5. METHODS 
5.1 -L-band SAR data pre-processing 
Each JERS-1 scene was submitted to a spatial filtering for 
speckle removal. Previous studies (Shi and Fung, 1994) 
recommend the use of adaptive filters because they smooth the 
image without removing edges and sharp features. In this 
study a spatial filter based on Lopes et al. (1990) algorithm 
and known as FGAMMA filter was used. The algorithm 
assumes a gamma distribution for the image and performs 
spatial filtering on each individual pixel using the digital 
number in a square window surrounding each pixel. In this 
study a 3 by 3 pixels window was used and provided a good 
result. 
After the speckle removal, the JERS-1 scenes were mosaicked 
and then geometrically corrected using TM/Landsat image as 
reference. This image had ben previously ortho-rectified using 
the method developed by Toutin (1995). For further discussion 
on this subject refers to Costa (1995). 
The geometrically corrected image was resampled to a 
resolution of 12 m by 12 m using the 8 point Sin (x)/x method 
as suggested by Shlien (1977). This algorithm determines the 
digital number from the weighted average of 64 closest pixels 
to the specified input coordinates and assigns the value to the 
output coordinates. The resulting image is sharper than that 
obtained by bilinear interpolation. 
5.2 C-band SAR data pre-processing 
The antenna pattern correction applied to the mosaic produced 
by the Canadian Centre for Remote Sensing was not sufficient 
to remove the range effect on the surface backscatter. To 
eliminate this effect, a correction factor was applied to the 
data as follows: 
e selection of a homogenous 50 lines swath along the range 
direction; 
e computation of a average correction factor applied to 
every pixel along the range direction. 
This procedure also accounted for most of the image speckle 
what prevented further filtering. 
The next step was to register the C-band SAR image onto the 
L-band SAR image using the same resampling algorithm to 
produce a 12m by 12 m pixel image. 
5.3 - Sample Acquisition 
The sample acquisition of the different targets was done using 
visual interpretation of aerial photographs and the Landsat 
image. Masks were created to sample areas corresponding to 
the classes of interest in each image. The masks were used to 
compute the mean digital number, the standard deviation of 
each class. 
5.4 - Digital number normalization 
Because of the uncertainty of the absolute calibration of the C- 
Band SAR data available (Costa, 1995), it was decided to use 
the digital number for both data set. The data were converted 
submitted to a normalization. 
For the normalization procedure, it was assumed that: a) for a 
calm open water surface the backscatter from both wavebands 
should be low and equal; b) calm open water would present 
the lowest digital numbers in both images; c) the ratio 
between the average digital number of open water at C band 
and L band would provide a normalization factor which would 
allow a quantitative analyses of the macrophyte backscattering 
properties. 
5.5 - Generation of SAR multiband composition 
A series of contrast stretch functions (linear, equal and root) 
were tested to produce multiband color composites using the L 
and C band images. With aid of ground information it was 
selected the best combination of color and contrast to 
maximize the visual discrimination among the macrophyte 
genus and the visual discrimination between the terrestrial and 
the aquatic ecosystem. 
6. RESULTS 
6.1 - Multiband Composition 
Figure 3 shows : a) the multiband composition of L band 
image displayed as red (R) and C band image displayed as 
cyan (B and G), both submitted to linear contrast stretch; b) 
the C band and L band images submitted to linear contrast 
SH n; c) L band image submitted to squared root contrast 
stretch. 
The most obvious difference between the SAR images is the 
low return from the macrophyte stands displayed in the L 
band. As seen in figure 2, the reservoir water level was 71.50 
in March 7 and 71.65 in April 15. This difference in water 
level (15 cm) is not sufficient to affect the area covered by 
macrophytes. Therefore, the differences in the return can be 
explained by the larger L band penetration depth which 
exceeds the aquatic plant canopy height (1m in average). As a 
result most of the radiation interacts directly with the water 
surface being reflected in the forward direction. At C band, 
the penetration depth is smaller allowing for multiple 
scattering within the canopy and a stronger backscatter. 
The multiband composition clearly shows the area covered by 
macrophyte stands in cyan in the lower reach of the Pucuruí 
inlet. Towards upstream, where the macrophyte stands are 
thicker, an increase in L band backscatter can be observed. 
The limit between the terrestrial and the aquatic environment 
is more evident in the multiband composition than it is in the 
individual bands. The low incidence angle of the C band casts 
long shadows which prevent setting a precise boundary 
between land and water. Besides that, the high backscatter 
from the thicker stands makes it difficult to set the limit 
between the aquatic vegetation and the terrestrial vegetation. 
In L band, the limit between land and water is clear, but the 
limit between open water and macrophyte stands is not as 
evident as in C band. 
In figure 3 c one can observe the L band image submitted to 
Square root contrast stretch. In this case, the macrophyte 
stands are enhanced and can be mapped. It indicates that, the 
L band is perhaps more sensitive to the stands height and 
density than the C band. The combination of both wavebands 
can highlight differences in stand density, biomass and dossel 
Structure as seen in more detail in figure 4. 
The differences in waveband sensitivity of canopy height and 
density can be better observed in the graphic of figure 5. The 
first striking feature is that, as expected from the theory, the 
normalized digital number (DN) for the Forest class is much 
higher in L band than in C band. These results are in 
agreement with the backscatter statistics provided by Dobson 
et al. (1995). Using data from the SIR-C/X- SAR experiment 
in which L band and C band data were acquired under equal 
image parameters (look angle, pixel spacing and number of 
looks) for boreal forest. The authors related total dry biomass 
to image backscatter. For a constant biomass of 10 kg m* , the 
average backscatter for L band was around -8 dB, whereas for 
C band, the average value was around -10 dB. 
194 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996