2. DESCRIPTION OF THE TEST-SITE, REMOTE
SENSING DATA, AND GROUND TRUTH
Within a relatively small area, the Freiburg-Black Forest test-site
in Southwest Germany represents in a unique manner landforms
as well as geographical and climate units, typical for temperate
Central European landscapes. The study area (about 30x30 km)
stretches from France over the fertile Rhine valley with its sub-
mediterranean climate and its variety of agricultural crops, vine-
yards and forests, to the city of Freiburg in the centre of the area.
From there it passes the western slope of the Black Forest,
mostly covered with forest.
ERS-1 acquisitions (14 Single Look Complex (SLC) data sets
delivered by the D-PAF, or by ESA/ESRIN) from the commis-
sioning phase in 1991 and from the multidisciplinary phase in
1992/93 were selected to cover a variety of seasonal and there-
fore phenological stages and different meteorological conditions.
All data were acquired at 10h20 GMT during ascending passes
of the ERS-1 satellite. SPOT/XS from September 12, 1991 were
also available, and have been used for the present study.
Meteorological data (e.g. precipitation, temperature, relative
humidity) were compiled continuously from July 1991 onwards,
by three meteorological stations of the German Meteorological
Service. A digital elevation model (DEM) from the German
Geodetic Survey was also available for the whole area.
Data analysis was made by selecting a wide range of well docu-
mented ground samples:
The analysis of multitemporal ERS-1 slant range SAR data
was conducted using 40 test areas (forest, agriculture, grassland,
built-up areas, water bodies). All test areas, each greater than 2.5
ha, were selected on flat terrain within the test site.
After geocoding, the multisensor (ERS-1, SPOT) analysis
was performed using a forestry GIS from the inventory used for
regular forest taxation in 1990 for a forest district in the Rhine
valley and on the western slopes of the Black Forest. This data
base covers about 6000 ha of forest including stand descriptions,
species composition and age class. 960 forest stands (polygons)
located on flat terrain were considered for data analysis.
3. PROCESSING CHAIN FOR ERS-1 DATA AND
MULTISENSOR DATA FUSION.
The overall processing chain for ERS-1 (SLC) SAR data detailed
below is especially designed to monitor the changes occurring to
the scene at a high spatial resolution (Kattenborn et al., 1993). In
developing this processing chain, emphasis was put on efficiency
in terms of preservation of radiometric quality and spatial reso-
lution, computation time, and data storage.
ERS-1 SAR data coregistration: For the first step of analysis,
coregistration of the ERS-1 data was performed by shifting the
frames in range and azimuth.
Data calibration: Changes in calibration constants introduced in
the ESA/ESRIN and D-PAF ERS-1 SAR processors, on April
6th, and November 15th, 1992 were taken into account. Recent
work shows that uncertainties in ERS-1 calibration are within
x0.8 dB (Lavalle, 1993). Antenna pattern and range spreading
loss corrections (Laur et al., 1993) were also performed.
332
Spatial multilooking: The multilooking operation is done sp.
tially by averaging the intensities of 5 consecutive pixels in the
azimuth direction, then converting the resulting pixel value ty
amplitude by taking its square root. An overlapping of 1 pixel in
azimuth is introduced, in order to preserve thin features present
in the 1-look SLC data. The final equivalent number of looks
(ENL) after this operation is L=4.8 looks in the resulting image,
The pixel sampling is then approximately 16x16 meters, with 5
spatial resolution comparable to that of ERS-1 PRI data, but
with a much better signal to noise ratio.
Data compression: Data compression is made by storing the
multilooked image on linearly rescaled 8-bit amplitude data. The
scaling factor, which is kept for further treatment in order to con-
serve data calibration, is determined using global statistics on
strong scatterers. This way, saturation occurs only for very
strong scatterers, mainly located within the urban areas which are
not of interest for our study. Given the low values of these scal-
ing factors (of the order of 1.5), the loss of radiometric accuracy
during this operation is negligible within natural areas.
Restoration of the radar reflectivity: The next processing step
consists of adaptive speckle filtering by means of the feature re-
taining Gamma-Gamma Maximum A Posteriori adaptive speckle
filter, using an 11x11 processing window size for structures de-
tection and an 9x9 processing window size for speckle filtering
(Lopes et al., 1993). This operation allows a drastic speckle re-
duction from L=4.8 looks to an ENL of about L=300 looks. It
restores the radiometric information with an error not exceeding,
with 90% confidence level, +0.35 dB in homogeneous
(textureless) areas of the scene.
Geocoding and coregistration of ERS-1 SLC and SPOT-XS
data: Geocoding of the ERS-1 SLC data was performed using
orbit parameters provided with the ERS-1 data and the DEM
with a resampling to SPOT pixel size of 20x20 m. Using a set of
ground control points, the accuracy of geocoding was estimated
to be 21.8 m in range and 9.8 m in azimuth direction, i.e. about
one pixel in the map projected images. The SPOT data had been
delivered in map projection (ortho-image) by ISTAR Ltd.
4. CHANGES IN RADAR BACKSCATTER WITH
ENVIRONMENTAL PARAMETERS
Since some observational data suggest that the variability of
SAR backscatter is related to changing environmental conditions
(Dobson et al., 1991, Moghaddam et al., 1993, Way et al, 1993,
Pulliainen et al., 1993), a statistical analysis was undertaken by
calculating correlations between mean backscatter of test areas
and meteorological parameters. In table 1, the linear correlation
coefficients between precipitation measurements and mem
backscatter values of three vegetated areas (grassland, agricul-
ture and forest) and an urban area are presented. Precipitation
measurements for time frames of 2, 4, 6, 8, 10 and 20 days be
fore each ERS-1 acquisition were averaged, in order to character.
ise wet and dry periods before the acquisitions. This should allow
to study the influence of varying moisture conditions 0!
backscatter behaviour of different kinds of landuse classes.
These correlations show that the vegetated areas seem to rea
generally with an increase of backscatter (coefficients of correla:
tions around 0.8 for precipitation averaged for 20 days) whereas
the non-vegetated urban area shows no significant correlation
with precipitation. There is possibly also a difference in the m“
response of backscatter to increasing available moisture, as C
be seen from the higher correlation of agriculture for average
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
e adu ma lian ld uni t i seh ar
