IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 51, NO. 4, APRIL 2013 2417
A Change Detection Approach to Flood Mapping
in Urban Areas Using TerraSAR-X
Laura Giustarini, Renaud Hostache, Patrick Matgen, Guy J.-P. Schumann, Member, IEEE,
Paul D. Bates, and David C. Mason
Abstract—Very high resolution synthetic aperture radar (SAR)
sensors represent an alternative to aerial photography for delineating
floods in built-up environments where flood risk is highest.
However, even with currently available SAR image resolutions
of 3 m and higher, signal returns from man-made structures
hamper the accurate mapping of flooded areas. Enhanced image
processing algorithms and a better exploitation of image archives
are required to facilitate the use of microwave remote-sensing data
for monitoring flood dynamics in urban areas. In this paper, a
hybrid methodology combining backscatter thresholding, region
growing, and change detection (CD) is introduced as an approach
enabling the automated, objective, and reliable flood extent extraction
from very high resolution urban SAR images. The method
is based on the calibration of a statistical distribution of “open
water” backscatter values from images of floods. Images acquired
during dry conditions enable the identification of areas that are
not “visible” to the sensor (i.e., regions affected by “shadow”)
and that systematically behave as specular reflectors (e.g., smooth
tarmac, permanent water bodies). CD with respect to a reference
image thereby reduces overdetection of inundated areas. A case
study of the July 2007 Severn River flood (UK) observed by
airborne photography and the very high resolution SAR sensor on
board TerraSAR-X highlights advantages and limitations of the
method. Even though the proposed fully automated SAR-based
flood-mapping technique overcomes some limitations of previous
methods, further technological and methodological improvements
are necessary for SAR-based flood detection in urban areas to
match the mapping capability of high-quality aerial photography.
Index Terms—Algorithms, flood mapping, image processing,
satellites, synthetic aperture radar (SAR).
I. INTRODUCTION
THE support of remote sensing for mapping changes in
water surface extents and elevations has been demonstrated
widely (for detailed reviews, see [1]–[5]). Recently,
Manuscript received October 21, 2011; revised March 2, 2012 and June 7,
2012; accepted July 23, 2012. Date of publication September 7, 2012; date
of current version March 21, 2013. This work was part of the HYDRASENS
project, supported by the National Research Fund of the Grand Duchy of
Luxembourg and the Belgian Federal Science Policy Office in the framework
of the STEREO II research program (Contract nr. SR/00/100).
L. Giustarini, R. Hostache, and P. Matgen are with the Département Environnement
et Agro-biotechnologies, Centre de Recherche Public - Gabriel Lippmann,
4422 Belvaux, Luxembourg (e-mail: giustari@lippmann.lu; hostache@
lippmann.lu; matgen@lippmann.lu).
G. J.-P. Schumann and P. D. Bates are with the School of Geographical
Sciences, University of Bristol, Bristol BS8 1SS, U.K. (e-mail:
guy.schumann@bristol.ac.uk; paul.bates@bristol.ac.uk).
D. C. Mason is with the Environmental System Science Center, University
of Reading, Reading RG6 6AL, U.K. (e-mail: dcm@mail.nerc-essc.ac.uk).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TGRS.2012.2210901
the 2009–2010 Data Fusion Contest, organized by the Data
Fusion Technical Committee of the IEEE Geoscience and Remote
Sensing Society, focused on the evaluation of existing
algorithms for flood mapping through change detection (CD)
[6]. The success of these research studies together with recent
public and political awareness for quantifying global environmental
change has led to a significant increase in the number
of satellites dedicated to flood monitoring and hydrology in the
wider sense. Importantly, flood monitoring from space has the
advantage of large area coverage and relatively fast response
services (see for example the International Charter “Space and
Major Disasters” initiated by major space agencies: http://www.
disasterscharter.org/).
The vast majority of a flooded area is rural rather than urban,
and accordingly most literature on remote-sensing-based flood
detection to date has focused on the rural case. However, it is
perhaps more important to detect the urban flooding because
of the increased risks and costs associated with it. Flood extent
can be detected in rural floods using synthetic aperture radars
(SARs) such as ERS and ASAR, but these have too low a
resolution (25 m) to detect flooded streets in urban areas.
However, a number of SARs with spatial resolutions as fine as
3 m or better have recently been launched and are potentially
capable of detecting urban flooding. They include TerraSAR-X,
RADARSAT-2, and the four COSMO-SkyMed satellites.
In an operational context, [7] proposed a hybrid methodology
which combines radiometric thresholding and region growing
as an approach enabling the automated, objective, and reliable
flood extent extraction from SAR images. First results
on moderate- and low-resolution image data indicate that the
proposed method may outperform manual approaches if no
training data are available, even if the parameters associated
with these methods are determined in a non-optimal way. The
results demonstrate the algorithm’s potential for accurately
processing data from different SAR sensors.
Notable examples of research into automatic near realtime
flood detection algorithms using single-polarization highresolution
(greater than a few meters) SAR imagery have
been shown by [8] and [9] on TerraSAR-X data and [10] on
COSMO-SkyMed data. The algorithms by [8] and [9] search
for water as regions of low SAR backscatter using a regiongrowing
iterated segmentation/classification approach, whereas
the technique by [10] is based on a fuzzy logic approach
which integrates theoretical knowledge about the radar return
from inundated areas based on backscattering models, with
simple hydraulic considerations and contextual information.
Both algorithms are very effective at detecting rural floods, but
0196-2892/$31.00 © 2012 IEEE
2418 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 51, NO. 4, APRIL 2013
would require substantial modification to work in urban areas
containing radar shadow and layover.
A semiautomatic algorithm for the detection of floodwater in
urban areas using TerraSAR-X has been developed by [11]. It
uses a SAR simulator [12] in conjunction with LiDAR terrain
data to estimate regions of the image in which water would
not be visible due to shadow or layover caused by buildings
and taller vegetation. Ground will be in radar shadow if it
is hidden from the radar by an adjacent intervening building.
The shadowed area will appear dark and may be misclassified
as water even if it is dry. In contrast, an area of flooded
ground in front of the wall of a building viewed in the range
direction may be allocated to the same range bin as the wall,
causing layover which generally results in a strong return and a
possible misclassification of flooded ground as unflooded. The
algorithm proposed by [11] is aimed at detecting flood extents
for validating an urban flood inundation model in an offline
situation and requires user interaction at a number of stages.
Follow-up work from this was carried out by [13]. Here,
the objective was to build on a number of aspects of the
existing algorithms to develop an automatic near real-time
method for flood detection in urban and rural areas. In
the urban area, 75% of the urban water pixels visible to
TerraSAR-X were correctly detected, though this percentage
reduced somewhat if the urban flood extent visible in the
aerial photos and detected by TerraSAR-X was considered,
because flooded pixels in the shadow/layover areas not visible
to TerraSAR-X then had to be taken into account. Better flood
detection accuracy was achieved in rural areas, with almost
90% of water pixels being correctly detected by TerraSAR-X.
The algorithm assumes that high-resolution LiDAR data are
available for at least the urban regions in the scene, so that a
SAR simulator may be run in conjunction with the LiDAR data
to generate maps of radar shadow and layover in urban areas. It
is therefore limited to urban regions of the globe that have been
mapped using LiDAR.
In an operational flood management perspective, an ideal
flood-mapping system operating in near real time should be
fully automatic, computationally efficient, independent of the
content of local geo-information databases, and, most importantly,
capable of providing accurate and reliable results.
To contribute to the recent developments in highperformance
flood detection algorithms to obtain timely and
more accurate flood warnings, we propose an effective technique
based on image differencing as proposed by [7], which
may compete with existing algorithms in terms of accuracy and
level of automation. For this, we also focus on high-resolution
SAR data for flood detection inside urban areas and use the
TerraSAR-X image of the England summer 2007 floods as
demonstration. Although this is only a single test, and different
results may be obtained for other urban areas where the built
environment is different to the UK case studied here, it does
provide a first demonstration of the potential of the method.
II. METHODOLOGY
Martinis et al. [8] recently highlighted an apparent lack
of traceability and standardization in many SAR-based floodFig.
1. General scheme of the three processing steps of the flood detection
algorithm M2b.
mapping methodologies. This concern has led to the introduction
of two variants of an automated and physically based
SAR-based flood-mapping algorithm [7]. Both variants, which
are termed M1 and M2a, respectively, exploit the statistics
of backscattering coefficients retrieved from SAR to segment
an image into its flooded and non-flooded parts. While M1
only considers a single SAR flood image to extract pixels
corresponding to “open water” via thresholding and region
growing, M2a adds CD with respect to a non-flood reference
image to improve the algorithm’s performance. In this paper,
we introduce an enhanced version of M2a, which we term M2b.
This method addresses some of the shortcomings of M2a that
[7] identified in two representative case studies.
This section provides a detailed overview for all processing
steps of the flood extraction algorithm M2b, together with
the associated parameters defining each process and a list of
differences with respect to the M1 and M2a algorithms previously
introduced. In addition to standard pre-processing steps
commonly involved with Level 1 SAR data, the M2b algorithm
consists of four processing steps (Fig. 1).
A. Statistical Distribution of the “Open Water” Backscatter
The flood extraction algorithm uses as input Level 1 SAR
data that are geocoded, coregistered, and calibrated. The first
step is the estimation of the probability density function (PDF)
of backscattering values associated with “open water.” The
aim of this processing step is the calibration of a theoretical
PDF that optimally fits the empirical distribution of backscatter
values from “open water” inferred from the SAR image. According
to [14], the backscatter variability on a homogeneous
surface is mainly due to speckle and the theoretical PDF that
best describes the distribution of backscatter originating from a
homogeneous surface is a Gamma PDF. Here, we hypothesize
“open water” to be a homogeneous surface, which means that
a potential limitation of the approach, and SAR mapping of
inundated surfaces in general, relates to the possible roughening
GIUSTARINI et al.: CHANGE DETECTION APPROACH TO FLOOD MAPPING 2419
of “open water” caused by emerging vegetation, wind, or rainfall.
Alternative PDFs have been parameterized and tested: the
K-distribution and the RiIG distribution functions (see, e.g.,
[15]). However, the goodness of fit provided by the three PDFs
was found to be almost equivalent, with the Gamma PDF
slightly outperforming the other two functions in this particular
case study. Moreover, the Gamma PDF has only two parameters
(compared to three parameters for the other PDFs) and the
additional advantage of a physically based interpretation for
homogeneous areas with fully developed speckle [14]. The
latter can be considered a reasonable assumption for “open
water.” Consequently, the Gamma PDF was preferred over
other competing PDFs for approximating the distribution of
backscatter values corresponding to “open water”
fσ0
m
(σ0
/k) =
(σ0
− σ0
1)(k−1)
σ0
m−σ0
1
k−1
k
· Γ(k)
e
−
(σ0−σ0
1
)(k−1)
(σ0
m−σ0
1
)
(1)
where k is the shape parameter of the gamma distribution and
σ0
m is the gamma distribution mode. The parameter σ0
1 is the
minimum backscatter value in the SAR image, which needs
to be applied so that the gamma distribution is, therefore, only
computed for positive values.
Two parameters thus need to be optimized to identify the
theoretical gamma function f that best fits the empirical distribution
of backscatter values from “open water” h (i.e., image
histogram). The optimization of the two parameters k and σ0
m
consists in minimizing the root mean squared error (RMSE)
between the image histogram and the gamma distribution, for
backscatter values lower than σ0
thr, with the parameter σ0
thr ≥
σ0
m representing the point where the distributions f and h
start deviating. The optimization is performed with sequentially
increasing values of σ0
m and σ0
thr. For both parameters, the
proposed sampling step is 0.1 dB. The optimization process is
initiated with a first-guess value for σ0
m of −25 dB. For each
tested mode value, sequentially increasing σ0
thr values, higher
than the corresponding tested σ0
m value, are selected. For each
set of σ0
m and σ0
thr values, the parameter k is optimized using
the nonlinear fitting process of [16], i.e., the nonlinear regression
based on the Levenberg–Marquardt algorithm for nonlinear
least squares. The RMSE between the theoretical density
function f and the empirical density distribution h is calculated
for each parameter set and over all backscatter values lower
than σ0
thr. Finally, the parameter set (σ0
m, k, σ0
thr) providing the
lowest RMSE is set as optimal. In case the image histogram
is not bimodal, an appropriate option is available for the user
to manually set a range of plausible values, inside which the
algorithm tests different modes searching for the optimal one.
B. Backscatter Thresholding
The aim of the this step is to extract seeds of “open water”
areas from the flood image, being either individual pixels or regions.
The parameter σ0
thr represents the maximum backscatter
value for which the fit between the theoretical and empirical
PDF is satisfactory. For backscattering values higher than σ0
thr,
the distribution functions f and h start deviating. As a matter
of fact, σ0
thr is considered the maximum backscatter value
for which there is no significant overlap between radiometric
distributions corresponding to water bodies and other land use
types. Since the backscatter values from water surfaces are
comparatively low, this value is used to extract the seeds of
water bodies by selecting the pixels having backscatter values
lower than σ0
thr. This thresholding yields a preliminary flood
inundation map that represents the seed region for a subsequent
region growing process.
Moreover, to be able to map permanent water bodies, the
threshold computed on the flood image, σ0
thr, is also applied on
the reference SAR image to classify seeds of permanent water
bodies. It is worth noting that these seeds include, in addition
to permanent water bodies, other smooth surfaces with a water
surface-like radar response as well as all shadow-affected areas.
The issue related to smooth surfaces will be discussed in more
detail in the following sections.
C. Region Growing
Next, the extracted water bodies, representing the seeds, are
dilated using the region growing approach of [17]. The procedure
iteratively grows the seeds until a given tolerance level is
reached. The sequence of thresholding and region growing only
adds pixels to the seeds that are located in the vicinity of the preliminary
flood extent, thereby limiting the risk of overdetection
in areas distant from the flooded area (i.e., misclassification of
“dry” pixels as “wet”). The tolerance parameter characterizes
the regional homogeneity of the backscattering behavior. The
tolerance criterion adopted here is based on the percentiles of
the theoretical gamma distribution of “open water” pixels.
The iterative procedure incorporates pixels with backscatter
values lower than σ0
rg, corresponding to a given percentile,
RG%, of the theoretical gamma distribution of “water” pixels in
the image. In this paper, we propose a simultaneous calibration,
recently advocated by [7]. The approach optimizes the tolerance
criterion together with the CD parameter introduced in the next
section.
Region growing, with the same threshold value σ0
rg, is also
applied to dilate the seeds of permanent smooth surfaces obtained
from the reference image. The approach provides a mask
of water surface-like radar response areas that is used to limit
the region growing applied on the flood image, thereby preventing
the spreading of flooded areas into permanent smooth areas.
D. CD
Matgen et al. [7] argued that flood maps resulting from region
growing should include all “open water” pixels connected
to the seeds. The region growing should thus extend into the
high percentiles of the gamma distribution. However, the resulting
overdetection needs to be removed by the subsequently
applied CD step. CD thus aims at removing pixels from the
flood extent map that do not correspond to flood water. To do
so, only pixels that significantly change their backscatter values
with respect to their baseline backscatter values are kept in
the flood extent map, while pixels that did not decrease their
backscatter values by a minimum amount are removed. This
2420 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 51, NO. 4, APRIL 2013
Fig. 2. (a1) Flood image (July 25, 2007) and (b1) postflood reference image
(July 22, 2008). Zoom in to the area of interest (city of Tewkesbury) for (a2)
the flood image and (b2) the reference one.
means that the main river channel, which is a permanent water
body, is not any longer an integral part of the flooded area.
The specific parameter of the CD is Δσ0
, defined as the
required minimum change in backscatter between the reference
and the flood image for a pixel being considered as flooded.
To determine the optimal criterion for the required minimum
change in backscatter, an iterative procedure is adopted. As
mentioned earlier, the two parameters σ0
rg and Δσ0
are optimized
through a simultaneous calibration, minimizing the
RMSE computed over the whole range of backscatter values
in the flood image between the theoretical gamma distribution
and the empirical distribution of “open water” pixels. This
means that different threshold values, σ0
rg, which correspond
to different percentiles of the theoretical gamma distribution,
are sequentially selected from an interval of plausible values,
and a corresponding minimum CD parameter Δσ0
is optimized
for each tested threshold value. We consider as plausible values
all values that are greater than σ0
thr and which increase with
a sampling step of 1% up to the value of 99% of the Gamma
PDF and then with an increment of 0.1% up to the value of
99.9% of the Gamma PDF. For every parameter set (σ0
rg, Δσ0
),
the sequence of region growing and CD processes is applied
on the area conditioned by the permanent smooth area mask.
At the end of each iteration, the histogram of “flood water” pixels
is computed. The corresponding empirical PDF is compared
against the initially calibrated theoretical gamma distribution
(1). The parameter set (σ0
rg, Δσ0
) providing the lowest RMSE
value is set as optimal.
To summarize, M2b essentially represents an improved
version of the M2a method introduced by [7]. The two
algorithms both take into account a reference SAR image
and include four inter-related processing steps (i.e., calibration
of gamma distribution function, radiometric thresholding,
region growing, and CD). However, while M2a predefines
the region growing parameter as the 99% percentile of the
“water” backscatter gamma distribution, M2b adds flexibility to
the optimization process by calibrating the tolerance criterion
that, together with the associated CD parameter Δσ0
, minimizes
the RMSE between empirical and theoretical distribution
functions.
This modification implemented in M2b constitutes an important
change as it renders the algorithm fully automated, without
any requirement of manual user inputs. Therefore, the mapping
process is believed to be entirely objective. Another important
improvement is that M2b, unlike M2a, makes use of the reference
image to build a mask of permanent water surface-like
radar response areas. Indeed, to render the algorithm suitable
for urban flood mapping, it is necessary to mask out not only
smooth surfaces like tarmac, paved roads, and parking lots, but
also all regions in shadow-affected areas unseen by the satellite.
In urban areas, the latter are particularly important as they
potentially lead to a significant part of overdetected flooded
areas. This issue will be thoroughly discussed in Section IV-B2.
It should also be noted that method M2a and its enhanced
version M2b both rely on the availability of reference images
acquired from the same orbital track, with the same incidence
angle, polarization, and resolution, and prior to the onset of
flooding. Moreover, the adequate choice of a season-dependent
reference image might help reducing the effects of changes
in vegetation, as argued recently by [18]. With the advent of
relatively new sensors such as TerraSAR-X, it can be difficult
to find an image that satisfies these selection criteria. However,
as image archives are gradually being built up, this should be
less of a problem in the near future. If no reference image is
available, method M1 can be applied.
III. STUDY AREA AND AVAILABLE DATA SET
This section describes the study area, the flooding event and
the available remote-sensing images for testing and evaluating
the proposed automated flood delineation algorithm.
The image data used for this study were acquired for
the approximately 1-in-150-year flood that took place around
Tewkesbury, U.K., in July 2007 [11]. Extreme rainfall intensities
resulted in substantial flooding of urban and rural areas;
about 1000 properties in the town of Tewkesbury were affected
[19]. Tewkesbury lies at the confluence of the River Severn,
flowing in from the northwest, and the River Avon, flowing
in from the northeast. Bankfull discharge is approximately
350 m.s−1
(or 4.5 m in gauged level) at the Saxons Lode
gauging station ∼7 km upstream of Tewkesbury. The summer
2007 event was unusual for the study site in that the majority
of the flow derived from local rainfall. On the 20th July, two
days prior to flood peak, more than 12 cm of rain fell on the
surrounding area. The flood peak of 5.43 m Ordnance Data
Newlyn was measured at Tewkesbury on July 22 with both
rivers exhibiting a more rapid increase in flow than a typical
autumn or winter event that may build over many weeks, with
flows increasing from 100 m.s−1
to > 500 m.s−1
in 57 h,
between the 20th and the 22nd July. The river did not return
to below bankfull until July 31. In the region of interest (red
GIUSTARINI et al.: CHANGE DETECTION APPROACH TO FLOOD MAPPING 2421
TABLE I
CHARACTERISTICS OF THE AVAILABLE IMAGES (STRIPMAP MODE) FOR THE ANALYZED FLOOD EVENT
box in Fig. 2), an area of 1.5 km2
was possibly flooded at the
time of the TerraSAR-X overpass, according to a 2 m resolution
hydraulic model [20].
The majority of the buildings are two storey houses. In the
area of interest, there are some industrial warehouses but no
large-size factories. Overall, the study area is rather representative
of a typical urban landscape in the UK. To demonstrate
the applicability of our proposed SAR image segmentation
algorithm, we define urban area as the zone inside and in the
vicinity of the built-up region of the town of Tewkesbury.
A. TerraSAR-X Images
A unique data set consisting of numerous types of remotely
sensed images over one single event hydrograph were acquired
over the selected study area [20]. From this data set, a stripmap
TerraSAR-X image acquired on July 25, 2007 (at 06:34 GMT,
Wednesday) was selected (see Fig. 2). The image is a multi-look
ground range spatially enhanced scene with 1.5 m pixel spacing
and has a mean incidence angle of 24◦
. Its H/H polarization
mode arguably allows for the best discrimination between a
SAR image’s flooded and non-flooded parts [8]. At the time of
the satellite overpass and image acquisition, there was relatively
low wind speed and no rain [11]. Moreover, no rainfall was
recorded in the 30 h preceding the TerraSAR-X acquisition, as
well as during the satellite overpass itself.
In their flood delineation study, [11] used the single
TerraSAR-X flood image together with airborne scanning laser
altimetry (LiDAR) data. Here, we also consider a dry reference
image which is a postflood image acquired from the same
orbit track and with the same polarization as the flood image.
This way, geometric problems related to coregistration can be
limited, and baseline backscatter values can be inferred. A
single scene having these imaging characteristics and covering
all of the azimuth extent of the target is available in the current
TerraSAR-X image archive. It was acquired on July 22, 2008
(at 06:34 GMT, Tuesday), almost exactly a year after the flood
event had occurred. The flood and non-flood images have both
been acquired in the same month of the year. Hence, it can be
assumed that the state of vegetation is similar in both images.
This is important as decreases in backscatter values between
any two images are caused not only by flooding, but also by
changes in vegetation.
The two images, whose characteristics are listed in Table I,
have been georeferenced and calibrated. These two processes
are important to preserve a backscatter consistency and an
accurate coregistration between the images. Hence, it can
be avoided to erroneously consider as flooding related those
changes in backscattering values that are due to differences in
image acquisition characteristics.
Fig. 3. (a) Flood validation map obtained from high-resolution aerial photography
on July 24, 2007 at 11:30 GMT: The permanent water bodies are also
displayed; (b) comparison between the LISFLOOD computed flood extent at
the time of aerial photographs acquisition (July 24, 2007 at 11:30 GMT) and
TerraSAR-X overpass (July 25, 2007 at 6:34 GMT).
The TerraSAR-X images used in this study have a 1.5 m pixel
spacing and a ground resolution of the order of magnitude of
3 m. This means that each pixel represents a 1.5 × 1.5 m2
area
on the ground and that only individual objects of dimensions
bigger than 3 m can be discriminated in the image. In an
operational context, the reference image would ideally consist
of a preflood satellite acquisition. However, in this particular
case, given the relative novelty of a sensor such as TerraSAR-X,
it was not possible to find a reference image, prior to the onset
of flooding, acquired from the same orbital track and with same
polarization. Therefore, a postflood image was selected.
Particular attention has been given to an adequate coregistration
of the images, as an accurate overlapping is a prerequisite
for detecting flooding-related changes in the backscattering
behavior. The accuracy of the georeferencing is of subpixel
precision. Next, the images have been filtered with a 5 ×
5 Gamma-MAP filter to decrease the speckle contribution. This
2422 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 51, NO. 4, APRIL 2013
filter smoothes out the speckle granularity while preserving
details, such as the contours of buildings and flooded areas [21].
It also impacts the parameterization of the gamma PDF (1) by
reducing the spread of backscattering values associated with
“open water.” The red box area in Fig. 2 presents the area of
interest for the city of Tewkesbury: it refers to a rectangular area
of 1135 × 998 pixels (1.5 m pixel spacing) for a total surface
of ∼3 km2
.
B. Validation Data Set
The validation data set, consisting of very high resolution
0.2 m aerial photographs acquired during the flooding event
in July 2007, enables a comprehensive evaluation of the algorithm’s
performance in terms of SAR-based flood delineation.
An aircraft operated by the Environment Agency of England
and Wales carried out the overflights.
The flood extent was obtained through manual photointerpretation
[Fig. 3(a)]. Taking advantage of existing landuse
maps of the area, permanent water bodies associated with
rivers and canals have been removed from the validation map.
While in general the delineation of flood boundaries from such
high-resolution optical products is relatively straightforward,
it is important to note that the flooding of densely vegetated
and built-up environments can lead to some ambiguities. For
instance, in the case of bare soil fields, the accurate positioning
of the separation line between muddy flood waters and nonflooded
areas is nontrivial. However, for this case study, difficulties
in shoreline delineation were encountered only in a
limited number of locations.
In addition, it is important to bear in mind that the aerial
photographs were acquired on July 24 (at 11:30 GMT) while
the TerraSAR-X image was obtained 19 h later on July 25
(at 06:34 GMT). Although there was no significant decrease
in gauged water levels between the acquisition time of aerial
photographs and the TerraSAR-X overpass [22], this time gap
might be responsible for some discrepancies between the aerial
photography-derived and SAR-derived flooded areas. To estimate
the potential variation of flood extent between the two
acquisition times, simulations with a previously calibrated hydraulic
2 m LISFLOOD-FP flood model [23] have been carried
out both at the acquisition time of the aerial photographs and
at the TerraSAR-X overpass. It is here assumed that the model
provides a satisfactory representation of the time variation of
the flood extent, since an evaluation of the model results showed
that the model predicts water levels with a mean error of less
than 30 cm [20]. The simulations show a reduction of the
flooded area of approximately 5% between the two time steps.
In particular, Fig. 3(b) shows the differences between the two
simulated flood inundation maps. The most notable differences
can be observed on a triangular-shaped field (see middle-right
part in the domain of interest) from which, according to the
model simulations, flood water was drained between the two
overpasses. This location was also problematic in terms of
identifying its flooding status through photo-interpretation, as
explained in more details in the discussion section. These
factors, all unrelated to the processing of the SAR images, need
to be taken into account during the analysis, as all the observed
Fig. 4. Optimization of the parameters of the automated algorithm for M2b
method. (a) RMSE calculated by comparing the empirical image histogram
with several competing Gamma PDFs, obtained with different values for the
mode, σ0
m. (b) RMSE calculated by comparing the optimized Gamma PDF
and the empirical image histogram after region growing and change detection,
for different tested values of the region growing parameter, σ0
rg. (c) Example of
RMSE calculated by comparing the optimized Gamma PDF and the empirical
image histogram obtained, for a given fixed region growing parameter, σ0
rg,
testing several change detection values, Δσ0. (d) Empirical image histogram
and optimized Gamma PDF. (e) Empirical image histogram with the optimized
region growing parameter, σ0
rg, displayed. (f) Example of the effect deriving
from the application of the optimized change detection value, Δσ0, for a given
region growing parameter, σ0
rg.
differences may not be necessarily due to the inability of the
proposed algorithms for accurately extracting the flood extent
from SAR imagery.
IV. RESULTS AND DISCUSSION
This section assesses the classification accuracy obtained
with the fully automated flood detection algorithm M2b and
contrasts its performance with those of the previously introduced
M1 and M2a algorithms. It also provides insights into the
added value of reference images for flood delineation in urban
areas.
A. Extraction of Flooded Areas
The flood extent has been extracted from the TerraSAR-X
image using the three methods M1, M2a, and M2b.
In particular, for method M2b, Fig. 4 illustrates the optimization
of the four parameters: the mode of the “open water”
backscatter gamma PDF, σ0
m, the backscatter threshold, σ0
thr,
the tolerance criterion for the region growing step, σ0
rg, and the
minimum CD value, Δσ0
.
Panel (a) reports the optimization of the mode parameter,
while panel (d) provides the corresponding optimized gamma
PDF in red, together with the histogram of the backscatter
values in the flood image. Panel (d) displays the value of the
second parameter, σ0
thr (i.e., in this case study equal to
−15.5 dB) as the maximum backscatter value for which there
was no overlap between the empirical histogram and the theoretical
gamma PDF. The optimized value is also provided in
GIUSTARINI et al.: CHANGE DETECTION APPROACH TO FLOOD MAPPING 2423
Fig. 5. (a) Backscatter histogram of the reference image with superimposed
threshold value σ0
thr, computed on the flood image; (b) reference mask.
Fig. 5(a), together with the backscatter histogram of pixels in
the reference image. This value is used to derive the reference
mask of permanent water surface-like radar response areas
through thresholding of the reference image [Fig. 5(b)]. From
Fig. 4(d) and Fig. 5(a), it can be observed that for high backscatter
values, there is a systematic noise in the return signal of both
TerraSAR-X images. This is arguably due to the high complexity
of urban topography and its considerable impact on the highresolution
backscattering signal. However, at this stage, this
is only a hypothesis that could not be verified: other inherent
technological reasons related to very high resolution SAR data
could be at the origin of the noisy data. Considering the region
growing step of method M2b, it is important to mention that the
parameters σ0
rg and Δσ0
are optimized together. Therefore, the
subplot in panel (b) of Fig. 4 illustrates the impact that different
σ0
rg parameters (each associated with a corresponding optimal
Δσ0
value) have on the RMSE. Similarly, panel (c) provides
an example of the performance of Δσ0
values for a given σ0
rg
value: Fig. 4(c) refers to the optimal σ0
rg value for the case
study. The backscatter value corresponding to the optimized
σ0
rg value is also displayed in panels (e) and (f). Finally, panel
(f) shows the empirical histogram of flood pixel values before
and after CD: these histograms are computed only with the
pixels inside the SAR-derived flooded area. The reduction of
the distribution tail and the related reduction of overdetection
are indicated by the empirical histogram approaching the theoretical
gamma PDF.
B. Evaluation at City Level (Quantitative Analysis)
1) Overview of Flood Maps: Three flood extent maps were
obtained through the application of the three image processing
algorithms. The corresponding contingency matrices were
computed using the evidence provided by aerial photography.
The binary pattern of flooded and non-flooded pixels was compared
against the reference flood map (in this case, see Fig. 3).
The result is a matrix (or contingency table) of four possible
outcomes. With respect to the reference flooded area, there
are two ways for a remote-sensing-derived flooded area to be
correct (either by correctly representing flooded or non-flooded
pixels) and two ways to be incorrect (either by erroneously
under- or overpredicting the observed inundation extent). The
values of the contingency matrix for all methods are reported
in Table II (and also displayed as contingency maps in Fig. 6)
for a quantitative evaluation of the performances. Moreover,
TABLE II
QUANTITATIVE EVALUATION OF TERRASAR-X DERIVED FLOOD EXTENT
the optimized (and/or fixed) parameter values for the region
growing and the CD are indicated.
From Table II, it can be concluded that in the present case
study, the three algorithms provide very similar performance
levels. When the evaluation is carried out at a regional scale
(i.e., at city level), the differences seem to be marginal. Methods
M2a and M2b slightly outperform method M1 with respect
to the main performance measures provided in Table II. For
example, the total error is lower when CD is applied. While
the underdetection obtained with M2b is slightly higher than
with M1 and M2a, this is compensated with a lower overdetection.
Overall, M2b and M2a perform better than M1, as it
was expected, suggesting that CD with respect to a non-flood
reference image does provide some advantages.
The results do not reflect the added value that we expected
from the methodological improvements of method M2b. This
result is due to the fact that, in this particular case study, the
optimized region growing threshold σ0
rg equals 98%, which is
very close to the pre-defined 99% value that [7] proposed for
M2a. A similar result was obtained in a pretest of the fully
automated algorithm with an ENVISAT ASAR WSM image
available for the same flood. These two case studies on the
one hand suggest that the threshold chosen by [7] is rather
plausible and, on the other hand, validate the applicability of the
procedure to images with different resolution. Here, the tradeoff
involves a controlled growing of the seed region to be able to
limit the overdetection of flooded areas. While the latter can
be partly removed by the subsequent CD, the results indicate
that the reduced overdetection comes at the cost of an increased
underdetection of flooded areas. The results also indicate that
method M2b, which provides an optimal empirical distribution
with respect to the targeted gamma distribution of “open water”
backscatter values, does not necessarily generate a more accurate
flood inundation map than M2a. On a more positive note,
it can be observed that the simultaneous optimization of region
growing threshold σ0
rg and CD parameter Δσ0
, computed by
minimizing the RMSE between empirical and theoretical distribution
functions, nearly led to the maximum value of correctly
detected pixels (81.7% as reported in Table II). Moreover, from
Fig. 7, it can be observed that in this case study, the optimum
parameter set also yields the best performance with respect to
the validation data.
A comparison with the flood extent detected for the same
test case with the semiautomatic procedure of [11] cannot be
carried out in a very meaningful way due to the fact that the
input data sets in the two studies are different. Mason et al. [11]
took advantage of a regional DEM, so that SAR-derived water
ground heights smoothly vary along the river reach, but did
2424 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 51, NO. 4, APRIL 2013
Fig. 6. Contingency map deriving from method: (a) M1, (b) M2a, (c) M2b.
For the sake of clearness in the representation, the displayed maps have been
cleaned by neighborhood analysis in post-processing step.
not make use of a preflood reference image. However, the
comparison of contingency matrices, on a common area of
interest and validation data set, reveals that the M2b method
performance of 81.7% of correctly detected pixels is rather
close to the percentage of 85.4% obtained with the flood
inundation map provided by [11]. This result indicates that
Fig. 7. RMSE values computed for different region growing thresholds (M2b
method) during the optimization process and corresponding performances in
terms of correctly predicted pixels (as flooded and as nonflooded).
topography data could be used more efficiently than preflood
reference images for increasing the accuracy of SAR-derived
flooded areas. However, this assessment needs to be confirmed
in future studies.
To better appreciate the advantages of these methodological
enhancements, it is worth analyzing the PDF of backscattering
values associated with pixels located inside the flood extents,
the latter corresponding either to the validation map obtained
from aerial photographs or the flood extents computed with
the different versions of the image-processing algorithm. The
different PDFs are displayed in Fig. 8. It becomes evident from
the panels that the PDF of “flood water” pixels from highresolution
photos is reasonably close to a gamma distribution,
albeit characterized by a heavy tail end. M1 does enable the
identification of a majority of water pixels but it misses out
the tail of the distribution. M2a yields a better performance,
as it adds more pixels to the flood extent, thereby reducing the
number of underdetected flood pixels. However, there is still
a tendency to slightly overestimate part of the tail of the theoretical
gamma PDF. Because of its enhanced CD procedure,
M2b method is capable of growing the seeds further into the
high percentiles of the gamma distribution, reducing the heavy
tail end and keeping the PDF of detected water pixels closer
to the theoretical one. It is worth noting here that the PDF of
backscattering values inside the area delimited by the highresolution
photos exhibits a particular tail that is missed by all
three versions of the flood detection algorithm. Surprisingly, the
high number of pixels with associated high backscatter values
is not only due to the expected increased backscatter response
from urban structures. In fact, the probability distribution from
pixels located in rural areas exhibits the same heavy tail end as
the one of pixels located in urban areas (Fig. 9).
Further research is needed to get a better understanding of
the particular shape of the PDF of backscatter values. One
possible explanation can be that some of the pixels changed
their status from “flooded” to “nonflooded” between the two
data acquisitions, as already shown in Fig. 3(b).
From this first overview of results, it can be concluded that
the strength of CD is that it allows the region growing to extend
further into the high percentiles of the gamma distribution
GIUSTARINI et al.: CHANGE DETECTION APPROACH TO FLOOD MAPPING 2425
Fig. 8. Backscatter probability density function of water pixels from the highresolution
photographs and water pixels from the method: (a) M1, (b) M2a,
(c) M2b. The histograms refer to the algorithm output, with no post-processing
cleaning step included.
as it efficiently removes part of the resulting overdetection.
More case studies are needed before a generalization of these
findings can be done in a meaningful way. Nevertheless,
these preliminary results corroborate those reported by [7] in
case studies dealing with moderate- and low-resolution SAR
imagery.
On the selected domain (i.e., rectangular area of 1135 ×
998 pixels), the running time of the complete M2b process on
an Intel Core 2Quad CPU, 2.66 GHz and 3.24 GB RAM is less
than 30 min, indicating the appropriateness of the method also
for near real-time applications.
Fig. 9. (a) Main rural and urban areas overlapped on the pixels in the flood
image covered by water according to the high-resolution aerial photographs;
(b) corresponding backscatter probability density functions.
To summarize, it can be observed that the results obtained
with the three versions of the algorithm are rather similar, with
the performance of the fully automated M2b being comprised
between those obtained with the simplified but fully automated
approach M1 and the more subjective one M2a. Moreover,
since the correct classification rate of the proposed methods are
comparable to those obtained by [11], it could be argued that
the main part of the 18% misclassification still remaining can
be imputed to limitations of the SAR imaging technique.
2) The Problem of Unseen Regions: One feature that requires
special attention relates to regions in a SAR image
that cannot be seen by the radar sensor because of its sidelooking
nature. The affected regions are commonly referred
to as “shadow” and “layover.” In the context of urban flood
mapping, “shadow” and “layover” are due to geometric distortions
caused mainly by the presence of buildings. Their
impact on SAR-based flood mapping is twofold. First, flooding
does not impact the radar response from shadow areas and,
consequently, SAR-based detection of flooded shadow areas is
not possible (i.e., problem of underdetecting floods). Second,
the low radar response from shadow regions might erroneously
lead to their classification as “flooded” even in the case they
are not (i.e., problem of overdetecting floods). Here, we assume
that the resulting overdetection might be addressed through
CD since urban shadow areas do not change between two
images acquired with the same imaging characteristics. This is
confirmed by the reduction in overdetection shown in Table II.
2426 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 51, NO. 4, APRIL 2013
Fig. 10. Mask of regions unseen by TerraSAR-X due to shadow and layover,
from [11].
On the other hand, due to double-bounce reflection effects, the
urban layover backscatter could be different in the flood and reference
images. This phenomenon typically occurs in vegetated
areas, where flooding yields an increased backscatter due to the
double bouncing between the flooded ground and branches or
leaves, resulting in a higher return signal in the flood image.
A similar mechanism of multiple reflections between flooded
streets and walls could potentially result in a brighter backscatter
in the urban areas covered by the flood image. Classification
errors are expected to be higher in urban areas than in forested
regions [24], and therefore the layover contribution should be
taken into account when mapping flooded urban areas. In fact,
the inherent underdetection problem can only be addressed
through technological advances (e.g., look angles closer to
nadir) or the use of ancillary data (e.g., topography data).
For the imaging characteristics of the July 25, 2007
TerraSAR-X image, [11] computed a mask of areas affected by
“shadow” and “layover” in the city of Tewkesbury (Fig. 10).
They used the German Aerospace Center (DLR) SAR endto-end
simulator in conjunction with airborne scanning laser
altimetry (LiDAR) data to estimate regions of the image in
which water would not be “visible” to the instrument. In this
paper, we made use of the shadow/layover mask from [11] to
evaluate the risk of misclassifying pixels in areas not “visible”
to the SAR sensor. In the area of interest (red box in relevant
figures), the shadow/layover mask of [11] covers a total area
of ∼1 km2
, which represents a significant percentage (∼39%)
of the total area. However, according to the validation map, the
flooded area not visible to the satellite reduces to 0.25 km2
,
over a total flood extent of 1.22 km2
(see Table III). Moreover,
due to the 24◦
look angle, in this particular case study, the
effect of layover is greater than shadow, as it covers a much
larger flooded area. This is mostly due to the diffuse presence
of hedges along the borders of the different fields in the rural
areas.
As flooding in shadow/layover areas is undetectable for
SAR, the corresponding regions would need to be delineated
a priori and considered as areas with an “unidentifiable status
of flooding.” In fact, even if a flood is correctly classified in a
TABLE III
QUANTITATIVE ANALYSIS OF WATER PIXELS (PIXEL SIZE 1.5 m)
IN THE SHADOW REGIONS [11]
shadow region, this result should be viewed as an error as the
right answer is obtained for the wrong reason. As the objective
of the developed method was to generate a mask of surfaces
that produce a radar signal response similar to that of inundated
areas to constrain the flood extent outside the shadow areas, in
the following, we focus only on the overlap between the obtained
flood extent and the shadow areas derived by [11]. Note
that the shadow mask itself, obtained with the SAR simulator
and the LiDAR data, might contain some degree of uncertainty.
In general, the overlap between the SAR-derived flood extent
and the shadow mask is restricted to the border regions of large
clusters of pixels, which were correctly classified as “flooded.”
The number of such pixels is not significant in comparison
to the total number of extracted flood pixels (see Table III):
furthermore, it can be observed that M2b helps in significantly
reducing the number of pixels classified as “flood water” in
the shadow regions. This is due to the fact that parts of the
shadow-affected areas are included in the mask of permanent
water surface-like radar response areas described earlier. By
considering a reference image acquired from the same orbital
track as the target image, M2b method reduces the risk of
classifying “shadow” areas as “flooded.”
C. Evaluation at Street Level (Qualitative/Thematic Analysis)
The benefits of using a reference image, including the masking
of permanent smooth areas representing in this case study
∼20% of the area of interest, become obvious when looking
at the spatial distribution of errors (Fig. 6). The application of
algorithms M2a and M2b leads to the expected reduction of
misclassified pixels in urban areas. Numerous scattered clusters
of pixels that were initially erroneously classified as “flooded”
could be removed, thereby significantly reducing overdetection.
In Table IV, a thematic analysis with a special focus on urban
features complements the quantitative analysis presented earlier
(Table II).
The objective is to understand the advantages and limitations
of the three variants of the SAR-based flood delineation
algorithms for correctly identifying flooding in urban areas.
From the results shown in Table IV, it can be observed that in
spite of the high-resolution SAR imagery used in this study,
the detection of flooding in built-up environments remains a
very challenging task. All algorithms struggle to recognize
the flooding status of many small-scale features that might
be crucial, as their state of flooding could mean significant
interruptions of everyday life. However, overall, the enhanced
algorithm M2b performs best, with a slightly reduced number
of misclassified areas. In particular, M2b enables the a priori
GIUSTARINI et al.: CHANGE DETECTION APPROACH TO FLOOD MAPPING 2427
TABLE IV
IMPROVEMENT DERIVING FROM THE USE OF A REFERENCE IMAGE: COMPARISON OF THE DIFFERENT METHODS FOR SOME REGIONS, WITH A SPECIAL
EMPHASIS ON URBAN FEATURES. SEE FIG. 6 FOR THE LOCATION OF STREETS, CROSSINGS, AND URBAN/RURAL AREAS
(THE CORRECTLY CLASSIFIED REGIONS OF EACH METHOD ARE IN BOLD FONT)
delineation of areas characterized by specular-like reflections
(i.e., areas with permanent water surface-like radar responses).
This is helpful given that smooth areas (e.g., R2 & R3) tend
to be systematically classified as flooded by M1 and, to a
lesser extent, by M2a. On the other hand, more open areas,
such as the main roads R12 and R13, are correctly classified
by all three methods. It is worth mentioning that, despite
these somewhat encouraging results, M2b fails to correctly
delineate flooding in many densely vegetated and built-up
environments.
These errors will be analyzed in more detail in the following
sections. In this analysis, we will consider ancillary data (e.g.,
land use map, oral communications from local experts) to better
understand the reasons that are at the origin of the remaining
misclassifications. Moreover, the mask of regions unseen by
the satellite, i.e., shadow and layover, has also been taken into
account for error detection at street level.
1) The Problem of Overdetection: As it can be seen from
the urban flood maps presented in Fig. 6, both over- and underdetection
are reduced as a result of applying the M2b algorithm
rather than its predecessors. This is particularly evident in the
case of the large shopping mall labeled R3. Due to the flatness
of its roof and resulting specular reflection, it was erroneously
classified as flooded by M1, while M2a and M2b correctly
excluded it from the flooded area. This emblematic example
best illustrates the potential added value of reference images, as
they enable the a priori identification of the majority of smooth
areas.
Similarly, other wide flat regions, such as parking lots and
airfields, are recognizable in the reference image. For instance,
the region labeled R1 corresponds to a large parking
lot composed of three parts. M2b completely removes one of
them from the flood extent map, while the two other parts
are significantly reduced in size. The sub-optimal performance
of M2b is arguably due to a difference in the number and
placement of vehicles at the time of the two satellite overpasses.
In very high resolution SAR imagery the presence or not of
an object like a car inevitably impacts the radar response. This
necessarily influences the capability of the M2b algorithm to
reliably identify areas of smooth tarmac and unfortunately may
not be resolvable at all, for obvious reasons.
The region labeled R2, an area both flat and made of tarmac
but not used as a parking space, shows the capability of
M2b to avoid the typical misclassifications of smooth areas as
flooded.
Finally, the thematic analysis confirms the algorithm’s ability
for identifying permanent water bodies. The permanently
flooded bed of the River Avon and some adjacent boat marinas
are removed from the flood extent map when taking into account
the reference image: this becomes evident when looking
at the areas of overdetection in the panels (b) and (c) of Fig. 6.
Clearly, this result is not achievable with a single flood image,
as M1 would invariably classify permanent water bodies as
flooded (see panel (a) of Fig. 6).
Despite the previously mentioned ability of M2b to detect
areas with permanent low backscatter values, there are still
some shadow-affected areas that are erroneously classified as
“flooded.” A typical example is a large inclined rooftop in
region R9. This is classified as “flooded” by all three methods
due to the fact that one side is not “visible” to the SAR
sensor. Other examples of this behavior can be found on various
inclined rooftops in the R7 region.
2428 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 51, NO. 4, APRIL 2013
To summarize, some risk of overdetecting flooded areas in
built-up environments inevitably remains. Non-flooded areas
that appear smooth and water-surface like at radar wavelengths
as well as areas unseen by the satellite because of the sidelooking
nature of SAR systematically produce very low signal
returns and are not easily distinguishable from flooded areas.
The results of this study suggest that taking into account the
baseline backscatter values from “dry” reference images partly
addresses the problem. Wide, open areas of tarmac or concrete
(roads, parking lots, airfields, etc.) can be identified and removed
from the final flood map (or, alternatively, categorized
as areas impossible to classify), while the situation is more
problematic with shadow areas. To check the plausibility of
both types of regions to be flooded, we expect that the use
of high-resolution high-precision DEM data may be helpful.
More research on the integration of additional data sources into
the image-processing algorithm is needed for this to provide
significant advantages.
2) The Problem of Underdetection: With respect to the
problem of underdetecting the true flood extent using SAR
observations, the results shown in Table II indicate that method
M2b leads to a decrease in performance. In fact, the percentage
of underdetected flood pixels rises from 15.6% obtained with
the initial M2a method to its M2b-related value of 16.2%. The
increase in underdetection between M2a and M2b is mainly
due to the fact that the optimized value of RG% is lower than
the a priori one. However, it has to be underlined that this
type of error is generally to be found on the edge of inundated
fields or in the vicinity of the main riverbed with tall vegetation
surrounding the areas. While the algorithm accurately retrieves
most of the flooded areas in wide, open areas, it can be observed
that it systematically fails to retrieve flooding under the
vegetation canopy. These errors are not related to the imageprocessing
algorithm; rather they are due to the fact that with
X-band radar systems volume scattering originating from the
vegetation canopy causes increased signal return. Furthermore,
as already mentioned in Section III-B, it cannot be ruled out
that the validation flood extent itself is affected by a slight
overestimation, as it was acquired closer to peak discharge than
the satellite images. This could also at least partly explain the
underdetection documented in the contingency matrix. Also,
the uncertainties in the delineation of the flood validation extent
from aerial photography are expected to have some marginal
effect.
For example, an important area of apparent underdetection
is the triangular shaped field labeled R15. However, due to the
time difference between the acquisitions of aerial photographs
satellite imagery, it is likely that most of the floodwater was
drained from the field in the 19 h preceding the TerraSAR-X
acquisition. This hypothesis is confirmed by hydraulic model
simulations [see Fig. 3(b)].
The roughening of water surfaces due to wind is another
inherent and potentially significant limitation of the algorithm
proposed in this study. When there are regular waves on the
surface of the water, Bragg resonance can result in very high
signal returns [25]. The misclassification of the area labeled
R18 as non-flooded represents a typical example. From the air
photos and model simulations, there can be no doubt about the
flooding of the area. However, waves are clearly identifiable
on the standing water. This renders accurate flood detection
extremely difficult (if not impossible), as it violates the algorithm’s
main underlying assumption of flooded areas behaving
as specular reflectors.
V. CONCLUSION AND PERSPECTIVES
This study proposes a promising methodology that is shown
to be capable of providing satisfactory results in mapping, in
a completely unsupervised way, flood extent in a challenging
case study, such as an urban flooding.
A. New Findings and Conclusion
The proposed algorithm is based on the calibration of a
PDF on the backscatter values associated with “open water.”
Next, a sequence of optimized backscatter thresholding, region
growing, and CD is applied on the flood image and a preor
postflood reference image acquired with the same imaging
characteristics. Since no manual (and subjective) input is
required from the end user, the algorithm enables automated,
objective, and repeatable flood detection. It operates with minimum
data requirements, considering as input data a flood image
and a reference image acquired before or after the flooding, also
maintaining the option of functioning with only a crisis image.
The algorithm is efficient in both the two fully automated
versions (M1 and M2b), as the processing time on an Intel(R)
Core (TM) 2Quad CPU, 2.66 Ghz, and 3.24 GB RAM for
a study area of 1135x998 pixels is less than 30 min. With a
classification accuracy of around 82%, the algorithm yielded
satisfactory results with respect to aerial photography-derived
flooded areas in an urban case study. Since the classification
accuracies of the proposed methods are comparable to those
obtained by [11], it could be argued that the main part of
the 18% of the remaining misclassification can be imputed to
limitations of the SAR imaging technique.
The difficulty of detecting flooded areas in a built-up environment
has been partially addressed by a CD approach
that makes use of pre- or postflood reference images available
in the data archives of satellite data providers. In particular,
the shadow effect stemming from man-made structures can
be taken into account through a mask of permanent water
surface-like radar response areas. This approach overcomes
the need of a high-resolution DEM and a SAR simulator for
determining shadow regions that are not visible to the satellite.
On the other hand, it requires a reference image with the same
imaging characteristics as the flood image. While the number
of suitable candidate images can be very limited in case of
relatively new satellites, such as TerraSAR-X, it is important
to note that image archives are gradually being built up, which
will progressively increase the likelihood of finding adequate
reference images in the online archives.
B. Future Research
Some further improvements are still necessary before the
deployment of a fully automated SAR-based flood delineation
GIUSTARINI et al.: CHANGE DETECTION APPROACH TO FLOOD MAPPING 2429
algorithm operating in a near-real time can be envisaged. Our
results indicate that in spite of the high-resolution SAR imagery
used in this study, the detection of flooding in built-up environments
remains a very challenging task. To further improve
the method, we aim at taking advantage of topographic and
land use data, which are now becoming more readily available
at global scale, albeit with variable accuracy and resolution.
We hypothesize that such ancillary data will help reduce the
elevation curvature along the flood edges as argued by [11] and
identify parts of the underdetection caused by emerging objects
such as trees or buildings. In addition, future research should
investigate the usefulness of alternative distribution functions
for optimally fitting the distribution of backscatter values corresponding
to “open water” over different study areas.
We consider this study to be timely because there is a
clear need for rapidly acquiring, processing, and distributing
hydrology-related information derived from SAR imagery. In a
crisis management context, where the situation on the ground
can change very fast, data are more valuable if available shortly
after the acquisition. The lag time between satellite acquisition
and availability of information for efficient data assimilation
can be variable from hours, in case of basins with small
contributing areas and low response times, to days for much
larger catchments. In fact, for near real-time applications in
hydrology, where flood extent data is systematically assimilated
into hydrologic-hydraulic models, the value of remote-sensing
data is much higher if rapidly accessible [26], [27].
APPENDIX
OBTAINING THE CODE
The flood-mapping code here described is a science tool
developed by a team of researchers, and we are happy to
provide a copy of it for non-commercial studies. The code is
reasonably well documented and has been tested in different
case studies; however, it has not been through the same quality
control procedure you would expect for a commercial software
package. Using the code also requires some basic computing
expertise. If you would like to obtain a copy of the code or
are interested in collaborative research, then please contact
us at one of the following addresses: giustari@lippmann.lu,
hostache@lipmman.lu, matgen@lippmann.lu
ACKNOWLEDGMENT
The authors wish to thank the Environment Agency of
England and Wales for the aerial photography of the
Tewkesbury flood and the UK Meteorological Office for granting
access to the NERC BADC for the rainfall records used in
the analysis. The TerraSAR-X flood image was received under
DLR SSS project HYD0363.
REFERENCES
[1] D. E. Alsdorf, E. Rodriguez, and D. P. L. Lettenmaier, “Measuring surface
water from space,” Rev. Geophys., vol. 45, p. RG2 002, May 2007.
[2] W. A. Marcus and M. A. Fonstad, “Optical remote mapping of rivers
at sub-meter resolutions and watershed extents,” Earth Surf. Process.
Landforms, vol. 33, no. 1, pp. 4–24, Jan. 2008.
[3] G. Schumann, P. D. Bates, M. S. Horritt, P. Matgen, and
F. Pappenberger, “Progress in integration of remote sensing-derived
flood extent and stage data and hydraulic models,” Rev. Geophys., vol. 47,
p. RG4 001, Nov. 2009.
[4] G. Nico, M. Pappalepore, G. Pasquariello, S. Refice, and
S. Samarelli, “Comparison of SAR amplitude vs. coherence flood
detection methods—A GIS application,” Int. J. Remote Sens., vol. 21,
no. 8, pp. 1619–1631, 2000.
[5] M. Chini, L. Pulvirenti, and N. Pierdicca, “Analysis and interpretation
of the COSMO-SkyMed observations of the 2011 Japan tsunami,” IEEE
Geosci. Remote Sens. Lett., vol. 9, no. 3, pp. 467–471, May 2012.
[6] N. Longbotham, F. Pacifici, T. Glenn, A. Zare, M. Volpi, D. Tuia,
E. Christophe, J. Michel, J. Inglada, J. Chanussot, and Q. Du,
“Multi-modal change detection, application to the detection of flooded
areas: Outcome of the 2009-2010 data fusion contest,” IEEE J. Sel.
Topics Appl. Earth Observ. Remote Sens., vol. 5, no. 1, pp. 331–342,
Feb. 2012.
[7] P. Matgen, R. Hostache, G. Schumann, L. Pfister, L. Hoffmann, and
H. H. G. Savenije, “Towards an automatic SAR-based flood monitoring
system. lessons learned from two case studies,” Phys. Chem. Earth,
vol. 36, no. 7/8, pp. 241–252, 2011.
[8] S. Martinis, A. Twele, and S. Voigt, “Towards operational near real-time
flood detection using a split-based automatic thresholding procedure on
high resolution TerraSAR-X data,” Nat. Hazards Earth Syst. Sci., vol. 9,
no. 2, pp. 303–314, Mar. 2009.
[9] S. Martinis, A. Twele, and S. Voigt, “Unsupervised extraction of floodinduced
backscatter changes in SAR data using Markov image modeling
on irregular graphs,” IEEE Trans. Geosci. Remote Sens., vol. 49, no. 1,
pp. 251–263, Jan. 2011.
[10] L. Pulvirenti, N. Pierdicca, M. Chini, and L. Guerriero, “An algorithm
for operational flood mapping from synthetic aperture radar (SAR)
data using fuzzy logic,” Nat. Hazards Earth Syst. Sci., vol. 11, no. 2,
pp. 529–540, Feb. 2011.
[11] D. C. Mason, R. Speck, B. Devereux, G. Schumann, J. Neal, and P. D.
Bates, “Flood detection in urban areas using TerraSAR-X,” IEEE Trans.
Geosci. Remote Sens., vol. 48, no. 2, pp. 882–894, Feb. 2010.
[12] R. Speck, P. Turchi, and H. Süß, “An end-to-end simulator for high resolution
spaceborne SAR systems,” in Proc. SPIE Defense Security, 2007,
vol. 6568, p. 656 80H.
[13] D. C. Mason, I. J. Davenport, J. C. Neal, G. J.-P. Schumann, and
P. D. Bates, “Near real-time flood detection in urban and rural areas using
high resolution synthetic aperture radar images,” IEEE Trans. Geosci.
Remote Sens., vol. 50, no. 8, pp. 3041–3052, Aug. 2012.
[14] F. Ulaby, R. Moore, and A. Fung, Microwave Remote Sensing, Active and
Passive: Volume Scattering and Emission Theory, Advances System and
Applications. Norwood, MA: Artech House, 1986.
[15] T. Eltoft, “The Rician inverse Gaussian distribution: A new model for nonRayleigh
signal amplitude statistic,” IEEE Trans. Image Process., vol. 14,
no. 11, pp. 1722–1735, Nov. 2005.
[16] G. A. F. Seber and C. J. Wild, Nonlinear Regression. Berlin, Germany:
Wiley, 2003.
[17] R. M. Haralick and L. G. Shapiro, “Image segmentation techniques,”
Comput. Vis. Graph. Image Process., vol. 29, no. 1, pp. 100–132,
Jan. 1985.
[18] R. Hostache, P. Matgen, and W. Wagner, “Change detection approaches
to flood extent mapping: How to select the best pre-flood reference image
from on-line archives?” J. Appl. Earth Observation Geoinf., vol. 19,
pp. 205–213, 2012.
[19] E. Agency, Review of 2007 summer floods, Environment Agency,
Bristol, U.K. [Online]. Available: http://publications.environment-agency.
gov.uk/PDF/GEHO1107BNMI-E-E.pdf
[20] G. J.-P. Schumann, J. C. Neal, D. C. Mason, and P. D. Bates, “The
accuracy of sequential aerial photography and SAR data for observing
urban flood dynamics, a case study of the UK summer 2007
floods,” Remote Sens. Environ., vol. 115, no. 10, pp. 2536–2546,
Oct. 2011.
[21] A. Lopès, E. Nezry, R. Touzi, and H. Laur, “Structure detection and
statistical adaptive speckle filtering in SAR images,” Int. J. Remote Sens.,
vol. 14, no. 9, pp. 1735–1758, 1993.
[22] H. Zwenzner and S. Voigt, “Improved estimation of flood parameters
by combining space based SAR data with very high resolution digital
elevation data,” Hydrol. Earth Syst. Sci., vol. 13, no. 5, pp. 567–576,
May 2009.
[23] J. C. Neal, G. Schumann, T. Fewtrell, M. Budimir, P. D. Bates, and
D. C. Mason, “Evaluating a new LISFLOOD-FP formulation with data
from the summer 2007 flood in Tewkesbury, UK,” J. Flood Risk Manage.,
vol. 4, no. 2, pp. 88–95, Jun. 2011.
2430 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 51, NO. 4, APRIL 2013
[24] E. Malnes, S. Solbo, I. Lauknes, G. J. Evertsen, T. A. Tllefsen, I. Solheim,
and M. Indregard, “Floodman-global near real-time flood monitoring for
hydrological users,” in Proc. Int. Conf. Innov., Adv. Implement. Flood
Forecast. Technol., 2005, vol. 4, pp. 1–9.
[25] D. O’Grady, M. Leblanc, and D. Gillieson, “Use of ENVISAT ASAR
global monitoring mode to complement optical data in the mapping of
rapid broad-scale flooding in Pakistan,” Hydrol. Earth Syst. Sci., vol. 15,
no. 11, pp. 3475–3494, Nov. 2011.
[26] P. Matgen, M. Montanari, R. Hostache, L. Pfister, L. Hoffmann, D. Plaza,
V. R. N. Pauwels, G. J. M. D. Lannoy, R. D. Keyser, and H. H. G. Savenije,
“Towards the sequential assimilation of SAR-derived water stages into
hydraulic models using the particle filter: Proof of concept,” Hydrol. Earth
Syst. Sci., vol. 14, no. 9, pp. 1773–1785, Sep. 2010.
[27] L. Giustarini, P. Matgen, R. Hostache, D. Plaza, V. R. N. Pauwels, G. J. M.
D. Lannoy, R. D. Keyser, L. Pfister, L. Hoffmann, and H. H. G. Savenije,
“Assimilating SAR-derived water level data into a flood model: A case
study,” Hydrol. Earth Syst. Sci., vol. 15, no. 7, pp. 2349–2365, Jul. 2011.
Laura Giustarini received the B.Sc. and M.Sc. degrees
in environmental engineering from the University
of Perugia, Perugia, Italy, in 2003 and 2006,
respectively.
She has worked for the National Research
Council, Research Institute for Geo-Hydrologicl
Protection, Perugia, Italy. Since 2010, she has been
with the Department of Environment and AgroBiotechnologies,
Public Research Center - Gabriel
Lippman, Belvaux, Luxembourg. Her current research
interests are the integration, through data assimilation
techniques, of radar remote sensing into coupled hydrologic and
hydraulic models for surface water management.
Renaud Hostache received the Engineering degree
in hydraulics and environmental sciences from
Grenoble Institute of Technology-National School
of Hydraulics and Mechanism of Grenoble, the
M.Sc. degree in mechanics of geophysical media
and environment from the Joseph Fourier University,
Grenoble, France, in 2003, and the Ph.D. degree in
water sciences from the National School of Agriculture,
Water, and Forest Engineering, Montepellier,
in 2006. During the Ph.D. work at the Cemagref,
Montepellier, France, he investigated the fine 3-D
characterization of flood hazard, through satellite imagery and its integration
in flood inundation models.
Since 2007, he has been with the Department of Environment and
Agro-Biotechnologies, Public Research Center - Gabriel Lippman, Belvaux,
Luxembourg. His research interests are focused on the integration of remotesensing
observations in hydraulic models and on the hydrologic and hydraulic
model uncertainty characterization and reduction.
Patrick Matgen received the M.Sc. degree in environmental
engineering from the Ecole Polytechnique
Fédérale de Lausanne, Lausanne, Switzerland. In
2011, he was awarded the Ph.D degree from the
water resources section of the Technical University
in Delft, The Netherlands, for his dissertation on
the retrieval of surface and subsurface water from
microwave remote-sensing observations and the integration
of the data with flood prediction systems.
Currently, he holds the position of Project
Leader responsible for microwave remote-sensing
and hydrologic-hydraulic modeling projects at the Centre de Recherche Public
- Gabriel Lippmann (Grand Duchy of Luxembourg).
Guy J.-P. Schumann (S’06–A’08–M’09) received
the M.A. degree in geography and environmental
science and the M.Sc. degree in remote-sensing image
processing and applications from the University
of Dundee, Dundee, U.K., in 2003 and 2005,
respectively, and the Ph.D. degree in collaboration
with the Public Research Centre - Gabriel Lippmann,
Belvaux, Luxembourg, in 2008.
From August 2008 to March 2011, he was a Fulltime
Lecturer and Research Fellow in hydrology at
the School of Geographical Sciences, University of
Bristol, Bristol, UK: he currently is an Honorary Research Fellow at the same
university. His research interests include flood hydrology and hydraulics and
remote sensing in hydrology, in particular, radar remote sensing for flood
hydrology and management.
Paul D. Bates received the Ph.D. degree from the
University of Bristol, Bristol, U.K., in 1993, with
support from a Natural Environmental Research
Council studentship.
Subsequently, he has worked with the University
of Bristol as a Postdoctoral Researcher and Lecturer
and has been a Full Professor since 2003. Currently,
he is the Director of the Hydrology Research Group
with the School of Geographical Sciences, Bristol.
He has been a Visiting Scientist with the Laboratoire
National D’Hydraulique, Princeton University, Paris,
and the European Union Joint Research Centre, Ispra, Italy. His research
concerns the development and analysis of numerical models for predicting river
flood flows, principally using data derived from remote-sensing sources. He has
specific interests in spatial prediction, risk, and uncertainty.
Prof. Bates is the Editor-in-Chief of the International Journal of River Basin
Management.
David C. Mason received the B.Sc. and Ph.D. degrees
in physics from the University of London,
London, U.K., in 1963 and 1968, respectively.
He has worked for the U.K. Medical Research
Council and Plessey Electronic Systems Research.
Since 1984, he has been with the Natural Environment
Research Council Environmental Systems Science
Center, University of Reading, Reading, U.K.,
carrying out research on the automated extraction of
information from remotely sensed data and linking
these data to environmental models. His current interests
include using remotely sensed data for validation and parameterization
of river flood models and assimilation into coastal orphodynamic models.