International Journal of Remote Sensing

ISSN: 0143-1161 (Print) 1366-5901 (Online) Journal homepage: https://www.tandfonline.com/loi/tres20

Flood boundary delineation from Synthetic

Aperture Radar imagery using a statistical active
contour model

M. S. Horritt , D. C. Mason & A. J. Luckman

To cite this article: M. S. Horritt , D. C. Mason & A. J. Luckman (2001) Flood boundary
delineation from Synthetic Aperture Radar imagery using a statistical active contour model,
International Journal of Remote Sensing, 22:13, 2489-2507, DOI: 10.1080/01431160116902

To link to this article: https://doi.org/10.1080/01431160116902

Published online: 25 Nov 2010.

Submit your article to this journal

Article views: 1415

View related articles

Citing articles: 24 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tres20
int. j. remote sensing, 2001, vol. 22, no. 13, 2489 –2507

Flood boundary delineation from Synthetic Aperture Radar imagery

using a statistical active contour model

M. S. HORRITT†, D. C. MASON
Environmental Systems Science Centre, University of Reading, 3 Earley Gate,
Whiteknights, P.O. Box 238, Reading RG6 6AL, England, UK

and A. J. LUCKMAN
Department of Geography, University of Wales, Swansea, Wales, UK

(Received 28 September 1998; in Ž nal form 22 November 1999)

Abstract. Flood extent maps derived from remotely sensed data can provide
distributed validation data for hydraulic models of  uvial  ow, and can be used
for  ood relief management and to develop spatially accurate hazard maps. A
statistical active contour model is used to delineate a  ood from the Ž rst European
Remote Sensing satellite Synthetic Aperture Radar (ERS-1 SAR) imagery as a
region of homogeneous speckle statistics. The segmentation uses both local tone
and texture measures and is capable of accurate feature boundary representation.
The results are assessed by comparison with simultaneous aerial photography,
the SAR segmentation scheme classifying 75% by area of the shoreline region
correctly. Seventy per cent of the shoreline coincides with the ground data to
within 20 m. The main error is due to un ooded vegetation giving similar radar
returns to open water.

1. Introduction
Recent research has identiŽ ed the potential of remotely sensed data to improve
our understanding of hydrological systems, especially surface phenomena such as
 ooding. Reviews of the use of remote sensing in  ood hydrology are given in Bates
et al. (1997) and Smith (1997), both identifying the all weather capability of Synthetic
Aperture Radar (SAR) imagery as a considerable advantage over other sensors
operating at visible or infrared wavelengths of the electromagnetic spectrum which
cannot penetrate the cloud cover often associated with  ood events. Flood extent
maps derived from remote sensing data can provide valuable distributed calibration
and validation data for hydraulic models of river  ow processes (Bates and Anderson
1995, Horritt 1998) and can be used for emergency  ood relief management and to
develop spatially accurate hazard maps.
The simplest model of SAR returns from  ooded areas assumes that the water
surface is generally much smoother than the surrounding dry land, and acts as a specular
re ector, giving low backscatter . This simplistic model is re ected in many previous
studies of  ood mapping through SAR imagery, which generally adopt a qualitative

†e-mail: [email protected]

Internationa l Journal of Remote Sensing

ISSN 0143-116 1 print/ISSN 1366-590 1 online © 2001 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
2490 M. S. Horritt et al.

approach (Badji and Dautreband e 1995, Calabresi 1995, Kannen 1995, Noyelle et al.
1995). Pre- ood, in- ood and post- ood imagery are combined in the red, green and
blue channels of a false colour composite image, in which the low returns from the
 ood waters generate a magenta hue, which is identiŽ ed and delineated manually.
ImhoV et al. (1987 ) uses a simple thresholding technique for automatic classiŽ cation of
 ooded regions in SAR imagery, and Oberstadler et al. (1997 ) use an evidential reasoning
technique for automatic classiŽ cation of the  ood area. This technique has the drawback
that the  ood boundary is poorly deŽ ned, and so small scale  ood features which the
high resolution of SAR systems is capable of detecting are ignored.
The imaging of the water surface is complicated by wind roughening and the
eVects of protruding vegetation, both of which may produce signiŽ cant radar returns.
Multiple re ections between the water surface and upright vegetation enhance back
scattering giving  ooded vegetation a bright radar return in the image, the magnitude
of this eVect being a function of radar look angle, wavelength and polarization.
Inundated forests have been identiŽ ed with high returns in the L-band (Richards
et al. 1987), but this eVect is reduced at shorter radar wavelengths due to increased
volume scattering in the canopy. Ormsby et al. (1985 ) found that the X-band gave
bright returns for  ooded marshland but no backscatter enhancement in forests.
Henderson (1995) detected high radar returns from  ooded forest and swamp and
Ramsey (1995) found an inverse relationship between SAR returns and the depth of
marsh  ooding. Solomon (1993) describes backscatter from rivers in tropical forest
areas and its variation with river orientation, with the water acting mainly as a
specular re ector when looking along the river but giving higher returns when
looking across due to multiple water/vegetation re ection. Wang et al. (1995) have
modelled these interactions mathematically, their results predicting that for C-band
SAR at an incidence angle of approximatel y 20° (the characteristics of ERS-1 and 2
SARs)  ooding under Amazonian forest canopies should increase backscatter by
approximatel y 2.6 dB. The eVects of wind on ocean surface roughness and radar
backscatter are also well known (Ulaby et al. 1986 ) but the eVect on terrestrial water
bodies is complicated by the variable fetch and eVects of local topography . Due to
the multiplicative nature of SAR noise, any increase in backscatter from the water
surface (whether due to wind roughening or vegetation) will also produce a higher
level of noise in the  ooded region.
Despite these complications, the use of SAR imagery compares favourably with
other remote sensing systems. ImhoV et al. (1987) compare Landsat and SIR-B
(Shuttle imaging radar) images with a control set of colour and infrared aerial
photographs of monsoon  ooding in Bangladesh. SAR data processed with a simple
density slicing and threshold technique provided a better identiŽ cation of the  ooded
area (85% correct) compared to Landsat (64% correct), when compared to the air
photograph control. Biggin and Blyth (1996 ) mapped a  ood from SAR imagery
of the River Thames, UK, validating the result against simultaneous aerial photo-
graphs. The images were analysed by eye with the SAR image correctly identifying
approximatel y 80% of the  ooded area.
The phase information in SAR imagery also has the potential to be used for  ood
mapping. Wegmuller et al. (1995 ) found that water could be identiŽ ed as regions of
low backscatte r and low interferometric phase correlation between 2 SAR scenes. Corr
et al. (1995 ) used coherence between ERS-1 images taken 3 days apart to distinguish
between Ž elds in winter and deciduous and coniferous forest, but achieved poor results
for 35 day repeat cycles. Crop growth and farming activities tend to reduce phase
 Flood boundary delineation 2491

correlation, making classiŽ cation more diYcult. Coherence mapping also tends to
generate results at a lower spatial resolution than the SAR image itself.
This research aims to rectify some of the current problems associated with  ood
mapping from SAR imagery. The development of an automatic segmentation algo-
rithm allows rapid development of objective  ood extent maps with accurately
deŽ ned (to within # 1 pixel ) boundaries in order to make best use of the spatial
resolution of SAR imagery. Comparison with simultaneous aerial photography data
generates a measure of conŽ dence in the maps derived from SAR imagery. The use
of coherence mapping for  ood detection is also assessed as an alternative to working
with intensity data alone.
This paper is organized as follows. In §2.1, a statistical active contour model for
SAR image segmentation is described, and the possibility of using coherence measure-
ments for  ood mapping is discussed in §2.2. Section 3.1 gives the results of applying
these techniques to imagery of a  ood on the river Thames, and compares them
with  ood maps derived from simultaneous aerial photographs . A Bayesian method
of estimating classiŽ cation accuracy is then developed in §3.2. Finally, the results are
summarized and future research needs outlined in §4.

2. Methodology
2.1. Segmentation using a statistical active contour model
Active contour models or snakes (Cohen 1991, Williams and Shah 1992) have
recently gained popularity as a means of turning incomplete and noisy edge maps
into smooth continuous vector segment boundaries. The snake method uses a
dynamic curvilinear contour to search the edge image space until it settles upon
image region boundaries, driven by an energy function that is attracted to edge
points, so that a snake will still settle on a line feature even if it is incomplete. The
contour is usually represented as a series of nodes linked by straight line segments.
The technique has been applied to SAR images of the North Sea shoreline on the
east coast of the UK (Mason and Davenport 1996). Edges are detected using a
constant false alarm rate contrast ratio technique to counteract the multiplicative
nature of SAR speckle, the snake being used as the edge linking stage in a much
larger image processing chain. Ivins and Porrill (1994) developed a statistical snake
that operates on the image itself, rather than an edge image, dispensing with the
need for a previous edge detection stage. It is suggested in their paper that mean
image intensity can be computed for a number of pixels lying along the contour,
rather than at a number of discrete points, which would give a level of noise reduction
but still result in accurate delineation of segment boundaries. This would seem an
ideal strategy for this problem, combating the eVects of SAR speckle but fulŽ lling
the aim of mapping the shoreline to within 1 pixel. This idea can be taken further,
as variance can be calculated along the contour, allowing local texture to be taken
into account in the segmentation process. It is stated in Luckman et al. (1997 ) that,
‘At C-band, SAR imagery often exhibits little variation in mean amplitude between
diVerent types of natural landcover’, and the failure of SAR amplitude to discriminate
strongly between diVerent surface features and cover types has led to texture being
used as the main classiŽ er in SAR imagery in many applications (Collins et al. 1997,
Henebry and Kux 1997, Soares et al. 1997 ). The capability of the statistical snake
to calculate image statistics over a number of pixels while maintaining well deŽ ned
segment boundaries makes it a potentially powerful tool for SAR image processing
for this application (and others), and worthy of further development.
A detailed description of the algorithm used here can be found in Horritt (1999 ),
2492 M. S. Horritt et al.

and only an overview of the main features follows. The statistical snake is formulated
as an energy minimization problem (Ivins and Porrill 1994), with the total snake
energy E(u(s)) being given by:

E(u (s))=E
tension
+E
curvature
­
PP G(I (x, y))dxdy (1)

Inside Contour
where u(s)=(x(s), y(s)) describes the contour position (x, y) in the two-dimensional
image space as a vector function of an arc length parameter, s. The contour is
represented in a piecewise linear fashion as straight lines linking a series of nodes.
E and E are energies generated by the model’s internal tension and
tension curvature
stiVness constraints, which favour a smooth contour made up of evenly spaced nodes
(this enhances the numerical stability of the model). G is a goodness functional that
assesses how well a set of image pixels (I(x, y)) meet certain criteria. The total energy
is minimized if the contour encloses a large area of good pixels, and in this respect
the model behaves as a region growing algorithm.
The driving term of the algorithm is the goodness functional which in this case
is based on the statistical properties of small samples taken from a random population
of image pixel intensity values. Image statistics are assumed to obey the simple
speckle model (Touzi et al. 1988, Oliver et al. 1994), with the order parameter equal
to the number of looks of the SAR system (three in the case of ERS-1 SAR PRI
(PRecision Image) data). The mean of small samples are assumed to be normally
distributed (according to the central limit theorem), and numerical experiment valid-
ates this assumption for sample size n 5 (see Ž gure 1 ). The measured variance is
assumed to be gamma distributed, with mean and order parameter found from
empirical models again determined by numerical experiment (see Ž gure 2).

Figure 1. Results of numerical experiment to Ž nd the distribution of the means of samples

of n=5, 10 and 20 pixels taken from a gamma-distributed population, along with the
normal distribution predicted by the central limit theorem. The mean and order
parameter of the original population were chosen to mimic typical ERS-1 three-look
amplitude imagery.
 Flood boundary delineation 2493

Figure 2. Results of numerical experiment to Ž nd the distribution of the variances of samples

of n=5, 10 and 20 pixels taken from a gamma-distributed population, along with the
distribution predicted by the empirical model.

If the mean and variance of pixel intensities along a segment of the contour are
measured, we can use this knowledge of how the mean and variance behave to assess
the probability that these pixels belong to the same region as that already inside the
contour. The goodness functional G is then based on the log of this probability, the
dependence on the measured mean mê , for example, having the following form:
n(mê ­ m)2
G(mê )=1­ (2)
vk2
where m is the mean of the seed population already enclosed within the contour, v
is the variance of the seed population, n is the sample size and k is a parameter that
can be adjusted to tune algorithm performance. This gives a measure which is equal
to 1 for a group of pixels with the expected mean, falling to 0 for pixels with a mean
diVering by kã (v/n) (i.e. k standard deviations) from the expected value. Taking the
energy to equal the log of the probability has parallels with the simulated annealing
formulation, where probability is exponentially related to the energy. The parameter
k is set at around 2 or 3 but may be increased further to allow for a level of statistical
inhomogeneity in the region being segmented. The dependence of the energy function
on the sample size, n, means that the false alarm rate (a false alarm occurring when
G<0 even for a population with mean m and variance v) is independent of sample
size. A large value of n, however, does narrow the region for which G>0, and so
enhances the model’s ability to discriminate between regions of diVerent speckle
statistics. Given this enhanced power of discrimination, the false alarm rate can
always be reduced by increasing k. The overall goodness functional (with components
based on both the measured mean and variance) is limited to a minimum value of
­ 1, as this makes interpretation of the model energy balance easier.
The roles of the tension and curvature constraints diVer between the usual edge
driven active contour model and the statistical version used here. If an edge detecting
2494 M. S. Horritt et al.

snake ‘snags’ on an isolated (noise) edge point, the resulting high curvature, tension
and in ation forces conspire to allow the snake to pass over the point. In this
algorithm, the curvature and tension constraints are viewed as a means of producing
a contour of appropriate smoothness with evenly spaced nodes, by a consideration
of the balance between image and curvature forces. The magnitudes of these energies,
which can be adjusted with a pair of weighting parameters, are usually determined
by a process of trial and error (Cohen 1991, Williams and Shah 1992, Ivins and
Porrill 1994 ), but here an analysis of the relationship between image and curvature/
tension forces in the model yields appropriate values for the weighting parameters.
The analysis places an upper bound on the curvature parameter for features of a
certain length scale r (curvature energy scales as 1/r), if the parameter is any larger,
the curvature will dominate the model energy and smooth out features of this length.
Similar analysis places an upper bound on the tension parameter (too large a value
will favour a short contour and sti e the growth of the snake), although a much
smaller value can be used to good eVect.
Since the  ooded region may not be simply connected (islands and isolated water
bodies form holes and outliers), the algorithm incorporates a method for dealing
with complex topology and snake self-intersection. An example of the algorithm
behaviour is shown in Ž gure 3, where the snake spawns a smaller sub snake to
represent an island.
The snake is represented as a series of nodes, each having a discrete location on
the image grid. The energy minimization scheme used is the greedy algorithm
(Williams and Shah 1992). At each iteration, change in energy, dE, is computed for
moves to all eight neighbours of each node in turn:
dE=­ GdA+dE +dE (3)
tension curvature
Obviously dE is equal to zero for no node movement. The lowest (most negative)
dE is chosen. The goodness functional G is calculated for pixels along the lines
linking the node with its two neighbours, and dA is the local change in area. If G is
positive, the snake is in a region of good pixels, a positive dA is favoured and the
snake expands. If G is negative, a negative dA is favoured, and the snake retreats
from the bad pixels. The statistics of the seed population, m and s, are calculated
from all pixels lying inside the contour, every 10 iterations. Where the contour is
artiŽ cially constrained by the image boundary, the curvature constraint is switched
oV, allowing the snake to expand freely along the edges of the image.

2.2. Coherence mapping

An algorithm based on backscatter intensity may have diYculty in situations
where wind- or turbulence-induced waves increase backscatter from water to the

Figure 3. The contour adapts its topology to describe an island. The snake meets itself,
spawning two new sub-snakes, one of which is recognized as an artefact and destroyed.
 Flood boundary delineation 2495

level of that from surrounding non- ooded areas. However, because the water surface
is in constant motion, it should exhibit strong temporal decorrelation of interferomet-
ric phase. A phase coherence map constructed from a pair of SAR images taken at
diVerent times should exhibit very low coherence over water bodies. Provided that
coherence from surrounding non- ooded areas remains suYciently high, this could
provide an additional means of distinguishing  ooded from non- ooded areas
(Wegmuller et al. 1995, Smith and Alsdorf 1997 ). Surfaces which change little with
time, such as bare soil, rock and buildings, tend to exhibit high coherences independ-
ent of the time between overpasses. The coherence measured for vegetated surfaces,
however, decays strongly as the time between overpasses increases as growth destroys
the correlation between the two scenes. Corr et al. (1995) measured coherence
between ERS-1 images taken 3 days apart (in winter, when vegetation growth is
minimal) and found high coherence values for Ž elds, which could be distinguished
from forest because volume scattering gave much lower coherences for the latter.
However, images with a 35-day repeat cycle also gave much lower coherences for
vegetation.

3. Results and analysis

3.1. T he snake algorithm and coherence mapping: comparison with airphoto data
Two reaches of the river Thames, both approximatel y 15 km in length, were
selected as test sites for the snake algorithm (see Ž gure 4). The dataset comprises
three ERS-1 SAR images, covering both test sites and acquired on three dates (see
table 1 ), with the image of 4 December 1992 capturing a  ood event. The pre- ood
and  ood SAR pair have a baseline appropriate for interferometric processing. A set
of oblique colour aerial photographs (Biggin and Blyth 1996), simultaneous with
the  ood SAR image (to within 2 hours), has also been acquired.
The  ood shoreline was delineated by eye from the aerial photographs and
vectorized. The shoreline vectors were then georeferenced using an orthographic

Figure 4. Schematic map of the Thames upstream of Oxford, showing the two SAR test
sites.
2496 M. S. Horritt et al.

Table 1. Details of ERS-1 SAR imagery used in the investigation.

Date Time (UTC) Frame Orbit Track Type

30 Oct 1992 11:01 2565 06752 00094 descending

4 Dec 1992 11:00 2565 07253 00094 descending
12 Feb 1993 11:00 2565 08255 00094 descending

Figure 5. Results of applying the snake algorithm to the Ž rst test site. The sub-regions shown
in Ž gures 7 and 8 are also shown (upper right and lower middle, respectively).
© European Space Agency.

Figure 6. Results of applying the snake algorithm to the second test site. © European Space
Agency.

transform parametrized by a least-squares method (Slama 1980) from 15–20 ground

control points for each photograph. Since some segments of the shoreline were visible
from more than one photograph, these could be used to assess the error (from both
georeferencing and delineation) in the shoreline location. Where this occurred, the
shoreline taken from the nearest Ž eld view was taken as the most accurate. The
error could also be veriŽ ed from segments where the shoreline was observed to lie
 Flood boundary delineation 2497

alongside a hedgerow or Ž eld boundary which could be located on 1:25 000 scale
maps, and was found to be less than 20 m.
The results of applying the snake algorithm to the in- ood SAR imagery are
shown in Ž gures 5 and 6 for the two test reaches. In applying the algorithm, a two-
stage strategy was adopted, the Ž rst stage using a nodal spacing of 12 pixels, no
tension or curvature constraints and with the texture component of the goodness
functional switched oV, lasting for 200 iterations. The snake was started as a narrow
strip lying along the course of the river channel, ensuring that it contained only
 ooded pixels. The Ž rst stage results in an approximate depiction of the  ood
boundary. The second stage used a nodal spacing of six pixels, and tension and
curvature constraints that allow the snake to depict features with a radius of half
the nodal spacing (this is generally the lowest useful value), and stabilized after 170
iterations for the Ž rst reach and 260 for the second. The constant k was set at 3.0 in
both stages to allow for some statistical inhomogeneity in  oodplain SAR statistics.
Figure 7 shows the 3 km×3 km detail highlighted in Ž gure 5, with the shorelines

(a)

(b)

Figure 7. Shorelines derived from aerial (red) and SAR (green) imagery for a 3 km×3 km
area, overlaid on the SAR imagery itself (a). Aerial photograph of the region, looking
north (b). © European Space Agency.
2498 M. S. Horritt et al.

found from SAR imagery and the aerial data overlaid on the SAR imagery and an
aerial view of the area. The scene shows well the sort of backscatter increase from
 ooded vegetation reported in the literature, for example in the system of hedgerows
present in the meander region (top right). Another problem is that some areas of
dry land give the same returns as the  ood, the island in the middle of the  ood, for
example. Figure 8 (the second region highlighted in Ž gure 5) illustrates this eVect for
another dry Ž eld, along with an aerial photograph of the area. Bare or sparsely
vegetated Ž elds (seen as brown in the photograph) exhibit the high backscatter
associated with wet soils (Giacomelli et al. 1995, Ulaby et al. 1996). A vegetated
canopy will reduce these returns, simple water cloud models (Prevot et al. 1993,
Taconet et al. 1996) predicting that the eVect will increase with canopy water content.
This is borne out by the green Ž elds in the scene which correspond with areas of
low SAR backscatter, and are likely to be misclassiŽ ed as  ooded.
Figure 9 shows a coherence map of the Ž rst test site (i.e. corresponding with the
image shown in Ž gure 5) with the vector shoreline found from the aerial photography.
The coherence map (# 80 m pixel size) was formed from the  ood and pre- ood

Figure 8. SAR and aerial view of a shoreline region, showing a vegetated Ž eld giving similar
radar returns to those from the  ood. The shoreline derived from aerial photography
(red) is shown on the SAR imagery, along with Ž eld boundaries (green). © European
Space Agency.
 Flood boundary delineation 2499

Figure 9. Coherence map for the Ž rst test site with the shoreline derived from the aerial
photography in red. © European Space Agency.

SLC (Single Look Complex) images, which were found to have a baseline suitable
for coherence mapping. Interferometric processing was carried out using the Atlantis
Earthview InSAR software. The  ooded region does indeed show consistently low
levels of coherence, but so do many other regions of the  oodplain, in agreement
with the Ž ndings of Corr et al. (1995) for imagery acquired under a 35-day repeat
cycle regime. Thus coherence mapping is of little use for this event.

3.2. Development of quantitative measures of algorithm performance

We Ž rst adopt a region based Bayesian approach in comparing the aerial and
SAR-derived vector shorelines, evaluating the probability of correct classiŽ cation
from the SAR imagery using the aerial photography as ground data. A Bayesian
method allows us to determine classiŽ cation accuracy given additional information,
such as landcover type and pre- or post- ood imagery, and to give a spatially
variable measure of conŽ dence in the SAR classiŽ cation.
Care is needed, however, in Ž xing the region over which this probability of
correct classiŽ cation is deŽ ned. Considering the whole test site, for example, may
bias the probability of dry regions having been identiŽ ed correctly towards 1. For a
small  ood in a large domain, classifying the whole region as dry is a reasonable
guess and would be considered a good classiŽ cation in terms of the probabilities.
This problem can be avoided if only the shoreline region is considered by limiting
the analysis to a strip containing the  ood boundary, but this raises the question of
how wide to make this strip. A narrow strip would exclude many misclassiŽ ed areas,
and a strip that was too wide would suVer the same problems as using the whole
domain in the analysis. A sensible compromise would be to make the strip width
equal to typical  oodplain Ž eld scales, so that it can include misclassiŽ ed Ž elds near
the shoreline but without introducing too much bias. Results are presented in table 2
for strips 300 m and 600 m wide along the shoreline and the whole domain for the
Ž rst test site and a strip 300 m wide for the second test site. These show only small
diVerences between the two sizes of strips, but a larger diVerence between statistics
for the shoreline region and the whole domain, which are biased towards 100% as
expected. The results imply that the statistics are reasonably stationary with respect
to strip width in the region 300–600 m, and can be used to predict the classiŽ cation
accuracy of the SAR segmentation algorithm in the shoreline region. They also show
2500 M. S. Horritt et al.

Table 2. Statistics of SAR algorithm performance.

Site no. Site no. Site no. 1, Site no. 2,

SAR Aerial 1300 m strip 1625 m strip full domain 300 m strip
Class status status (%) (%) (%) (%)

1 Wet Wet 37.5 38.5 22.6 33.3

2 Dry Dry 37.5 42.8 67.9 40.8
3 Wet Dry 15.5 11.2 5.4 19.7
4 Dry Wet 9.6 7.6 4.0 6.2
Correct 75.0 81.2 90.5 74.1

that the algorithm performs similarly in the two test reaches, identifying # 75% of
the area correctly.
It is possible that certain landcover types are more prone to misclassiŽ cation
than others ( because forest canopies may be impenetrable at C-band radar frequen-
cies, for example). Landcover classiŽ cation is available for most domestic  oodplains
in the form of the landcover map of Great Britain, developed from multitemporal
Landsat Thematic Mapper data (Fuller et al. 1994). The virtually guaranteed availab-
ility of this data would allow the application of this technique to many other test
sites in the UK.
To test this hypothesis, conditional probabilities of the form p(class|landcover
type) were calculated using:
p(class, landcover type)
p(class|landcover type)= (4)
p(landcover type)

Figure 10. Class probability against landcover type for the Ž rst test site.
 Flood boundary delineation 2501

Figure 11. Class probability against landcover type for the second test site.

Figure 12. Class probability against goodness functional for the pre- ood SAR imagery for
the Ž rst test site.
2502 M. S. Horritt et al.

This expression measures the probability of getting a particular class (in the range
1–4 as shown in table 2) and landcover type, but normalized by the prevalence of
that landcover type. Statistics were measured from the 300 m shoreline strip and
compiled into 2-D histograms and the results for the two test reaches are shown in
Ž gures 10 and 11. The results are inconsistent between the two sites, implying that
the results cannot be extended to other reaches. This is probably due to the lack of
correlation between landcover types and vegetation depth and density for the three
most common landcover types on the  oodplain: mown/grazed turf, meadow/verge/
semi-natural and tilled land. Tilled land, for example, will support a range of
vegetation depths depending on agricultural practices and the time of SAR image
acquisition.
Pre- ood SAR imagery may help in identifying areas prone to misclassiŽ cation,
such as Ž elds that give low returns even when dry, which could be erroneously
identiŽ ed as  ooded. A methodology identical to that used to analyse correlation
between misclassiŽ ed areas and landcover type was adopted. The goodness func-
tional, as used in the snake algorithm, was calculated for pre- ood imagery for both
test sites, thresholded into four classes, and statistics compiled from the 300 m strip.
The results are shown in Ž gures 12 and 13. The seed mean and standard deviation
used in calculating the goodness functional were taken from the  ood as found by
the snake algorithm for each reach, the goodness functional being calculated over
12 pixels as in the snake algorithm used to delineate the  ood (two line segments of
six pixels each). The results are similar for the two test sites, which implies that the
results should be applicable to other reaches, and the dependence on the goodness
functional is what would be expected. Areas with a high value of the goodness
functional in pre- ood imagery are more likely to be identiŽ ed as  ooded and fall
into classes 1 and 3.

Figure 13. Class probability against goodness functional for the pre- ood SAR imagery for
the second test site.
 Flood boundary delineation 2503

Probabilities like the ones displayed in Ž gures 12 and 13 can be inverted to give
a measure of conŽ dence in the segmentation produced by the snake algorithm based
on the goodness functional of the pre- ood imagery. For example, if an area has
been identiŽ ed as wet by the algorithm in a region with goodness functional G, the
probability that this identiŽ cation is correct will be given by:
p(wet, SAR wet, G)
p(wet| SAR wet, G)= (5)
p(SAR wet, G)
The probabilities on the right-hand side give the ratio of area identiŽ ed correctly as
wet from the SAR imagery to the total area identiŽ ed as wet from SAR, with the
given goodness functional. A similar expression is derived for areas identiŽ ed as dry
in the SAR scene. These probabilities derived from SAR imagery with ground data
can now be applied to imagery without referral to the airphoto data. The result is
a map of  ood probability, which is shown in Ž gure 14 for the Ž rst reach. The
probabilities of equation (5) were taken from both test reaches. The map shows
lower probabilities for regions identiŽ ed as  ooded but corresponding with darker
areas in the pre- ood imagery.
The shorelines derived from SAR and aerial photographic imagery can also be
compared directly in their vector form. The distance between each node of the
contour describing the SAR shoreline and the nearest point on the aerial shoreline
was measured and compiled into the histogram shown in Ž gure 15. The mean
distance between the two shorelines is 48 m, but this value is strongly in uenced by
the large tail of the distribution, which corresponds to regions where the SAR
segmentation has misplaced the shoreline by the order of a Ž eld length. The mean
of the distribution for distances less than 50 m is 20 m, which is a more representative
estimate of the shoreline error for 69% of its length.
Pre- ood imagery could be combined directly into the classiŽ cation process (as
in the multitemporal techniques referred to in the introduction), for example by
inhibiting the growth of the snake over areas that give low radar returns when dry.
The extension of the snake model to deal with multidimensional imagery is outlined
in Ivins and Porrill (1994). This would, however, ignore the uncertainty inherent in
classifying areas with similar SAR returns to the  ooded region, where there is little
contrast between  ooded and dry land. The region based approach adopted here

Figure 14. Map of  ood probability derived from in- and pre- ood SAR imagery.
2504 M. S. Horritt et al.

Figure 15. Distribution of distance error between SAR and aerial shoreline.

can deal with such areas, in a worse case scenario assigning a  ood probability of
50% to regions where the state of  ooding is unknown. The technique could also
be adapted to incorporate other remotely sensed or ground surveyed information
into the classiŽ cation process. There may well be a correlation between SAR classi-
Ž cation accuracy and vegetation depth which could be obtained from airborne
LiDAR (Light Detection And Ranging) survey (Lin 1997).

4. Conclusions and future research needs

A statistical active contour model has been applied to the problem of  ood
boundary delineation from SAR imagery. Two 15 km reaches of the Thames are
used to assess the performance of this algorithm by comparing the results with quasi-
simultaneous aerial photography. The eVects of  ooded vegetation explored in the
literature were conŽ rmed over the test reaches,  ooded hedgerows giving high returns,
probably due to a double re ection mechanism. A more serious practical diYculty
not previously identiŽ ed is that some regions of un ooded vegetation give very low
returns similar to open water, and are likely to be misclassiŽ ed as  ooded. These
regions can be identiŽ ed in imagery from before or after the  ood, and this multitem-
poral data used to generate a measure of conŽ dence in the derived shoreline. These
low returns were identiŽ ed with thickly vegetated areas, areas of bare wet soil giving
relatively high backscatter. Despite these diYculties, the SAR segmentation algorithm
classiŽ es 75% of the shoreline region correctly when compared with simultaneous
aerial photographic data. Furthermore, # 70% of the SAR shoreline is coincident
with the aerial shoreline to within 20 m.
Phase information in SAR imagery was also used in an attempt to map the  ood
as a region of low coherence. The 35 day interval between image acquisition does,
however, produce low coherence levels in many other regions of the image, in
agreement with the Ž ndings of Corr et al. (1995). This would seem to limit the use
 Flood boundary delineation 2505

of coherence mapping for  ood detection, unless imagery with a shorter repeat cycle
can be acquired, from the ERS-1/2 tandem mission, for example (Smith and
AlsdorV 1997 ).
The accuracy of the  ood mapping procedure may be improved by the adoption
of a more model based approach, rather than the somewhat heuristic method used
here, whereby the  ood is identiŽ ed as a region of (reasonably) homogeneous speckle
statistics. The real world situation is complicated by the vegetation cover on the
 oodplain. A qualitative comparison between the SAR imagery and aerial photo-
graphs has shown how bare soils produce high levels of radar backscatter, with
much lower returns coming from thickly vegetated areas. Enhanced backscatter from
 ooded vegetation is also observed, principally for linear features such as hedgerows
and the tree lined river banks. Vegetation mapping through airborne LiDAR and
multispectral scanners would enable these processes to be quantiŽ ed and catered for
explicitly in the segmentation process, perhaps via some form of physically based
model. This could permit the inversion of SAR returns into not only a binary
classiŽ cation of  ooded/un ooded, but also give information on water depths.
Flooding below vegetation may not be visible until the water depth approaches the
canopy depth. A region of enhanced backscatter may then occur, followed by open
water backscatter as the canopy is overtopped. Given the relative rarity of SAR
imagery of  uvial  ood events, it may be eYcacious to expand future studies to
coastal environments such as salt marshes where inundation occurs on a more
regular basis.

Acknowledgments
This work was done as part of a PhD studentship funded by the National Remote
Sensing Centre Ltd, Farnborough , UK. Thanks go to Ken Blyth of the Institute of
Hydrology, Wallingford, UK, for providing the SAR and photographic data, without
which this work would not have been possible.

References
Badji, M., and Dautrebande, S., 1995, Characterization of  ood inundated areas and delin-
eation of poor drainage soil using ERS-1 SAR imagery. Hydrological Processes, 11,
1441–1450.
Bates, P. D., and Anderson, M. G., 1995, Issues of hydraulic model validation and design
using remotely sensed data. Proceedings of the 1st ERS thematic working group meeting
on  ood monitoring, 26–27 June 1995 (Frascati, Italy: ESA-ESRIN).
Bates, P. D., Horritt, M. S., Smith, C. N., and Mason, D. C., 1997, Integrating remote
sensing observations of  ood hydrology and hydraulic modelling. Hydrological
Processes, 11, 1777–1795.
Biggin, D. S., and Blyth, K., 1996, A comparison of ERS-1 satellite radar and aerial photo-
graphy for river  ood mapping. Journal of the Chartered Institute of Water Engineers
and Managers, 10, 59–64.
Calabresi, G., 1995, The use of ERS data for  ood monitoring: an overall assessment.
Proceedings of the 2nd ERS applications workshop, 6–8 December 1995 (London: ESA)
pp. 237–241.
Cohen, L. D., 1991, On active contour models and balloons. CV GIP: Image Understanding,
53, 211–218.
Collins, M. J., Livingstone, C. E., and Raney, R. K., 1997, Discrimination of sea ice in the
Labrador marginal ice zone from synthetic aperture radar image texture. International
Journal of Remote Sensing, 18, 535–571.
Corr, D. G., Keyte, G. E., and Whitehouse, S., 1995, Studies of decorrelation in multi-
2506 M. S. Horritt et al.

temporal SAR imagery. International Geoscience and Remote Sensing Symposium, 1995
(Piscataway, NJ: IEEE), pp. 1026–1028.
Fuller, R. M., Groom, G. B., and Wallis, S. M., 1994, The landcover map of Great Britain:
an automated classiŽ cation of Landsat Thematic Mapper data. Photogrammetric
Engineering and Remote Sensing, 15, 1357–1362.
Giacomelli, A., Bacchiega, U., Troch, P. A., and Mancini, M., 1995, Evaluation of surface
soil moisture distribution by means of SAR remote sensing techniques and conceptual
hydrological modelling. Journal of Hydrology, 166, 445–459.
Henderson, F. M., 1995, Environmental factors and the detection of open surface water using
x-band radar imagery. International Journal of Remote Sensing, 16, 2423–2437.
Henebry, G. M., and Kux, H. J. H., 1997, Lacunarity as a texture measure for SAR imagery.
International Journal of Remote Sensing, 16, 565–571.
Horritt, M. S., 1998, Enhanced  ood  ow modelling using remote sensing techniques. PhD
thesis, University of Reading.
Horritt, M. S., 1999, A statistical active contour model for SAR image segmentation. Image
and V ision Computing, 17, 213–224.
Imhoff, M. L., Vermillion, C., Story, M. H., Choudhury, A. M., Gafoor, A., and Polcyn,
F., 1987, Monsoon  ood boundary delineation and damage assessment using space
borne imaging radar and Landsat data. Photogrammetric Engineering and Remote
Sensing, 53, 405–413.
Ivins, J., and Porrill, J., 1994, Statistical snakes: active region models. Proceedings of the 5th
British Machine V ision Conference, September 1994, York ( London: BMVA), vols 1
and 2, ch. 79, pp. 377–386.
Kannen, A., 1995, Use of SAR data for  ood mapping and monitoring in Thuringen, Germany.
Proceedings of the 2nd ERS applications workshop, 6–8 December 1995 (London: ESA)
pp. 237–241.
Lin, C. S., 1997, Waveform sampling LiDAR applications in complex terrain. International
Journal of Remote Sensing, 18, 2087–2104.
Luckman, A. J., Frey, A. C., Yanasse, C. C. F., and Groom, G. B., 1997, Texture in airborne
SAR imagery of tropical forest and its relationship to forest regeneration stage.
International Journal of Remote Sensing, 18, 1333–1349.
Mason, D. C., and Davenport, I. J., 1996, Accurate and eYcient determination of the shoreline
in ERS-1 images. IEEE T ransactions on Geoscience and Remote Sensing, 34, 1243–1253.
Noyelle, J., Delimiere, S., and Marinelli, L., 1995, IdentiŽ cation of  ooded areas in the
Rhone. Proceedings of the 2nd ERS applications workshop, 6–8 December 1995 (London:
ESA), pp. 237–241.
Oberstadler, R., Honsch, H., and Huth, D., 1997, Assessment of the mapping capabilities
of ERS-1 SAR data for  ood mapping: a case study in Germany. Hydrological
Processes, 11, 1415–1425.
Oliver, C. J., Blake, A., and White, R. G., 1994, Optimum texture analysis of synthetic
aperture radar images. Algorithms for Synthetic Aperture Radar Imagery, 2230, ch. 34,
389–398.
Ormsby, J. P., Blanchard, B. J., and Blanchard, A. J., 1985, Detection of lowland  ooding
using active microwave systems. International Journal of Remote Sensing, 5, 317–328.
Prevot, L., Dechambre, M., Taconet, O., Vidalmadjar, D., Normand, M., and Galle, L.,
1993, Estimating the characteristics of vegetation canopies with airborne radar
measurements. International Journal of Remote Sensing, 14, 2803–2818.
Ramsey, E. W., 1995, Monitoring  ooding in coastal wetlands by using radar imagery and
ground based measurements. International Journal of Remote Sensing, 16, 2495–2502.
Richards, J. A., Woodgate, P. W., and Skidmore, A. K., 1987, An explanation of enhanced
radar backscattering from  ooded forests. International Journal of Remote Sensing, 8,
1093–1100.
Slama, C. C., 1980, Manual of Photogrammetry, 4th edn (Falls Church, VA: American Society
of Photogrammetry).
Smith, L. C., 1997, Satellite remote sensing of river inundation area, stage and discharge: a
review. Hydrological Processes, 11, 1427–1439.
Smith, L. C., and Alsdorff, D. E., 1997, Flood mapping from phase decorrelation of tandem
ERS data: Ob river, Siberia. Proceedings of the 3rd ERS symposium on space at the
service of our environment, Florence, Italy, 17–21 March 1997 (Paris: ESA), pp. 537–539.
 Flood boundary delineation 2507

Soares, J. V., Renno, C. D., Formaggio, A. R., Yanasse, C. C. F., and Frey, A. C., 1997, An
investigation of the selection of texture features for crop discrimination using SAR
imagery. Remote Sensing of Environment, 59, 234–247.
Solomon, S. I., 1993, Methodological considerations for the use of ERS-1 SAR imagery
for the delineation of river networks in tropical forest areas. Proceedings of the 1st
ERS-1 symposium, Frascati, Italy, 26–27 June 1995 (London: ESA), pp. 595–600.
Taconet, O., Vidalmadjar, D., Emblanch, C., and Normand, M., 1996, Taking into account
vegetation eVects to estimate soil moisture from C-band radar measurements. Remote
Sensing of Environment, 56, 52–56.
Touzi, R., Lopes, A., and Bousquet, P., 1988, A statistical and geometrical edge detector for
SAR images. IEEE T ransactions on Geoscience and Remote Sensing, 26, 764–773.
Ulaby, F. T., Moore, R. K., and Fung, A. K., 1986, Microwave Remote Sensing, vol. 3
(Norwood, MA: Artech House).
Ulaby, F. T., Dubois, P. C., and van Zyl, J., 1996, Radar mapping of surface soil moisture.
Journal of Hydrology, 184, 57–84.
Wang, Y., Hess, L. L., Filoso, S., and Melack, J. M., 1995, Understanding the radar
backscattering from  ooded and non- ooded amazonian forests: results from canopy
backscatter modelling. Remote Sensing of Environment, 54, 324–332.
Wegmuller, U., Werner, C. L., Nuesch, D., and Borgeaud, M., 1995, Forest mapping using
ERS repeat-pass SAR interferometry. Earth Observation Quarterly, 49, 4–7.
Williams, D. J., and Shah, M., 1992, A fast algorithm for active contours and curvature
estimation. CV GIP: Image Understanding, 55, 14–26.