This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING 1

Multi-temporal Synthetic Aperture Radar Metrics

Applied to Map Open Water Bodies
Maurizio Santoro, Member, IEEE, and Urs Wegmüller, Senior Member, IEEE

Abstract—Multi-temporal synthetic aperture radar (SAR) parameters, herewith referred to as multi-temporal metrics, with
metrics are assessed to map open water bodies. High temporal acknowledged potential in thematic mapping [10][11].
variability and low minimum value in a time series of Envisat The major limitation of single SAR backscatter images to
Advanced SAR (ASAR) Wide Swath Mode (WSM) backscatter
measurements characterize open water bodies with respect to map water bodies relies in the dependence of the backscattered
other land cover types. Confusion occurs in the case of steep signal upon the surface conditions of the water body. Thresh-
terrain (slope angle > 10°), less than 10 backscatter observations olding approaches or supervised approaches applied to a single
and for mixed pixels with a water fraction. The behavior of the image were sufficient to detect and delineate lakes and rivers in
two SAR multi-temporal metrics is consistent at six study areas
in Europe and Central Siberia. A simple thresholding algorithm
C- and X-band co-polarized data as long as the backscatter was
applied to the multi-temporal SAR metrics to map open water overall low with respect to other land surfaces [12]–[17]. Sev-
bodies performs with overall accuracies above 90% in the case of eral authors reported false detections of water as land in the case
pure pixels of water or land. The accuracy decreases when mixed of rugged water surface [12]–[15], which could be compensated
pixels are accounted for in the reference dataset and for increasing for to a certain extent by using active contour methods [12],
land fraction in the reference samples. An overall accuracy of
approximately 80% was obtained for a 50% threshold of the water [18]. The generation of a global Water Indication Mask (WAM)
fraction. Omissions of water areas occur mostly along shorelines. from X-band TanDEM-X/TerraSAR-X image pairs [15], [18],
Specific conditions of the land surface can distort the minimum, [19] profits from a combination of classifications based on SAR
causing commission in the water class. The use of a low order rank amplitude and interferometric SAR coherence using individual
or percentile instead of the lowest backscatter value can reduce
such commission error.
threshold-based approaches on each observable [19]. Classifica-
tion accuracy reported in terms of correctness and completeness
Index Terms—Envisat ASAR, minimum backscatter, SAR was between 51% and 72%, and 60% and 81%, respectively,
backscatter, temporal variability, water bodies, wide swath mode.
for three latitudinal transects. Slightly higher accuracy was ob-
tained when using coherence data only [19].
I. INTRODUCTION Multi-temporal observations were used to understand and
quantify dynamics of water bodies [5], [8], [9], [13]; a gen-

S PACEBORNE Synthetic Aperture Radar (SAR) data are

available from a number of satellites operating at different
wavelengths, with multi-mode image configurations and are op-
eral conclusion was that the temporal sampling even in the
case of very frequent observations as in the case of Envisat
ASAR ScanSAR images was not optimal to track dynamics
erated with different acquisition strategies. Typically, the acqui- in a sufficiently detailed manner. Multi-temporal observations
sition of high resolution SAR systems (1–20 m) targets specific furthermore revealed changes of the SAR backscatter over
areas. Sensors with acquisition modes at moderate resolution lakes and rivers covered with ice [7], [20]. For bottom-fast ice,
(100–1,000 m) are instead operated to acquire data on a global low backscatter was observed; in the case of liquid water under
scale in a repeated manner. In view of generating estimates of the ice, the SAR backscatter level depended on the composition
a land surface parameter for large areas, moderate resolution of the ice layer [21]–[24].
image data products become the only practical alternative if a The objective of this paper is to investigate the properties of
mapping solution based on SAR data is explored. The avail- metrics derived from multi-temporal SAR data and demonstrate
ability of repeated acquisitions is of advantage since multi-tem- their usefulness in the context of land surface characterization
poral observations allow reduction of speckle noise [1], detec- with particular regard to the detection of open water bodies. Ul-
tion of trends in land surface parameters such as soil moisture timately, the goal was to assess the potential of moderate res-
[2], wetlands [3]–[5], cropland [6] and water bodies [7]–[9]. In olution SAR data to complement classification of water bodies
addition, multi-temporal data allow the generation of additional in global land cover products [25]. To this scope, we aimed at
setting up and validating a water body mapping algorithm that
could be straightforward and robust at the same time.
Manuscript received May 03, 2013; revised July 04, 2013; accepted October The SAR dataset consisted of images of the radar backscat-
29, 2013. This work was supported by the European Space Agency within the tered intensity acquired by the Envisat Advanced SAR (ASAR)
Climate Change Initiative Land Cover (CCI-LC) project.
M. Santoro and U. Wegmüller are with GAMMA Remote Sensing AG, instrument. To assess the consistency of the multi-temporal
CH-3073 Gümligen, Switzerland (e-mail: [email protected]; weg- metrics and the robustness of the water body mapping ap-
[email protected]). proach from the SAR data here considered, investigations
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. were undertaken at several study areas characterized by dif-
Digital Object Identifier 10.1109/JSTARS.2013.2289301 ferent typologies of land cover and water bodies, terrain

1939-1404 © 2013 IEEE

 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

TABLE I
LIST OF STUDY AREAS. CORNER COORDINATES REFER TO TOP-LEFT (TL) AND BOTTOM-RIGHT (BR)
COORDINATES EXPRESSED IN THE FORM OF LATITUDE AND LONGITUDE

conditions and seasonality. Section II provides an overview

of the study areas. Section III describes the SAR datasets
and the reference datasets. Section IV presents the signature
analysis of multi-temporal metrics for different land cover
classes. Section V describes the approach developed to map
water bodies using the multi-temporal SAR metrics found
to be most suitable for this application. Classification results
and agreement statistics with respect to the reference datasets
are discussed in Section VI. Finally, a set of conclusions are
presented in Section VII.
Fig. 1. Histograms of number of ASAR observations per pixel for each study
area. Number of observations have been binned in classes of 10 observations
II. STUDY AREAS each.

Multiple sites were considered in order to obtain broad

understanding of SAR metrics potentially suitable for mapping
then 30 days until April 2012). A WSM image was character-
water bodies and to come up with a robust mapping algorithm,
ized by look angles spanning between 18 (near range) and 42
capable to withstand effects of seasonal conditions on the SAR
(far range). The data were acquired in single-polarization mode,
backscatter and different land cover composition. Six study
either Horizontal-Horizontal (HH) or Vertical-Vertical (VV).
areas were selected (see Table I) following these requirements:
For this study, a one-year dataset of all Envisat ASAR WSM
(i) diverse landscape in terms of land cover, water bodies and
images over each of the study areas was considered. It was
topography;
assumed that the number of backscatter observations collected
(ii) diverse seasonal and environmental conditions;
within one year would have been sufficient to provide reliable
(iii) availability of large number of SAR backscatter measure-
multi-temporal SAR metrics while there would have been
ments within a one-year period;
only minor large scale changes of the landscape. To avoid
(iv) availability of reliable datasets to be used as reference.
large differences in time with the date of the reference datasets
While this selection captured a wide diversity of land cover
(Section III-B), the SAR dataset consisted of all images ac-
types, we did not explicitly consider areas characterized by tem-
quired in 2005. Table II presents an overview of the number of
poral dynamics (e.g., due to inundation) because beyond this
images covering each study area. Images were available from
first evaluation of multi-temporal SAR metrics in water body
ascending and descending orbits and were acquired mostly in
characterization. The size of the study areas ranged between
VV polarization. Fig. 1 shows the histogram of the number
km and km (Table I). In total, detection
of ASAR observations per pixel. For each site, the smallest
of water bodies with multitemporal SAR metrics was tested for
number of observations were located in correspondence of the
more than km .
edges.
The ASAR WSM data consisted of images of the backscat-
III. DATASETS tered intensity (Level 1P products) with a pixel size of
m m, whereas the spatial resolution is approximately
A. SAR Dataset
120 m both in range and in azimuth direction [27]. The SAR
The SAR dataset consisted of images acquired by Envisat images were first calibrated using factors provided by ESA in
ASAR in the Wide Swath Mode (WSM). The ASAR instrument the image data file. A buffer zone along the perimeter of an
operated at C-band (wavelength of 5.7 cm) with a multi-mode image was removed because of occasional processing artifacts.
configuration [26]. Because of the large swath of the WSM Multi-looking, i.e., spatial averaging, in a window was
mode (approximately 400 km), any point on the ground could applied to obtain images with a pixel size of 300 m m.
be observed several times throughout the repeat-pass interval of This resolution was selected to be in line with the resolution of
the satellite (35 days between January 2002 and October 2010, global land cover mapping efforts [25], [28]. However, a
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SANTORO AND WEGMÜLLER: MULTI-TEMPORAL SYNTHETIC APERTURE RADAR METRICS APPLIED TO MAP OPEN WATER BODIES 3

TABLE II To quantify the uncertainty of the backscatter measurements,

NUMBER OF ENVISAT ASAR WIDE SWATH IMAGES FOR EACH OF THE STUDY the Equivalent Number of Looks (ENL) [34] was estimated. The
AREAS. EACH IMAGE CORRESPONDS TO A 400-KM-WIDE IMAGE TAILORED
IN LENGTH BY ESA TO COVER THE STUDY AREA. FOR THE NUMBER OF ENL was computed in polygons that visually showed small spa-
OBSERVATIONS PER PIXEL, REFER TO FIG. 1. tial variability of the backscatter. To understand the impact of
the number of observations available per pixel on the ENL esti-
mates, polygons were located in areas including at least 20 ob-
servations of the backscatter. The proportion of pixels with less
than 20 backscatter observations was small; such pixels were
mostly located at the edge of a study area.
The ENL of Envisat ASAR WSM Level 1P images was re-
ported to be approximately 15 [27]. As a result of multi-looking,
the ENL of the backscatter images with a pixel size of 300 m was
mostly between 40 and 60 (corresponding to 0.6–0.7 dB uncer-
tainty of the backscatter). This result is explained by considering
that Level 1P data is oversampled by factors close to 2 in range
and in azimuth; therefore, multi-looking in a window cor-
responds to averaging over approximately 4 independent looks.
multi-looked version of the original dataset with an output pixel After speckle filtering, the ENL was mostly between 130 and
size of m m was also generated for the study area of 170, corresponding to a backscatter uncertainty of 0.3–0.4 dB.
the Netherlands to investigate the impact of spatial resolution The ENL estimates did not show dependence upon the number
on water body mapping. of observations per pixel. Based then on the results of the sig-
For each study area, the corresponding SAR images were nature analysis (see Section IV), it was reasonable to assume
terrain geocoded [29] to the geographic projection using the that the backscatter uncertainty would play a minor role in this
3-arcsec Shuttle Radar Topography Mission (SRTM) Digital study.
Elevation Model [30] south of 60 or digitized terrain eleva- As a final step, compensation of the SAR backscatter in the
tion information [31] north of 60 . A look-up table describing case of sloped terrain was applied as described in [35], [36]
the transformation between the radar and the map geometry using an estimate of the pixel area and the local incidence angle.
was generated based on orbital parameters and the DEM. To
correct for geolocation errors due to errors of orbital parameters B. Land Cover Datasets
or the SAR image metadata, a refinement of the lookup table
was applied. The refinement consisted of estimating the offset Two land cover datasets with proven thematic accuracy, geo-
between the SAR backscatter image to be geocoded and a metric precision and reliability were selected to act as reference
reference image for the output geometry [32]. The reference throughout this study: the CORINE (Coordination of Informa-
image was typically a SAR backscatter image simulated from tion on the Environment) land cover dataset [37] and the SRTM
the DEM. Because of the predominantly flat terrain within the Water Bodies Dataset (SWBD) [38].
study areas of the Netherlands and Poland, the latter image The CORINE land cover dataset was produced for the large
did not present any feature that would match with the SAR majority of the member states of the European Union with a
backscatter image. Here, the reference consisted of a mosaic spatial resolution of 100 m and 250 m for the years 1990, 2000,
of Landsat images obtained through the Global Land Cover and 2006 from high-resolution optical imagery. For our work,
Facility (GLCF, http://glcf.umd.edu), resampled to the pixel the 250 m dataset of 2006 (referred to as CLC2006) was con-
size of the SAR data. sidered [39].
The registration error of the SAR images with respect to the The SRTM Water Bodies Dataset (SWBD) is a vector dataset
reference datasets was less than 1/3rd of the pixel size. The SAR derived from the SRTM data acquired in February 2000. It
images with 300 m pixel size were geocoded to 1/360th of a de- shows shoreline of water bodies (lakes, rivers and oceans) at the
gree, in accordance with the GlobCover product [25]. SAR im- time of acquisition of the radar data [38]. The dataset covers all
ages with a pixel size of 150 m were geocoded to corresponding land masses between 60 and 60 with a spatial resolution
pixel sizes in degrees, i.e., 1/720th of a degree. of 90 m. The time difference between the SRTM and the ASAR
The co-registered SAR images were filtered with a multi-tem- dataset (five years) was taken into account by cross-checking
poral approach to decrease speckle noise [1]. The filtered images the SWBD dataset with Google Earth imagery in the case of
were obtained from the original dataset by means of a linear large differences between SWBD and the mapping result from
combination with weights corresponding to the local spatial av- the ASAR dataset (e.g. recent water reservoirs).
erage of the individual (unfiltered) intensity channels. Over tex- For the comparison with the SAR datasets and the corre-
tured terrain, however, spatially adaptive filters allow better es- sponding water body maps, the CLC2006 and SWBD datasets
timates of the radar cross section; hence, it was chosen to de- were transformed to the map geometry and the pixel size of the
fine the weights starting from a speckle-reduced image obtained geocoded SAR data. The CLC2006 land cover was resampled
with the GAMMA MAP filter [33] rather than from the original with a nearest neighbor approach. The SWBD dataset was ras-
unfiltered image. The advantage of the multi-temporal approach terized to the pixel size of the SAR datasets. The CLC2006 and
compared to a spatial filter is preservation of spatial resolution. the SWBD datasets complemented each other to cover all study
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

areas. For study areas where both datasets were available, the
agreement in terms of water and land classes was above 90%.

C. Samples Based on Google Earth Imagery

Samples extracted from high-resolution imagery in global
image and map viewers such as Google Earth are an alterna-
tive approach to generate reference information. In this study,
a stratified random sampling approach has been developed to
select land and water samples in equal manner. Polygons corre-
sponding to a pixel in the SAR image were overlaid onto Google
Earth image and the land cover therein was interpreted. In the
case of mixed water/land cover, e.g., in proximity of shorelines,
the water cover fraction within the polygon was also estimated.
The estimates were reported in intervals of 5%; a value of 1%
refers to pixels edging a shoreline but not including any signifi-
cant portion of water. To limit bias due to the operator, the poly-
gons were revisited after some time and a second estimation of
fractional water cover was performed. The estimates of water
fraction differed at most by 10% from which we concluded that
the estimates of water fraction were sufficiently reliable to be
used as reference. The number of reference samples was deter- Fig. 2. Images of TV (top) and MB (bottom) for the study area of Västerbotten.
mined with respect to a minimum required product accuracy and
confidence level [40]. To obtain a classification accuracy of at
least 85% with a confidence level of 0.05, 93 was the minimum MB was lowest over water because of the repeated occurrence
number of samples required. To allow a correct representation of of specular scattering in the forward direction (i.e., calm wind
the three possible classes involved in our study (“pure” water, conditions).
“pure” land and mixed pixels), 279 samples were selected for In this study, we assumed that the effect of noise in areas
each study area. of very low backscatter was negligible. The Noise Equivalent
When labeling each sample, the year of the image in Google Sigma Zero (NESZ) of Envisat ASAR in Wide Swath Mode was
Earth was taken into account. The majority of the samples between 21 and 26 dB [41]. Such values occurred seldom
were evaluated from high-resolution imagery acquired in 2005 and only for water areas (see Fig. 3). The effect of NESZ was
and 2006. Samples for which data were available for 2009 and found to be negligible when the MB was set equal to a low rank
2010 only were cross-checked with the CLC2006 or the SWBD parameter of the backscatter histogram rather than to the lowest
datasets. The screening did not reveal any difference between value.
the Google Earth samples and the raster dataset. The different behavior of TV and MB over water and land is
further shown by the density plots in Fig. 4 for the study area of
the Netherlands and the major four land cover classes therein.
IV. SIGNATURE ANALYSIS
Land cover classes were identified using the CLC2006 product.
The multi-temporal SAR metrics considered in this study A one-pixel edge-eroded version of the CLC2006 product was
were the minimum backscatter (MB) and the temporal vari- used to focus on pure pixels of a certain land cover type.
ability (TV) of the backscatter defined as the standard deviation While the range of values of the individual parameters over-
of the backscattered intensities in the logarithmic decibel (dB) lapped, the combination of TV and MB showed a clear sepa-
scale. The use of the dB scale for the latter parameter enhanced ration between the water and the land classes. Water presented
the contrast with respect to a standard deviation based on high TV and low MB in consequence of the strong variability of
intensities in the linear scale. the SAR backscatter in time and the very low return under calm
Fig. 2 shows the images of TV and MB for the study area conditions resulting in specular reflection in the forward direc-
of Västerbotten. Inland water bodies and the Gulf of Bothnia tion, respectively. Other land cover classes were mostly charac-
in the east of the study area were characterized by the highest terized by lower TV and higher MB.
TV and lowest MB. To get an understanding for the behavior Density plots with respect to the most frequent land cover
of TV and MB of water and land surfaces, Fig. 3 shows the classes were investigated for all study areas. Water bodies were
time series of the SAR backscatter for three pixels labeled in characterized by values of TV mostly above 2 dB and minimum
CLC2006 as water body, arable land and forest, respectively. backscatter below 16 dB. Other land cover classes presented
The continuous variation of the SAR backscatter in time over TV values between 1 and 3 dB and minimum backscatter above
open water implied the highest TV among the three cases here 15 dB. The wetlands class consisted of inland marshes, peat
considered. The TV of arable land was affected by changes bogs and salines. Marshes and peat bogs were characterized by
of the backscatter during the growing season, located approx- a rather constant backscatter in time except for areas affected
imately in the middle of the time series. The TV of forest was by seasonal flooding, e.g., Doñana wetlands in the south of An-
very low since the backscatter was rather constant in time. The dalusia [4] where the TV was between 2 and 3 dB and the MB
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SANTORO AND WEGMÜLLER: MULTI-TEMPORAL SYNTHETIC APERTURE RADAR METRICS APPLIED TO MAP OPEN WATER BODIES 5

Fig. 5. Scatterplots of TV and MB for water and other land cover classes (pure
pixels) and with respect to the number of SAR observations available per pixel.
Fig. 3. Time series of SAR backscatter for three pixels labeled in CLC2006 Study area: Andalusia. Pixel size: 300 m 300 m. Land cover mask: CLC2006
as water body (top), arable land (middle) and forest (bottom), respectively. TV (one-pixel edge eroded).
and MB estimates are presented above the corresponding panel. Study area: the
Netherlands. Pixel size: 300 m 300 m.

[42]. A J-M distance close to zero indicates no separability be-

tween the classes being compared whereas a distance of 2 in-
dicates total separability. J-M distances were computed for the
water class with respect to each of the four major classes in the
CLC2006 product (see Fig. 4), for either the TV and the MB.
The measured J-M distances were between 1.9 and 2.0 regard-
less of the metric and the land cover class being compared to.
These results confirmed very high separability between water
and land classes. Nevertheless, there are reasons for which TV
and MB over land and water were similar as discussed in the
remainder of this section.

A. Number of SAR Backscatter Measurements

Since the TV is defined as a standard deviation, a reliable es-
timation of the TV requires a large enough number of measure-
ments. Especially over water the TV estimation is affected if the
available number of measurements was below 10. Fig. 5 shows
four scatterplots of TV and MB for water and other land cover
classes, each being characterized by a threshold on number of
observations per pixel. The TV and MB of the land pixels were
Fig. 4. Density plots of TV and MB for pixels belonging to the water class with unaffected by the number of observations (red circles vs. grey
respect to major land classes for the study areas of the Netherlands. Increase of
density is shown by colors changing from blue to red. The density plot for the
dots). On the contrary, the TV of water pixels based on less than
water class is shown in each scatterplot for clarity reasons. Pixel size: 300 m 10 backscatter measurements was smaller compared to the es-
300 m. Land cover mask: CLC2006. timate obtained for water pixels with more backscatter obser-
vations (dark circles vs. cyan dots). When considering a higher
threshold for the number of observations per pixel (panel for
between 13 and 15 dB, i.e. in between features of perma- thresholds of 15 and 20 in Fig. 5), the behavior of TV did not
nent land and permanent water surfaces. The MB and TV of differ with respect to pixels characterized by more observations.
salines instead was similar to the observations over open water. Although, the interpretation of the different behavior of TV for
In this study, salines were considered as “land” because, in gen- the threshold of 10 observations is supported by a very small
eral, they may not be filled with water for a certain period of number of pixels and there were not less than 10 observations
time. per pixel at the other study areas, we interpreted the result as
To quantify the separability between the water class and a a warning that the TV over water could be unreliable if based
land class, the Jeffreys-Matusita (J-M) distance was computed upon a small number of measurements.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Fig. 6. Histograms of TV (left panel) and MB (right panel) for pixels identified as water, land with slope angle 10 and land with slope angle 10 in the
Switzerland study area. Pixel size: 300 m 300 m. Land cover mask: SWBD (one-pixel edge eroded). Slope angle derived from SRTM DEM.

B. Topography
Areas of layover and shadow were identified during SAR
processing using information from the DEM, orbital data and
look geometry parameters [29] and were masked out from
the geocoded SAR images. For other areas on sloped terrain
showing modulation of the backscatter after the compensation
for pixel area and local incidence angle, we explained such
distortions as a consequence of imperfect characterization of
slopes in the DEM or orientation-dependent scattering effects
[43], [44]. The effect of sloped terrain on TV and MB was
investigated by creating histograms for water and land with
slope angle either below or above a certain threshold. For a
threshold of 10 , we saw the clearest indications on the effect
of sloped terrain on the two parameters. The histograms of TV
and MB for water and land surfaces with gentle to moderate Fig. 7. Scatterplots of TV and MB for pure water, pure land and mixed pixels.
slopes ( 10 ) were almost disjoint (Fig. 6). The histograms Study area: Andalusia. Pixel size: 300 m 300 m. Land cover mask: SWBD.

for land surfaces with steep slopes ( 10 ) instead committed

to more joint parts when compared with the histograms of land
surfaces with slopes .

C. Mixed Pixels
The signatures of TV and MB were ambiguous in the case of
pixels located along shore- and coastlines with a certain water
fraction, i.e., mixed pixels. Fig. 7 shows a scatterplot of TV and
MB for pure water (cyan dots), pure land (red dots) and mixed
pixels (black circles) in the case of the study area of Andalusia.
The values of the two parameters for mixed pixels were located
between the clusters of values characteristic of pure water and
pure land. The effect of water fraction on TV and MB is fur-
ther shown in Fig. 8. For a water fraction of 0% (i.e., pure land)
or 100% (i.e., pure water), the ranges of values of the two pa- Fig. 8. TV (top) and MB (bottom) with respect to water fraction for the study
rameter were disjoint. For increasing water fraction, the TV in- area of the Netherlands. Pixel size: 300 m 300 m. Number of samples: 279.
creased whereas the MB decreased with an almost linear trend.
For a given water fraction, both parameters presented certain
variability, which could be explained as a consequence of the D. Conditions of the Land Cover
type of scatterers within the fraction of land surface inside the Special land cover conditions that occur for a limited time,
area covered by the pixel. TV was lower when the land part of such as wet snow, flooding of a field etc. [45], can have a strong
the pixel consisted of forest rather than cropland and bare soil; influence on the backscatter and, consequently, on the multi-
similarly, the MB was higher. For mixed pixels with a small temporal SAR metrics. Herewith, we focus on two conditions
fraction of the area corresponding to a single strong scatterer that occurred at several study areas.
over land (e.g., from infrastructures), the TV was low and the In the case of wet snow, the strong absorption of the mi-
MB was high in consequence of the dominance of this strong crowaves causes very low backscatter, which in turn increases
and temporally stable scatterer. the temporal variability of the backscatter and produces a very
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SANTORO AND WEGMÜLLER: MULTI-TEMPORAL SYNTHETIC APERTURE RADAR METRICS APPLIED TO MAP OPEN WATER BODIES 7

Fig. 9. Time series of ASAR WSM backscatter for a pixel over barren ground within the study area of Västerbotten. Pixel size: 300 m 300 m Thawing conditions
of snow cover registered on 08 May. 2.7 dB and 20.3 dB.

Fig. 10. Time series of ASAR WSM backscatter for a pixel on a glacier within Fig. 11. Time series of ASAR WSM backscatter for a pixel over a rice field at
the study area of Switzerland. 7 dB, MB 20.8 dB. Pixel size: the study area of Andalusia. Pixel size: 300 m 300 m.
300 m 300 m.

minimum and then increased during fall until the level before
low MB. A first example is shown in Fig. 9 for a time series of summer. For rice fields, temporary flooding caused sudden
backscatter measurements over barren ground. The backscatter decrease of the backscatter, which then increased steadily until
dropped on 08-May when weather data reported snow cover harvest (Fig. 11). The consequence was very low minimum
and indicated thawing conditions. The TV and the MB for backscatter and very large temporal variability (above 4 dB).
the backscatter time series were 2.7 dB and 20.3 dB, thus
falling within the range of values typical of water (see Fig. 4). E. Using Multiple Orbital Tracks
Any other measurement of the backscatter was above 16 dB In the case of a multi-temporal dataset including observa-
whereas the TV was 2.5 dB when the observation of 08 May tions from multiple orbital tracks, the TV and MB are influ-
had been excluded from the time series. The multi-temporal enced by the different look angles. Discussing look angle de-
SAR metrics were therefore in the range of values obtained pendence of the SAR backscatter for different scattering objects
for land surfaces. A second example is shown in Fig. 10 for is beyond the scope of this paper. Here, we focus on a compar-
a time series of backscatter measurements over a glacier. The ison between single- and multi-track TV and MB values to un-
backscatter was low during late spring and summer (from derstand the impact of look angle on the two parameters. This
May until September) because of the wet conditions of the investigation required that for several tracks, a certain number
snow cover. The TV and the MB were 7 dB and 20.8 dB. of backscatter measurements were available to be able to limit
Disregarding data acquired under wet snow conditions, the TV bias and uncertainties in the track-wise estimation of TV and
and the MB (2.5 dB and 6.0 dB, respectively) were within MB. Because of the 35-day repeat-pass cycle of Envisat, the
the range of values of a land surface. maximum number of observations per track in one year was 11.
Depending on the type of crop, the polarization of the A relatively large number of observations (i.e., more than 5)
microwave and cultivation practice, there can be substantial for several tracks were available only for the Poland study area.
absorption in the vegetation layer, leading to low minimum Fig. 12 shows an example of track-wise TV and MB for water
backscatter estimates very close to the level observed for water. and cropland. Among the land classes in the study area, crop-
This issue could be observed at the study areas of Andalusia, the land presented the largest variability of backscatter. Despite this
Netherlands, Poland, and Central Siberia. During the summer variability, the single-track TV was always smaller than the cor-
months, the backscatter of cropland decreased and reached a responding value of water; similarly, the MB was always higher
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Fig. 12. Single-track TV and MB for two pixels located in areas labeled in the CLC2006 dataset as water and land (cropland), respectively. For each satellite
track, the corresponding incidence angle in degrees at the location of the pixel is displayed next to the metric of the water pixel.

for land than for water. Fig. 12 also shows that TV and MB reliable estimates of TV and MB was the Gulf of Bothnia in the
were not much affected by the look angle, here expressed in east of the Västerbotten study area. Here, the polarization-spe-
the form of the local incidence angle to be more precise. The cific TV and MB values did not present any difference also when
multi-track TV was equal to 4.4 dB (water) and 1.7 dB (land), compared to the corresponding values obtained disregarding the
only slightly higher than the single-track values. The MB was polarization. It is, however, necessary to investigate this aspect
equal to 22.3 dB (water) and 15.1 dB (land), equal to the ab- taking into account that current and future SAR missions will
solute minimum over all tracks. acquire multi-temporal observations either in dual- or in full po-
In general, TV and MB derived from a multi-track dataset of larization.
observations appeared as slightly enhanced with respect to the
single-track case. Although, the two parameters can be consid- V. WATER BODY CLASSIFICATION METHODOLOGY
ered being affected primarily by a temporal component, both
also include a geometric component. Our interpretation applies The scatterplots of TV and MB of land and water showed
to the land cover types in the six study areas. In polar and desert symmetry of TV and MB for land and water with respect to
regions, the backscatter from volumes (snow, ice, or sand layers) a diagonal line represented by a linear equation of increasing
or rugged terrain differs depending on the look angle [46], [47] MB for increasing TV. A simple thresholding rule in the fea-
thus introducing a more relevant geometric component in the ture space of TV and MB seemed to be sufficient to map water
estimation of TV and MB. and land. In this study, we defined the thresholding rule as the
diagonal line that was at equal distance from pre-defined clus-
F. Summary ters of “pure” land and “pure” water in the case of the study
area of the Netherlands. The training dataset consisted of 10%
The outcome of the signature analysis is summarized below of the pixels belonging to each of the two classes according to
and served to set up a water body mapping algorithm based on the CLC2006 dataset.
multi-temporal SAR parameters. Equation (1) corresponds to the diagonal line representing the
1. TV and MB are sufficient for separating water and land threshold in the feature space of TV and MB:
surfaces in the case of pure pixels.
2. Water bodies and land surfaces present almost the same (1)
signatures regardless if the SAR datasets consists of multi-angle
or single-track observations. Here, represents the TV in dB and the MB in dB. This
3. Mixed pixels might be assigned to one or the other class thresholding rule was found to yield a very good separation be-
depending on water fraction and, to a certain extent, on the scat- tween pure land and pure water for the Netherlands (Fig. 13)
tering objects on land. and for all other study areas as well. The observations of TV
4. Topographic information in form of slope angle needs to and MB of mixed pixels were instead crossed by the diagonal
be accounted for to avoid false detections of land as water. line. Ultimately, we preferred setting up a simple classification
5. The SAR metrics are ambiguous in the case of a small approach to understand the potential of the TV and MB to sepa-
number of backscatter measurements. rate water and land rather than proceeding with a more complex
6. The definition of MB should rely on a low order rank of algorithm already at the beginning of our classification exercise.
the backscatter histogram rather than on the absolute minimum The signature analysis furthermore revealed that TV below
backscatter to avoid that specific events within the time span of 1.5 dB and MB above 16 dB are not realistic for water (see
the backscatter measurements distort the metric. Fig. 4). Combining the three thresholding corresponded to
7. The impact of polarization on the signatures of TV and MB defining two portions of the feature space associated to water
could not be assessed because the large majority of the ASAR and land respectively (Fig. 13).
images were acquired in VV-polarization. The only area cov- To avoid false detections of land as water in the case of steep
ered by a number of HH-polarized images sufficient to derive topography, pixels characterized by a slopes steeper than 10
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SANTORO AND WEGMÜLLER: MULTI-TEMPORAL SYNTHETIC APERTURE RADAR METRICS APPLIED TO MAP OPEN WATER BODIES 9

Fig. 13. Illustration of the water body mapping algorithm. Decision rules are
represented by the black diagonal line and the two dashed lines. The water and
the land regions in the feature space of TV and MB are marked accordingly in
light blue and green, respectively.

have been labeled automatically as land. To avoid false classifi-

cation in the case of a small dataset of backscatter observations,
pixels with less than 10 backscatter measurements have been
discarded. In this study, only the westernmost edge of the An-
dalusia study area could not be classified, corresponding to 5%
of the total number of pixels in the study area.

A. Assessment of Water Body Mapping Accuracy

Verification of the water body maps obtained from the ASAR
data is provided in the form of percentages of agreement with
respect to (i) the CLC2006 product re-coded to water and
non-water, or the SWBD product if the former was not avail-
able, and (ii) the samples extracted from Google Earth imagery.
The wetland classes “marshes”, “peat bogs” and “salines” of
the CLC2006 dataset were re-coded as land. The wetland class
“inter-tidal flats” was re-coded as water in accordance with the
definition of these in the GlobCover dataset. Regardless of the
reference dataset, we use the terms of user’s and producer’s
accuracy (UA and PA), overall accuracy (OA) and Kappa
coefficient as defined in [48] to quantify the agreement between
the SAR-based classification and a reference dataset. The UA
estimates the error related to an inclusion of samples in a given
class (commission error); conversely, the PA estimates the
error related to an exclusion of sample units from a given class Fig. 14. ASAR-based water body maps using as MB the lowest backscatter
(omission error). The OA and the Kappa coefficient indicate value, the backscatter fifth rank and the backscatter 10th percentile. For refer-
the overall agreement between the SAR-based classification ence, the CLC2006 dataset expressed in terms of a water and a land class is
shown in the bottom panel. Grey is used for land, white is used for water. Study
and the reference data. The latter statistic is used to detect a area: Västerbotten. Pixel size: 300 m 300 m. Note the larger amount of water
possible agreement by chance between two datasets, which is bodies detected between 14 and 16 in the top panel with respect to the
embedded in the definition of the OA. To assess the impact of other panels.
shorelines on the percentages of agreement with respect to the
raster datasets, both the original and the one-pixel edge-eroded fifth rank, and (iii) the backscatter 10th percentile. While using
version were used. the rank as MB is more adequate for large backscatter datasets,
the percentile is better suited when the multi-temporal dataset
VI. CLASSIFICATION RESULTS AND DISCUSSION is small. The choice of definitions for the MB is not exhaustive;
The classification algorithm was applied to the SAR data pro- however, it was not the scope of this analysis to investigate the
cessed at 300 m for each of the study areas. To assess the robust- optimal selection of the MB channel. With (ii) and (iii), the aim
ness of the mapping algorithm to seasonal conditions and par- was to demonstrate possible approaches to avoid that measure-
ticular land-cover characteristics, three different definitions for ment noise and sporadic events would affect the MB value to
MB were tested: (i) the lowest backscatter, (ii) the backscatter cause false detections.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

TABLE III
USER’S AND PRODUCER’S ACCURACY (UA AND PA), OVERALL ACCURACY (OA) AND KAPPA COEFFICIENT WITH RESPECT TO THE CLC2006 DATASET RECODED
TO WATER AND LAND CLASSES OR THE SWBD PRODUCT, DEPENDING ON AVAILABILITY. THE SHADED ROWS REFER TO FIGURES RELATIVE TO THE ONE-PIXEL
EDGE-ERODED VERSION OF EACH PRODUCT. ACCURACIES ARE REPORTED FOR THE MAP OBTAINED USING AS MB (I) THE LOWEST BACKSCATTER, (II) THE
BACKSCATTER FIFTH RANK AND (III) THE BACKSCATTER 10TH PERCENTILE. PIXEL SIZE: 300 m 300 m. ACCURACIES ARE IN THE FORM OF % VALUES

A comparison of water body maps with respect to the refer- on row (i) in Table III). The commission error for the water
ence dataset (CLC2006 re-coded to water and land) is shown in class decreased because false detection of land as water could
Fig. 14 for the study area of Västerbotten. Major water bodies be avoided by neglecting data acquired under special environ-
and the Baltic Sea were mapped correctly regardless of the defi- mental conditions that could distort the MB (Table III). The
nition of MB. The MB set equal to the lowest backscatter value agreement statistics for the UA of the water class however did
(Fig. 14, top panel) caused false detections of land surfaces as not give a clear indication on which of the two definitions here
water in correspondence of the western edge of the study area chosen for the MB metric (5th rank or 10th percentile) should
between 14 and 16 . Here, wet snow conditions caused ultimately be preferred (Tables III and IV). For classification
high TV and low MB when all measurements were considered based on a histogram parameter, the UA was above 92% for all
(see also Fig. 2). Using a low-order parameter of the backscatter study areas (rows (ii) and (iii) in Tables III and IV) except for
histogram rather than the absolute minimum implied that such Andalusia. The lower agreement was a consequence of a large
special conditions were disregarded and the corresponding portion of salines (wetlands) being detected as water. Because
water body map (Fig. 14, central panels) presented stronger of the fragmentation of salines, the UA in the case of pure pixels
agreement with the reference (Fig. 14, bottom panel). was higher than when all pixels were considered (90% vs. 80%).
Table III reports agreement statistics for each of the six study Herewith, we note that wetlands in the form of marshes and peat
areas when using the raster datasets (CLC2006 or SWBD) as bogs were detected as land in at least 95% of the cases. Inter-
reference. For study areas where both datasets were available, tidal flats were detected as water in more than 99% of the cases.
we note that the difference between accuracies using one or the Mixed pixels, i.e., pixels along shorelines, were often labeled
other dataset was less than a few percent units. Table IV reports as land while they were defined as water in the reference dataset.
the agreement numbers in the case of the Google Earth samples Omissions of water areas were more frequent in study areas
with pure land and pure water information. characterized by narrow water bodies (Andalusia, Poland, and
Mapping of water bodies performed better when using as Västerbotten). The PA of water was mostly in the 30% to 50%
MB a low-order parameter of the backscatter histogram (cf. UA range (Table III). When shoreline pixels were excluded from
of rows (ii) and (iii) with respect to the corresponding UAs the analysis, the PA of water was mostly above the 90% level
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SANTORO AND WEGMÜLLER: MULTI-TEMPORAL SYNTHETIC APERTURE RADAR METRICS APPLIED TO MAP OPEN WATER BODIES 11

TABLE IV
USER’S AND PRODUCER’S ACCURACY (UA AND PA), OVERALL ACCURACY (OA) AND KAPPA COEFFICIENT WITH RESPECT TO GOOGLE EARTH SAMPLES
( number of samples). THE VALIDATION DATASET CONSISTED OF POLYGONS WITH WATER FRACTION OF 100% (WATER CLASS) AND (LAND CLASS).
ACCURACIES ARE REPORTED FOR THE MAP OBTAINED USING AS MB (I) THE LOWEST BACKSCATTER, (II) THE BACKSCATTER FIFTH RANK AND (III) THE
BACKSCATTER 10TH PERCENTILE. PIXEL SIZE: 300 m 300 m. ACCURACIES ARE IN THE FORM OF % VALUES.

except for Andalusia where PA was between 71% and 81% de-
pending on the definition of MB. The poorer performance of
the mapping algorithm in Andalusia was related to the irregular
shoreline of several water reservoirs which appeared smaller
compared to the reference dataset. This indication is supported
by the higher PA based on Google Earth samples (above 85%,
Table IV), where the proportion of samples in correspondence
of a shoreline was smaller compared to the CLC2006 dataset.
The overall classification accuracy was above 90% for all
study areas (Tables III and IV). The Kappa coefficient instead
depended on the definition used for MB and whether shoreline
pixels had been excluded from the reference dataset (Table III).
This in turn implied that size and shape of water bodies affected
the Kappa coefficient. The lowest coefficients were obtained in
Fig. 15. UAs, PAs and OA with respect to a threshold for water fraction for the
Andalusia (45% and 80%, average of the three MB cases). For study area of the Netherlands. Pixel size: 300 m 300 m.
Poland and Västerbotten, the Kappa coefficient was of the order
of 60% when shorelines were taken into account whereas for the
remaining study areas the coefficient was approximately 80% 90%. A threshold of 50% was instead characterized by OAs
(average of the three MB cases). When restricting to pure pixels between 75% and 85%. The differences of OA for a given
only, the Kappa coefficient was above 90%. threshold were related to the amount of water pixels omitted
To further quantify the impact of water fraction in a pixel on by the classification algorithm. Lower OAs were obtained for
the accuracy of the water body maps, we illustrate in Fig. 15 the study areas with small water bodies and irregular shorelines.
agreement statistics with respect to a threshold on water frac- The classification algorithm applied to SAR data with the
tion in the case of the Netherlands study area. The PA of the pixel size of 150 m mapped features with a size on the order of
water class and thus the UA of land improved for increasing 300 m which were omitted in the classification using the SAR
minimum water fraction within a pixel because the omission of data with the alike pixel size. As an example, in correspondence
water areas was stronger in mixed pixels with larger fraction of of the Rhine river delta in the study area of the Netherlands, the
land. Commission errors of water areas were very seldom (UA percentage of pixels labeled as water increased from 14.7% to
of water almost 100%) because in the case of mixed pixels, the 15.4%. To assess the impact of pixel size on the classification,
multi-temporal metrics were found to be affected primarily by we compared the agreement statistics between the ASAR-based
the properties of the radar backscatter from the land fraction water body maps and the reference datasets at 150 m and 300 m
within the pixel (see Section IV-C). in Table V. When restricting to pure pixels of water and land
The OAs for the six study areas presented similar trends, (CLC2006 dataset), all agreement statistics were above 95% and
indicating increased agreement between ASAR-based water the Kappa coefficient was above 0.97. Lower PA of water and in
body maps and reference datasets for threshold on water frac- turn lower OA (approximately 88%) occurred when shoreline
tion (Fig. 16). Defining as water a pixel with at least 75% water pixels were included in the reference dataset. Benchmarking
fraction implied an overall mapping accuracy of approximately against the Google Earth dataset indicated more omissions of
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

TABLE V
USER’S AND PRODUCER’S ACCURACY (UA AND PA), OVERALL ACCURACY (OA) AND KAPPA COEFFICIENT FOR THE 150 M PIXEL SIZE WITH RESPECT TO
THE 300 M PIXEL SIZE (SHADED CELLS) FOR DIFFERENT REFERENCE DATASETS. THE MB CORRESPONDED TO THE BACKSCATTER FIFTH RANK. ACCURACIES
ARE IN THE FORM OF % VALUES.

mixed pixels were taken into account. For a threshold of 50%

of water fraction, the overall accuracy was approximately 80%.
Omissions of water areas occurred in correspondence of
mixed pixels with a water fraction and for narrow water bodies
with irregular shorelines. Water bodies with a dimension less
than twice the pixel size of the SAR data were not detected.
High resolution images such as those provided by the forth-
coming Sentinel-1 mission might reduce the area affected by
omissions of water, particularly in the case of small-sized water
bodies; it is however expected that mixed water/land pixels
along shorelines are classified mostly as land since because
the TV and MB are affected by the scattering properties of the
Fig. 16. OA for each study area with respect to the threshold for water fraction. objects within the resolution cell rather than by the resolution
Pixel size: 300 m 300 m. of the sensor itself.
While this study focused on mapping permanent water
bodies, the approach presented here is in theory applicable also
water areas for the pixel size of 150 m than 300 m. At 300 m,
to monitor the dynamics of water bodies by using SAR data
several narrow-sized water bodies neither appeared in the refer-
from selected time windows. Assessing the detection of water
ence dataset nor were detected in the ASAR data. Conversely,
surfaces in consequence of temporary events like inundation
at 150 m they were still not detected in the ASAR data but ap-
and flooding would require dense time series of measurements
peared in the reference dataset although they were mostly char-
in correspondence of the specific event. The availability of
acterized by a certain land fraction which in turn caused the
daily observations of ASAR Wide Swath Mode observations
omissions.
at latitudes north 60 could be used to demonstrate the capa-
bility of time series of short-term TV and MB to track water
VII. CONCLUSIONS AND OUTLOOK
dynamics in tundra regions.
In this study, we looked at the potential of SAR multi-tem-
poral metrics for land cover characterization with particular ACKNOWLEDGMENT
regard to discriminate between open water bodies and land P. Defourny, S. Bontemps ,and J. Radoux, UCL, are ac-
surfaces. The SAR dataset consisted of C-band Envisat ASAR knowledged for advice on the classification approach. The
Wide Swath Mode images because of the large amount of SAR dataset was obtained through ESA’s Category 1 Project
data available worldwide, easy access and the moderate spatial ID 8073. K. Traut, FSU Jena, is acknowledged for support in
resolution. Although data from other spaceborne SAR sensors the selection and the ordering of the Envisat ASAR data. The
might be more suitable for such application (e.g., L-band, dual- authors are truly indebted to the anonymous reviewers for their
and full polarimetric data, shallow look angles, meter resolu- comments.
tion), none satisfies the requirements of easy access, frequent
and large-area coverage. REFERENCES
A simple thresholding algorithm based on the temporal [1] S. Quegan and J. J. Yu, “Filtering of multichannel SAR images,” IEEE
variability (TV) of the SAR backscatter and the minimum Trans. Geosci. Remote Sens., vol. 39, pp. 2373–2379, 2001.
[2] C. Pathe, W. Wagner, D. Sabel, M. Doubkova, and J. B. Basara, “Using
backscatter (MB) estimated from stacks including at least 10 ENVISAT ASAR global mode data for surface soil moisture retrieval
measurements could discriminate between permanent water over Oklahoma, USA,” IEEE Trans. Geosci. Remote Sens., vol. 47, pp.
bodies and land surfaces at 150 and 300 m pixel size in a 468–480, 2009.
[3] A. Bartsch, W. Wagner, K. Scipal, C. Pathe, D. Sabel, and P. Wolski,
consistent manner. Defining the MB as low order ranks or “Global monitoring of wetlands–The value of ENVISAT ASAR global
percentiles rather than the lowest backscatter served to avoid mode,” J. Environ. Manage., vol. 90, pp. 2226–2233, 2009.
commission errors for the water class. The overall classification [4] B. Marti-Cardona, C. Lopez-Martinez, J. Dolz-Ripolles, and E.
Bladè-Castellet, “ASAR polarimetric, multi-incidence angle and
accuracy was above 90% when restricting to pure pixels; the multitemporal characterization of Doñana wetlands for flood extent
accuracy decreased linearly with water fraction when also monitoring,” Remote Sens. Environ., vol. 114, pp. 2802–2815, 2010.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SANTORO AND WEGMÜLLER: MULTI-TEMPORAL SYNTHETIC APERTURE RADAR METRICS APPLIED TO MAP OPEN WATER BODIES 13

[5] J. Reschke, A. Bartsch, S. Schlaffer, and D. Schepaschenko, “Capa- [26] “ASAR product handbook,” ESA ESRIN, no. 2.2, Feb. 27, 2007.
bility of C-band SAR for operational wetland monitoring at high lati- [27] J. Closa, B. Rosich, and A. Monti Guarnieri, “The ASAR wide swath
tudes,” Remote Sens., vol. 4, pp. 2923–2943, 2012. mode products,” presented at the IGARSS’03, Toulouse, France, Jul.
[6] A. Bouvet and T. Le Toan, “Use of ENVISAT/ASAR wide-swath data 21-25, 2003.
for timely rice fields mapping in the Mekong River delta,” Remote Sens. [28] M. A. Friedl, D. K. McIver, J. C. F. Hodges, X. Y. Zhang, D. Mu-
Environ., vol. 115, pp. 1090–1101, 2011. choney, A. H. Strahler, C. E. Woodcock, S. Gopal, A. Schneider, A.
[7] T. Geldsetzer, J. van der Sanden, and B. Brisco, “Monitoring lake ice Cooper, A. Baccini, F. Gao, and C. Schaaf, “Global land cover map-
during spring melt using RADARSAT-2 SAR,” Can. J. Remote Sens., ping from MODIS: Algorithms and early results,” Remote Sens. Env-
vol. 36(S2), pp. 391–400, 2010. iron., vol. 83, pp. 287–302, 2002.
[8] A. Bartsch, A. M. Trofaier, G. Hayman, D. Sabel, S. Schlaffer, D. [29] U. Wegmüller, “Automated terrain corrected SAR geocoding,” pre-
B. Clark, and E. Blyth, “Detection of open water dynamics with EN- sented at the IGARSS’99, Hamburg, Germany, 28 Jun.–2 Jul. 1999.
VISAT ASAR in support of land surface modelling at high latitudes,” [30] B. Rabus, M. Eineder, A. Roth, and R. Bamler, “The shuttle radar to-
Biogeosciences, vol. 9, pp. 703–714, 2012. pography mission–A new class of digital elevation models acquired by
[9] C. Kuenzer, H. Guo, J. Huth, P. Leinenkugel, X. Li, and S. Dech, spaceborne SAR,” ISPRS J. Photogramm. Remote Sens., vol. 57, pp.
“Flood mapping and flood dynamics of the Mekong Delta: ENVISAT- 241–262, 2003.
ASAR-WSM based time series analyses,” Remote Sens., vol. 5, pp. [31] J. de Ferranti, “Digital Elevation Data,” 2009 [Online]. Available:
687–715, 2013. http://www.viewfinderpanoramas.org/dem3.html
[10] S. Quegan, T. Le Toan, J. J. Yu, F. Ribbes, and N. Floury, “Multi- [32] U. Wegmüller, C. Werner, T. Strozzi, and A. Wiesmann, “Automated
temporal ERS SAR analysis applied to forest mapping,” IEEE Trans. and precise image registration procedures,” in Analysis of Multi-tem-
Geosci. Remote Sens., vol. 38, pp. 741–753, 2000. poral Remote Sensing Images, ser. Remote Sensing, Bruzzone and
[11] U. Wegmüller, A. Wiesmann, T. Strozzi, and C. Werner, “Forest Smits, Eds. Trento, Italy: World Scientific, Sep. 13-14, 2001, vol. 2,
mapping with multi-temporal SAR,” presented at the ForestSAT 2002 pp. 37–49, 2002.
Conf., Edinburgh, U.K., Aug. 5-9, 2002. [33] A. Lopes, E. Nezry, R. Touzi, and H. Laur, “Structure detection and
[12] P. Ahtonen and M. Hallikainen, “Automatic detection of water bodies statistical adaptive speckle filtering in SAR images,” Int. J. Remote
from spaceborne SAR images,” presented at the IGARSS’05, Seoul, Sens., vol. 14, pp. 1735–1758, 1993.
Korea, Jul. 25–29, 2005. [34] C. Oliver and S. Quegan, Understanding Synthetic Aperture Radar Im-
[13] A. Bartsch, C. Pathe, K. Scipal, and W. Wagner, “Detection of perma- ages. Boston, MA, USA: Artech House, 1998.
nent open water surfaces in central Siberia with ENVISAT ASAR wide [35] A. Wiesmann, U. Wegmüller, M. Santoro, T. Strozzi, and C. Werner,
swath data with special emphasis on the estimation of methane fluxes “Multi-temporal and multi-incidence angle ASAR wide swath data for
from tundra wetlands,” Hydrol. Res., vol. 39.2, pp. 89–100, 2008. land cover information,” in Proc. 4th Int. Symp. Retrieval of Bio- and
[14] B. Brisco, N. Short, J. van der Sanden, R. Landry, and D. Raymond, Geophysical Parameters from SAR Data for Land Applications, Inns-
“A semi-automated tool for surface water mapping,” Can. J. Remote bruck, Austria, Nov. 16-19, 2004.
Sens., vol. 35, pp. 336–344, 2009. [36] M. Santoro, C. Beer, O. Cartus, C. Schmullius, A. Shvidenko, I. Mc-
[15] T. Hahmann, S. Martinis, A. Twele, and M. Buchroithner, “Strategies Callum, U. Wegmüller, and A. Wiesmann, “Retrieval of growing stock
for the automatic mapping of flooded areas and other water bodies volume in boreal forest using hyper-temporal series of Envisat ASAR
from high resolution TerraSAR-X data,” in Cartography and Geoin- ScanSAR backscatter measurements,” Remote Sens. Environ., vol. 115,
formatics for Early Warning and Emergency Management: Towards pp. 490–507, 2011.
Better Solutions. Brno: Masaryk University, 2009, pp. 207–214. [37] G. Büttner, J. Feranec, G. Jaffrain, L. Man, G. Maucha, and T. Soukup,
[16] M. Silveira, “Separation between water and land in SAR images using “The CORINE LAND COVER 2000 Project,” EARSeL eProceedings,
region-based level sets,” IEEE Geosci. Remote Sens. Lett., vol. 6, pp. vol. 3, pp. 331–346, 2004.
471–475, 2009. [38] “Shuttle radar topography mission water body data set 2005,” SWBD,
[17] V. Gstaiger, S. Gebhardt, J. Huth, T. Wehrmann, and C. Kuenzer, Feb. 2012 [Online]. Available: http://www2.jpl.nasa.gov/srtm/index.
“Multi-sensoral derivation of inundated areas using TerraSAR-X and html
ENVISAT ASAR data,” Int. J. Remote Sens., vol. 33, pp. 7291–7304, [39] “CLC2006 technical guidelines,” European Environment Agency,
2012. EEA, Copenhagen, Denmark, 2007.
[18] T. Hahmann and B. Wessel, “Surface water body detection in high- [40] M. E. Ginevan, “Testing land-use map accuracy: Another look,” Pho-
resolution TerraSAR-X data using active contour models,” presented togramm. Eng. Remote Sens., vol. 45, pp. 1371–1377, 1979.
at the EUSAR 2010 Conf., Aachen, Germany, Jun. 7-10, 2010. [41] Envisat [Online]. Available: https://directory.eoportal.org/web/eo-
[19] A. Wendleder, B. Wessel, A. Roth, M. Breunig, K. Martin, and S. Wa- portal/satellite-missions/e/envisat
genbrenner, “TanDEM-X water indication mask: Generation and first [42] P. H. Swain and S. M. Davis, Remote Sensing: The Quantitative Ap-
evaluation results,” IEEE J. Sel. Topics Appl. Earth Observ. Remote proach. New York, NY, USA: McGraw-Hill, 1978.
Sens., vol. 6, pp. 171–179, 2013. [43] T. Castel, A. Beaudoin, N. Stach, N. Stussi, T. Le Toan, and P. Du-
[20] K. D. Unterschultz, J. van der Sanden, and F. E. Hicks, “Potential of rand, “Sensitivity of space-borne SAR data to forest parameters over
RADARSAT-1 for the monitoring of river ice: Results of a case study sloping terrain. Theory and experiment,” Int. J. Remote Sens., vol. 22,
on the Athabasca River at Fort McMurray, Canada,” Cold Regions Sci. pp. 2351–2376, 2001.
Technol., vol. 55, pp. 238–248, 2009. [44] M. A. Tanase, M. Santoro, J. de la Riva, F. Pérez-Cabello, and T.
[21] M. O. Jeffries, K. Morris, W. F. Weeks, and H. Wakabayashi, Le Toan, “Sensitivity of X-, C- and L-band SAR backscatter to burn
“Structural and stratigraphic features and ERS-1 synthetic aperture severity in Mediterranean pine forests,” IEEE Trans. Geosci. Remote
radar backscatter characteristics of ice growing on shallow lakes Sens., vol. 48, pp. 3663–3675, 2010.
in NW Alaska, winter 1991-1992,” J. Geophys. Res., vol. 99, pp. [45] P. Defourny and S. Bontemps, “Revisiting land-cover mapping con-
22459–22471, 1994. cepts,” in Remote Sensing of Land Use and Land Cover: Principles and
[22] C. R. Duguay, T. J. Pultz, P. M. Lafleur, and D. Drai, “RADARSAT Applications, C. Giri, Ed. Boca Raton, FL, USA: CRC Press, 2012.
backscatter characteristics of ice growing on shallow sub-Arctic [46] H. Stephen and D. G. Long, “Microwave backscatter modeling of ERG
lakes, Churchill, Manitoba, Canada,” Hydrol. Process., vol. 16, pp. surfaces in the Sahara desert,” IEEE Trans. Geosci. Remote Sens., vol.
1631–1644, 2002. 43, pp. 238–247, 2005.
[23] M. Nolan, G. Liston, P. Prokein, J. Brigham-Grette, V. L. Sharpton, [47] W. G. Rees, Remote Sensing of Snow and Ice. Boca Raton, FL, USA:
and R. Huntzinger, “Analysis of lake ice dynamics and morphology on CRC Press, 2006.
Lake El’gygytgyn, NE Siberia, using synthetic aperture radar (SAR) [48] R. G. Congalton, “A review of assessing the accuracy of classifications
and Landsat,” J. Geophys. Res., vol. 108, 2003. of remotely sensed data,” Remote Sens. Environ., vol. 37, pp. 35–46,
[24] T. Hirose, M. Kapfer, J. Bennett, P. Cott, G. Manson, and S. Solomon, 1991.
“Bottomfast ice mapping and the measurement of ice thickness on
tundra lakes using C-band synthetic aperture radar remote sensing,”
J. Amer. Water Res. Assoc., vol. 44, pp. 285–292, 2008.
[25] P. Bicheron, P. Defourny, C. Brockmann, L. Schouten, C. Vancutsem,
M. Huc, S. Bontemps, M. Leroy, F. Achard, M. Herold, F. Ranera,
and O. Arino, “GLOBCOVER–Products description and validation re-
port,” Media France 2008.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Maurizio Santoro (M’04) received the M.S. degree Urs Wegmüller (M’94–SM’03) received the M.S.
in aerospace engineering from the University of and Ph.D. degrees in physics from the University of
Naples “Federico II,” Naples, Italy, in 1998, the Bern, Bern, Switzerland, in 1986 and 1990, respec-
Lic.Eng. degree from the Chalmers University of tively.
Technology, Gothenburg, Sweden, in 2001, and the Between 1991 and 1992, he was with the Jet
Ph.D. degree from Friedrich-Schiller University, Propulsion Laboratory, California Institute of Tech-
Jena, Germany, in 2003. nology, Pasadena, CA, USA. Between 1993 and
From 2004 to 2005, he held a postdoctoral posi- 1995, he was with the University of Zürich, Zürich,
tion with Friedrich-Schiller University. Since 2006, Switzerland. In 1995, he was a Founding Member of
he has been a Project Scientist with GAMMA Re- GAMMA Remote Sensing AG (GAMMA), Güm-
mote Sensing AG, Gümligen, Switzerland. His main ligen, Switzerland, a Swiss company active in the
duties include synthetic aperture radar (SAR) and interferometric SAR data pro- development of signal processing techniques and remote sensing applications.
cessing and applications of SAR interferometry for land cover mapping. He is As Chief Executive Officer of GAMMA, he has overall responsibility for
involved as Principal and Coinvestigator in several international projects on the GAMMA’s activities. He is/was the Principal Investigator for projects sup-
use of Earth observation data for land cover mapping and monitoring. His main ported by the European Space Agency and European Commission framework
research interests include characterization of land cover using SAR and inter- programs. At present, his main involvement is in the development of applica-
ferometric SAR data and retrieval techniques of forest biophysical parameters tions and the definition and implementation of related services in land surface
from SAR data. deformation mapping, hazard mapping, land use mapping, and topographic
mapping.