, 71,72, 75, 77, 79, 80, 81,82, 84,
Sequence 52, 86,138,163, 181,188,194, 233,234, 85, 8 6 , 87, 92,105,106,107,117,118,122,130,
243, 246, 266, 274, 288,302,339, 359, 376, 394, 131.134.137.149.150.151.152.153.154.155,
404, 440, 446, 452, 454 156,157,161,170,171,172,173,174,179,180,
Shadow 41,80, 84, 87,109,111,123,128,130,131, 182,184,185, 187, 201,202,203, 204, 205, 283,
134.135.138, 139,141,149,151,152,153,155, 284, 294, 295, 296, 304, 313, 314, 318,319, 320,
156,157,158,161,166,167,170, 171,172,174, 322, 328, 329, 349, 350, 353, 357, 373, 375, 376,
175,179,180,184, 200, 201,223, 242, 264, 270, 377, 378, 379, 380, 381,382,383, 385, 386, 387,
286, 316, 318, 320, 350, 378,380, 382, 388, 390, 388, 389,390, 391,392,393,394, 395,396, 397,
392, 394, 442, 452 398, 399,400, 404, 408,409,412, 413,421,423,
Shrimp culture 422, 424 424,432,433, 435,436,437,446, 447,452, 453,
Side-looking radar 404 454, 455
Sidereal day 41 Soil Science 43, 57, 63, 84, 97, 105, 112, 113, 400,
Signal 1 1 , 16, 2 2 , 24, 25, 26, 27, 28, 29, 30, 31, 32, Soil and vegetation 48, 81, 118, 134, 137, 150, 152,
34, 36, 50, 57, 8 8 , 89, 90, 94, 114, 122, 124, 158, 179,446
125.128.139, 208, 209,212, 213, 214, 219, 220, Soil cluster 8 6 , 157, 158, 170, 171, 172
 Index 483

Soil landscape 394, 396, 398, 454 Spectral reflectance factor 8

Soil mapping 105,378,394,452,454 Spectral resolution 36, 52, 90, 293. 294, 412, 418,
Soil moisture 386, 392, 436, 446 426
Solar angle 1 1 1 , 156, 316 Spectral response 8 8 , 193, 418, 420
Solar day 41 Spectral signature 72,176
Solar time 9, 41 Specular 7, 27, 78, 8 8 , 316, 388, 442
Somersault effect 234, 243 Specular reflection 7, 27, 78, 8 8
Source 3, 5, 6 , 7, 16, 22, 23, 24, 25, 26, 27, 31,32, SPOT 32, 36, 38, 41,42, 43, 44, 46, 50, 52, 6 8 . 8 6 ,
34, 36, 44, 48, 75. 97, 113, 155, 156, 208, 214, 87, 89. 94. 96,107,109,110,128,131,136,137,
221,227, 236, 238, 244,246, 252, 264,278, 292, 138,142,145,150,157.189,193,197, 219, 222,
294,296,298, 302,314,318,320, 326,330,340, 227, 228,229, 231,234,239,242, 246, 249,250,
342, 348, 350, 358, 370,371,376, 401,402,410, 251,252,258, 264, 266,274.276, 284, 286, 290,
412, 414, 436, 438, 440, 442, 444, 452, 454 294,296,304, 316, 318,320.328, 330,332,340,
Source image 221,236, 244 342,352,354, 356,360,368,372,376,380,382,
Soyuz 52 388,394,398,404,410,416,422,426,438,451,
Space shuttles 52, 442 452,454
Spatial analysis 70, 96,153,158,314,320, 322,359, Spring crops 70, 290, 334
452, 454 SPZ 366
Spatial conformity 245 Square grid 209, 2 1 2 , 247, 274, 316, 374
Spatial distribution 97, 101,103, 133,145, 146,250, Square root 118, 124, 135, 239
314, 318, 326, 359, 452 Stages 72, 81,86, 90, 190, 208, 236, 284, 294, 318,
Spatial field 99, 109, 153, 197, 283, 284, 314, 396, 346,348,350, 354, 356,368,373, 396,422,426,
454 442
Spatial filter 214, 290 Standard deviation 119,122,136,137,154,178,226,
Spatial filtering 214 288, 300, 302, 360, 362, 380, 454
Spatial integration 18, 334, 336 Statistic 54, 81, 108, 114, 115, 116, 117, 119, 121,
Spatial interpretation 153 122,123,136,141,142,143,146,148,150,156,
Spatial map 246, 422 161,162,174,177,178,181,186,187,188,189,
Spatial model 104, 107, 235, 313, 328 196,197,199, 216, 234,288, 298, 302, 304, 306,
Spatial organisation 96,153,158,159,161,172,174, 320, 328, 336, 344, 352,362,371,372, 373, 374,
175,189,194, 249, 283, 284,313, 314, 320, 326, 376, 394, 404, 426
336, 394, 396, 454 Statistical analysis 108, 150, 174, 178, 304, 328
Spatial resolution 25, 38, 43, 44, 46, 48, 50, 52, 94, Statistical interpretation 136, 148,150, 373
131,284, 326, 414,416,418,426, 434,436,438, Statistical method 115,117,161,288, 298,304,306,
444, 454 374, 376
Spatialisation 293, 358, 366, 368 Statistical tests 119,121
Specific indices 355 Stefan-Boltzmann 3
Speckle 440, 446 Stereo restitution 268, 276
Spectral bands 5, 1 0 , 13, 15, 36, 38, 42, 43, 44, 46, Stereoscope 43, 266, 268
48. 52, 67, 74, 75, 83, 89, 90, 127, 128, 129, Stereoscopic 43, 52, 106, 110, 1 1 1 , 258, 262, 266,
130,131,150. 219, 284,288,290, 294,296,298, 268, 270, 272, 274, 276, 316, 320, 326, 398,442
302, 378,392, 416,418,424,436,440,446,451, Stereoscopic model 442
454 Stereoscopic pair 43, 110, 258, 266
Spectral behaviour 72, 73, 75, 76, 77, 78, 79, 80, 81, Stereoscopic view 262, 268, 276, 316
82, 8 6 , 8 8 , 114, 127, 131, 175, 176, 349 Stereoscopic vision 106, 110, 111, 266, 268, 270,
Spectral characteristics 31,72,73, 74,75,78,79,82, 274, 320
84. 8 6 , 120, 122, 125, 126, 127, 128, 129, 133, Stereoscopy 42, 43, 44, 96, 255, 264, 268, 270, 274,
144,150,166,168,170,172,175,179,182,183, 278, 318, 320
184,188,191,198, 288,302,316, 348,350,352, Storms 414,418,444
358,364,366, 378,386,388,392,408,418,420, Stratified sampling 304, 306
422, 424, 452 Stress 19, 74, 8 6 , 87, 294, 374, 416, 430, 432, 434
Spectral classes 290 Stretching 123, 127, 244, 246
Spectral contrasts 408 Stripping 48, 454
Spectral emissivity 4,15 Structural 71
Spectral reflectance 7, 61,81,84, 85,173, 352,418, Structural analysis 115,116, 167,193, 194, 318
420 Structural classification 116,156,199, 298, 328
484 Processing of Remote Sensing Data

Structural element 115, 217, 218,219, 227, 322, 326, Synthetic descriptor 313, 314
454 Systematic errors 36, 302
Structural processing 71,94, 113, 193, 194, 313 Systematic verification 342
Structure 16, 17, 30, 52, 73, 76, 77, 78. 79, 80, 82,
84, 85, 91, 100, 106, 111, 115, 132, 144, 172,
194.198, 220, 222, 260, 268,284, 298, 314, 316,
322, 326, 328, 360, 372, 378, 382, 388, 390, Talweg 96, 103, 108, 111, 146, 172, 318, 320, 322,
394, 402,405,406,412, 414, 426, 434,442,446, 324, 394, 396, 398
454 TDRS 48, 454
Structure of vegetation cover 79, 82 Technology transfer 188
Subpixel composition 290 Tectonic 401,402, 404, 405, 410
Subsampling 284 Temperate zones 316, 344, 348, 394
Subtidal 414,418,420 Temperature 3, 4, 12, 13, 14, 15, 16, 17, 18, 19, 20,
Subtractive 57, 59, 60, 61,66 21,30, 40, 50, 8 8 , 374, 378, 404, 405, 408, 410,
Sun 4, 6 , 9,10,11,17, 32, 38, 40, 41,42,46, 48, 52, 414, 416, 422, 426, 430, 432, 434
58, 71, 75, 83. 8 6 , 87, 242, 264, 274, 294, 318, Temporal profiles 290
378,380, 390, 396,418, 454 Temporal resolution 46, 416
Sun-synchronous 40, 41,42, 46, 48, 274 Temporal sequence 8 6
Super-wide angle cameras 258 Temporal variations 16,342,446
Supervised 113, 115, 116, 144, 145, 150, 167, 175, TeraVue 113,122,132,146, 148,176,181,186,246,
189.191.198, 288, 290,298,328, 334, 336, 348, 334
350, 352, 356, 362, 364, 366 Terraces 111,316,318,396
Supervised classification 113, 115, 116, 144, 145, Territorial 111,414
191.198, 288, 298, 328,336, 348, 352, 356, 362, Terroirs 106
364, 366 Tessellations 235, 244
Supervised method 150, 350 Test zones 31
Surface 4, 6 , 7, 9, 1 2 , 13, 14, 15, 16, 17, 18, 19, 2 0 , Textural 94, 96, 114, 115, 117, 141, 167, 193, 194,
21,22, 23, 24, 25, 26, 27, 28, 29, 30, 31,32, 34, 198,229,296,320,326,398
36, 50, 64, 70, 76, 78, 80, 81, 83, 84, 8 6 , 8 8 , Textural analysis 115,193
89, 90, 131, 153, 165, 167, 172, 174, 179, 182, Textural environment 296
189, 195, 204, 241, 244, 245, 256, 260, 276, Textural processing 94, 96, 115, 117, 398
278, 283, 288, 294, 296, 302, 304, 314, 318, 322, Texture 114, 115, 194, 288, 342, 360, 378, 388, 390,
328, 332, 334, 344, 348,350,364, 376, 378, 380, 396, 454
382, 384, 386, 388, 390, 392,394, 396,401,404, Thematic group 116, 156, 179, 181, 183, 184, 186,
405,406,408, 414, 416,418,420, 422,424,426, 187,188,189,454
430, 432,434, 436,438,440,442, 444,446,452, Thematic Mapper (TM) 131
454 Thematic masks 142
Surface emissivity 422 Thermal 3, 4, 5, 9, 10, 11, 12, 13, 14, 16, 17, 18, 19,
Surface roughness 2 1 20, 21,34, 44, 46, 48, 50, 52, 90, 128, 129, 131,
Surface state 165,172,294,314,328, 378, 382,390, 260, 294, 296, 386, 390, 398,401,402,404,405,
392, 394, 422, 424, 444, 454 406,408,410, 416,418,420,422, 424,426,430,
Surface state of sea 424, 444 432, 434, 444
Surface state of soil 378, 394, 454 Thermal Inertia 294, 405, 408, 410
Surface temperature 12, 13, 14, 15, 16, 17, 18, 19, Thermal infrared 4, 9, 10, 11, 12, 13, 14, 16, 18, 20,
50, 404, 408, 416, 426, 430, 432, 434 21,44, 46, 48, 50, 52, 128, 129, 131,260, 294,
Swamp 321,365,418,421 296, 386, 398, 418, 420, 424, 430
Swath 36, 38, 50, 52, 436, 438 Thickening 218,219
Swells 424 Thiessen polygons 245
Symmetry 214, 238, 239 Thinning 218,219
Synergy of forms 404 Threshold 90,116,123,124,125,137,144,145,147,
Synoptic 8 6 , 106, 109,111,276, 313, 328, 394, 404, 149,177,181,182,186,205,221,224, 227, 228,
412, 416 229, 232,233, 284, 288, 294,302, 339, 362, 366,
Synoptic view 86,106, 109, 111,276, 313,328,404, 432, 454
416 Threshold indices 288
Synthesis of form 155 Thresholding of bands 362, 366
Synthesised generalisation 197, 206, 284 Tidal flats 28,29,332,418
 Index 485

Tilt 261,272 Vegetation 29, 44, 49, 51, 65, 6 8 , 72, 82, 117, 122,
Tin 406 126,128,130,131,132,133,135,139,149,150,
TIROS 40, 55 151,163, 167,171,318,322,355,435,443,448
Titanium 406 Vegetation (plant) cover 15, 26,30, 72, 80, 84, 85,94,
TM 53, 131 402
TM1 89 Vegetation coverage 405
TM2 89 Vegetation group 349
TM3 70, 89 Vegetation index 14-15, 85-87, 134, 138, 188, 291,
TM6 band 131 353-355, 375, 419, 430, 435
Tomography 455 Vegetation instrument 44, 451
Top-hat form 227, 232 Vein deposits 402
Topoclimates 433 Veins 402
Topography 442 Velocity 37, 42
Town 153, 319 Verification 191,375
Training zone 298, 455 Vertical exaggeration of relief 266
Transformation function 123-124 Vertical photographs 258
Transformations 123 Vertical view 313
Translation 237, 239, 243 Vertograph 269, 277
Transmittance 20, 31,73 VIGIE448
Transmission 73 Vineyards and orchards 327
Transverse resolution 438, 445 Vineyards 321,332, 432
Trench 99 Visible 5, 73, 80, 84, 294, 295, 350, 376, 378, 384,
Trimetrogon 259 386, 387, 422
Trimming 219 Visible spectrum 59
Trimodal 121 Vision 57
Turgescent 78 Visual 139, 174
Two (three)-dimensional histogram 305 Visual discrimination acuity 264
Two-dimensional histogram 130, 156, 157, 165, 363, Visual interpretation 71, 95, 97, 174, 266, 297, 321,
393 326,329, 333-334,342-343,350,360,369,372,
Typology 303 394

u Visualisation (3D) 241

VOISIN 194, 286, 291,297, 334, 336
Volcanic 446
UAA 347, 357 Volumes 302
Ultrametric 117,119 Voronoi diagram 244
Ultraviolet 5, 6 6 Vulnerability to soil erosion 329
Uniform film thickness 256
Universal time 9
Unsupervised 144
w
Unsupervised classification 115,144, 349, 363 Warping 246
Unsupervised methods 351 Water 10, 13, 21,26, 29, 30, 45, 6 8 , 72, 78, 8 8 , 110,
Uranium deposits (mineralisations) 402, 409, 411 122,128,130-133,135,139,148,150-151,155,
Uranium 402, 404, 406 157,161 -162,163,167,171,201,322, 338,365,
Urban 141 393, 411,424, 427, 444
User’s precision 303 Water bodies 126, 296
Water colour 90
V Water content 73, 294, 352, 356. 418
Water regime 295
Validation (verification) 292, 298, 309, 325, 333, 374, Water zones 317
398, 432 Watershed 415
Validity 117, 177 Water-retention capacity 388
Valley 201,317 Wavelength 3, 11,24, 38, 58, 70, 455
Value (Munsell) 63, 149, 379, 384, 386, 388, 451 Waves 26
Variance 287 Waxy cuticle 74
Vector data 248, 306 Wet (submerged) grasslands 361, 368, 369
Vectors 138 Wet soil 30
486 Processing of Remote Sensing Data

Wetlands 349, 359, 364-366, 399

Whiskbroom Scanning 35
White 59 Yellow 60-61, 127
Wien 3 Yield assessments 73
Wind 30, 51, 8 8 , 423, 424, 433, 434, 445 Yield estimation 372, 374
Window 194, 2 1 2
Window size 286, 336
Winter corn 81-82
Winter crops 70 Zenith 6
Winter 167, 395 Zenith angle 9
Zinc 406
Zoom 95, 234

X band 445, 448

Xanthophylls 73
 CD-Rom Image Index

CD 1.1, 4 CD 9.6, 196, 199, 200, 201

CD 2.1, 39, 51 CD 9.7, 201
CD 2.2, 46, 358 Cd 9.8, 201
CD 2.3, 46 CD11.1,221
CD 2.4, 51 CD 11.1 0, 226
CD 3.1, 70, 117 CD 11.11, 227
CD 11.12, 227
CD 4.1, 76
CD 11.2, 219,222
CD 4.2, 92, 197
CD 11.3, 219,222
CD 4.3, 50
CD 11.4, 222
CD 5.1, 102 CD 11.5, 218,223
CD 5.2, 114 CD 11.6, 224
CD 6.1, 126 CD11.7,224
CD 6.2, 47 CD 11.8, 218, 225
CD 7.1, 133 CD 11.9, 226
CD 7.2, 133 CD 12.1, 240, 241
CD 7.3, 134 CD 12.2, 237, 242
CD 7.4, 134, 135 CD 12.3, 247
CD 7.5, 136 CD 18.1, 346, 349, 355
CD 7.6, 136 CD 18.2, 357
CD 7.7, 139, 146 CD 18.3, 357
CD 7.9,147 CD 18.4, 357
CD 7.10, 148 CD 23.1, 436
CD 7.11, 149, 151, 152
CD 23.2, 438
CD 7.12, 15 312, 191
CD 24.1, 443, 444, 447, 449, 450, 452, 454
CD8.1, 157,158,159,161,164,165,169,170,
CD 24.2, 447
173
CD 24.3, 450
CD 8.2, 174 CD 24.4, 451, 453
CD 8.3, 175 CD 24.5, 453, 454
CD 8.4, 156, 175
CD 8.5, 176 CD 25.1, 467
CD 8.6, 177 CD 25.2, 468
CD 8.7, 179, 181 CD 25.3, 468
CD 25.4, 471
CD 9.1, 184, 185
CD 9.2, 186 CD 26.1, 50, 476
CD 9.3, 186, 187 CD 26.2, 478
CD 9.4, 189 CD 26.3, 50, 478
CD 9.5, 191 CD 26.4, 280