Landscape Ecol (2007) 22:171177

DOI 10.1007/s10980-006-9013-2

REPORT

Mapping spatial patterns with morphological image

processing
Peter Vogt Kurt H. Riitters Christine Estreguil
Jacek Kozak Timothy G. Wade James D. Wickham

Received: 11 April 2006 / Accepted: 6 May 2006 / Published online: 5 August 2006

Springer Science+Business Media B.V. 2006

Abstract We use morphological image process- results from image convolution, for a forest map
ing for classifying spatial patterns at the pixel le- of the Val Grande National Park in North Italy.
vel on binary land-cover maps. Land-cover
pattern is classified as perforated, edge, patch, Keywords Morphological image processing
and core with higher spatial precision and the- Spatial pattern Forest fragmentation
matic accuracy compared to a previous approach
based on image convolution, while retaining the
capability to label these features at the pixel level Introduction
for any scale of observation. The implementation
of morphological image processing is explained Indicators of forest fragmentation are reported by
and then demonstrated, with comparisons to several national and international environmental
programs (Montreal Process Liaison Office 2000;
Heinz Center 2002; Malahide 2004; USDA Forest
P. Vogt (&) C. Estreguil Service 2004) in assessments of forest health and
Land Management and Natural Hazards Unit biodiversity. Turner et al. (2001) describe ap-
(LMNH), European Commission DG Joint proaches to landscape and forest pattern analysis,
Research Centre, Institute for Environment and
and Li and Wu (2004) and Neel et al. (2004)
Sustainability (IES), T.P.261, Via E. Fermi 1, I-21020
Ispra, VA, Italy summarize some of the conceptual and practical
e-mail: [email protected] limitations. In practice, most forest fragmentation
indicators are motivated either by the concepts of
K. H. Riitters
adjacency and connectivity at the pixel level (e.g.,
US Forest Service, Southern Research Station, 3041
Cornwallis Road, Research Triangle Park, North Musick and Grover 1991) or by Formans (1995a)
Carolina 27709, USA landscape-level concepts of patch-corridor-matrix
and patch-mosaic. To satisfy requirements for
J. Kozak
comparability of data and indicators over large
Institute of Geography and Spatial Management,
Jagiellonian University, Gronostajowa 7, 30-387 geographic regions, the input data for assessments
Krakow, Poland are typically land-cover maps derived from re-
mote sensing.
T. G. Wade J. D. Wickham
An important aspect in the description of for-
US Environmental Protection Agency,
Environmental Sciences Division, Research Triangle est spatial pattern is the accurate identification
Park, North Carolina 27711, USA and mapping of internal and external

123
172 Landscape Ecol (2007) 22:171177

fragmentation (Zipperer 1993). Different causes map edge types according to the pair-wise iden-
of landscape change may have characteristic sig- tities of adjacent pixels forming an edge (Metzger
natures related to these types of fragmentation and Muller 1996) where the abundance of dif-
(Forman 1995b), and internal gaps in forest can- ferent edge types was used to estimate class-level
opies introduce biotic and abiotic edge effects and landscape-level indices of adjacency and edge
directly into interior forest. Bogaert et al. (2004) complexity. Metzger and Decamps (1997) used
developed a landscape-level classification algo- morphological operations to illustrate a proposed
rithm that identifies ten fragmentation categories, landscape-level habitat connectivity index called
including internal and external fragmentation, interior habitat percolation degree. However,
according to observed changes in patch area, indicators at landscape level provide only a single
number of patches, and patch perimeter in the value and are difficult to interpret independently
landscape. However, patch-based approaches are of composition (Neel et al. 2004). In addition,
difficult to implement in large-area assessments landscapes with substantially different arrange-
due to the large number of patches and the large ments of forest can have the same landscape-level
extent of the map. Reducing the number of pat- index value. In contrast, pixel-level classification
ches (e.g., Heilman et al. 2002) requires data permits mapping and monitoring of spatial pat-
modifications that can result in the loss of infor- terns at the pixel level which provides a greater
mation about small patches, and subdividing the sensitivity to pattern changes over time. To
map (e.g., Riitters et al. 2004) often truncates illustrate the approach, we present an algorithm
individual patches, which results in inaccurate that identifies the four types of forest patterns
estimates of patch size and shape. considered by Riitters et al. (2002), and we com-
Large-area, pixel-level mapping of internal and pare the results of the two methods obtained in
external fragmentation has been described by the Val Grande National Park (North Italy).
Riitters et al. (2000, 2002) and Civco et al. (2002).
These methods are based on image convolution
and do not require the identification of individual Methods
patches. Instead, a fixed-area window, or kernel,
is centered over each pixel on a forest map, and Matheron (1967) and colleagues introduced
an index is calculated according to the amount morphological image processing for the study of
and adjacency of forest in the window. This result the geometry of porous media. Recent textbooks
is then assigned to the forest pixel located at the describe mathematical morphology as a theory
window center, thus building a new map of the and technique based on set theory for analyzing
fragmentation index values. However, this ap- the shape and form of objects (Soille 2003). An
proach often misclassifies fragmentation (see Ri- algorithm to classify forest patterns is defined by a
itters et al. 2000) because: (1) it is based partly on sequence of logical operations such as union,
percolation theory, which applies strictly to ran- intersection, complementation, and translation
dom forest maps, and real landscapes are not using geometric objects called structuring ele-
random; (2) the fragmentation index is a cate- ments (SE) of pre-defined shape and size. We
gorical delineation of the continuous will provide only a verbal description of the
amountadjacency parameter space, and; (3) algorithm (see Soille (2003) for a formal mathe-
contextual information from outside the window matical language) and consider two SEs: an 8-
cannot be considered, resulting in unreliable in- neighborhood (SE1) and a 4-neighborhood
dex estimates which becomes increasingly appar- (SE2). We define forest connectivity in cardinal
ent for windows smaller than 9 9 pixels. directions only (SE2) and use the following two
In this paper, we use morphological image morphological operations. The erosion operator
processing (Soille 2003) as an alternate approach shrinks regions of forest and the dilation oper-
for reliable pixel-level classification and mapping ator expands them; the direction and extent of
of land-cover patterns. Morphological image these operations is defined by the shape and
processing has been used in landscape ecology to dimension of the SE.

123
Landscape Ecol (2007) 22:171177 173

comprises coherent forest regions that are too

small to contain core forest. Perforated forest
defines the boundaries between core forest and
relatively small perforations, and edge forest
includes interior boundaries with relatively large
perforations as well as the exterior boundaries of
core forest regions. We will demonstrate how the
SE can be adapted to tune the analysis scale to
specific definitions of far and large, which
permits differentiation of internal and external
fragmentation.

Step 1: Detect core forest

Fig. 1 Illustration of four types of spatial pattern on an
artificial map (see text for definitions)
Beginning with the forest map (Fig. 2, top row,
We consider four classes of forest pattern left panel), core forest (Fig. 2, top row, center
(Fig. 1). Core forest is relatively far from the panel) is obtained by applying an erosion with
forestnonforest boundary and patch forest SE1. The center pixel of SE1 is core forest if all

Fig. 2 Morphological erosion applied to the forestnon- (top row, right panel through bottom row, left panel) to
forest mask (top row, left panel) to identify core forest derive patch forest as the difference from the original
(top row, center panel). Next, dilations are executed until forest map (bottom row, right panel)
there is no difference between two consecutive dilations

123
174 Landscape Ecol (2007) 22:171177

Fig. 3 Morphological erosion and successive dilations forest regions showing patch and edge pixels (bottom row,
applied to the nonforest mask (top row, left panel through left panel) and edge after eliminating patch (bottom row,
center row, right panel). Final dilation into neighboring right panel)

eight neighbors are forest, resulting in all forest left panel). Here, patch pixels can never be added
regions being shrunk by 1 pixel. The difference because they are detached from the core-con-
between this core forest map and the original nected pixels. The first dilation adds forest pixels
forest map defines the pixels that are candidates that are directly connected to core forest, and
for the remaining classes of patch, perforated, and repeated dilations add forest pixels that are indi-
edge. rectly connected. The dilations stop when all
indirectly connected forest pixels have been ad-
Step 2: Detect patch forest ded. The difference between this map and the
original forest map is the set of patch pixels
Patch pixels are forest regions that do not contain (Fig. 2, bottom row, right panel).
core forest. They are identified after their com-
plement, the original forest map without patch Step 3: Detect edge forest
pixels, has been found. The latter is reconstructed
starting from the set of core forest pixels (Fig. 2, The detection of edge pixels starts from the
top row, center panel) and adding all forest pixels nonforest map (Fig. 3, top row, left panel). By
that are connected to this set which is achieved by analogy to Step 2, the nonforest patches are
repeated dilations with SE2 (Fig. 2, bottom row, identified and removed. Edge pixels can then be

123
Landscape Ecol (2007) 22:171177 175

Fig. 4 Final result

showing the four forest
spatial pattern classes

Fig. 5 Top: The forest

mask of the Val Grande
National Park with a
rectangular sub-region for
a magnified view of the
classification (Fig. 6).
Bottom: Comparison of
the forest class proportion
derived for the
convolution (C.) and the
morphological approach
(M.). The window
dimension (convolution)
or SE-size
(morphological) is shown
under each column

123
176 Landscape Ecol (2007) 22:171177

Fig. 6 Forest mask of the sub-region of Fig. 5 (left) and classification results for convolution (top) and morphological
approach (bottom)

identified by dilating the current nonforest areas bottom). The higher accuracy of the morpholog-
in all directions (using SE1) and looking for forest ical method at the pixel level is demonstrated by a
pixels instead of nonforest pixels. This step also direct comparison of the classified map (Fig. 6)
retrieves the forest patch pixels, which are re- for which the visual evidence shows clearly that
moved by subtraction (Fig. 3, bottom row, left image convolution is less accurate.
panel). With increasing SE (or window) size, both
methods increase the width of the perforated and
Step 4: Detect perforated forest edge regions at the expense of the core regions,
illustrating how changing the SE size tunes the
With knowledge of core, patch, and edge pixels, spatial scale of analysis in the same way as
the perforated pixels are obtained by subtraction changing the window size in image convolution.
as the only remaining unlabeled forest pixels With changes in SE size, the morphological ap-
(Fig. 4). proach remains accurate at the pixel level; small
patch regions remain patch regions and stay dis-
connected to neighboring core forest regions, and
Results and discussion continuous forest boundaries are labeled as a
single class. Because of the higher accuracy of
The 500 km2 Val Grande National Park is a Na- pixel-level mapping, summary statistics and trend
tura 2000 site in northern Italy. A 30-m resolution analyses at landscape level will also be more
forestnonforest raster map (Fig. 5, top, Estreguil accurate. These improvements will allow an
and Cerruti 2004) was used to compare the mor- unsupervised and precise spatial pattern analysis
phological approach with the image convolution at both, the pixel and landscape level.
approach. Four different scales of analysis were
investigated using four different structuring ele-
ment sizes and corresponding window sizes. Be- Summary
cause the two approaches use the same definition
of core forest, they yield identical core regions for Our intention in this brief note is to stimulate
the same SE (or window) size, but there are further interest in mathematical morphology for
important differences between the two ap- landscape pattern analysis. Previous applica-
proaches for the remaining three classes (Fig. 5, tions focused on landscape-level estimation of

123
Landscape Ecol (2007) 22:171177 177

percolation and adjacency indices (Metzger and Heinz Center (The H. John Heinz III Center for Science,
Muller 1996; Metzger and Decamps 1997), but Economics and the Environment) (2002) The state of
the nations ecosystems: measuring the lands, waters,
our approach may find more applications for and living resources of the United States. Cambridge
pixel-level classification and mapping of patterns, University Press, Cambridge UK
such as the identification of internal and external Li H, Wu J (2004) Use and misuse of landscape indices.
fragmentation that we demonstrated. We pur- Land Ecol 19:389399
Malahide (2004) Message from Malahide. The conference
posefully chose a simple example to enable Biodiversity and the EUSustaining Life, Sustaining
comparisons with an earlier approach for pixel- Livelihoods was held under the Irish Presidency in
level mapping in landscape assessments, and we Malahide, Ireland, 2527 May 2004. http://www.bio-
used verbal descriptions familiar to landscape diversity-chm.eea.eu.int/convention/F1117799202/
1112853936
analysts using geographic information systems or Matheron G (1967) Elements pour une theorie des milieux
image processing software. As a formal frame- poreux. Masson, Paris
work utilizing set theory and logic operations, Montreal Process Liaison Office (2000) Montreal process
mathematical morphology can address more year 2000 progress reportprogress and innovation in
implementing criteria and indicators for the conser-
complicated problems that may not be tractable vation and sustainable management of temperate and
otherwise. Our current research is extending the boreal forests. The Montreal Process Liaison Office,
concepts described in this brief note, and we are Canadian Forest Service, Ottawa, Canada
conducting additional studies needed to lay the Metzger JP, Muller E (1996) Characterizing the com-
plexity of landscape boundaries by remote sensing.
foundation for implementing the techniques in Land Ecol 11:6577
national and international assessments of forest Metzger JP, Decamps H (1997) The structural connectivity
health and biodiversity. threshold: an hypothesis in conservation biology at
the landscape scale. Acta Oecol Int J Ecol 18:112
Acknowledgments We thank Marco Cerruti for provid- Musick HB, Grover HD (1991) Image textural measures
ing data and in-depth knowledge on the Val Grande Na- as indices of landscape pattern. In: Turner MG,
tional Park. Many special thanks to Pierre Soille and Gardner RH (eds) Quantitative methods in landscape
Marcin Iwanowski from JRC/LMNH for the in-depth ecology. Springer-Verlag, New York
discussion on morphological image processing techniques. Neel MC, McGarigal K, Cushman SA (2004) Behavior of
class-level landscape metrics across gradients of class
aggregation and area. Land Ecol 19:435455
Riitters KH, Wickham JD, ONeill RV, Jones KB, Smith
References ER (2000) Global-scale patterns of forest fragmenta-
tion. Ecol Soc (formerly Cons Ecol) 4(2):3
Bogaert J, Ceulemans R, Salvador-Van Eysenrode D Riitters KH, Wickham JD, ONeill RV, Jones KB, Smith
(2004) Decision tree algorithm for detection of spatial ER, Coulston JW, Wade TG, Smith JH (2002) Frag-
processes in landscape transformation. Environ mentation of continental United States forests. Eco-
Manage 33:6273 systems 5:815822
Civco DL, Hurd JD, Wilson EH, Arnold CL, Prisloe Jr. MP Riitters KH, Wickham JD, Coulston JW (2004) A pre-
(2002) Quantifying and describing urbanizing land- liminary assessment of Montreal process indicators of
scapes in the northeast United States. Photogramm forest fragmentation for the United States. Environ
Eng Rem S 68:10831090 Monitor Assess 91:257276
Estreguil CM, Cerruti M (2004) Portfolio of Earth Soille P (2003) Morphological image analysis: principles
Observation based indicators for Biodiversity and and applications, 2nd edn. Springer-Verlag, Berlin
Nature Protection. EUR 21078/EN Turner MG, Gardner RH, ONeill RV (2001) Landscape
Forman RTT (1995a) Land mosaics. John Wiley and Sons, ecology in theory and practice: pattern and process.
New York Springer-Verlag, New York
Forman RTT (1995b) Some general principles of land- USDA Forest Service (2004) National report on sustain-
scape and regional ecology. Land Ecol 10:133142 able forests 2003. Report FS-766, USDA Forest Ser-
Heilman GE, Strittholt JR, Slosser NC, Dellasala DA (2002) vice, Washington
Forest fragmentation of the conterminous United Zipperer WC (1993) Deforestation patterns and their
States: assessing forest intactness through road density effects on forest patches. Land Ecol 8:177184
and spatial characteristics. BioScience 52:411422

123