is performed in order to eliminate speckle noise and to
make the data more feasible to handle. It is performed by
block averaging of the full resolution data within a moving
8 by 8 pixel window as follows:
DN400m =Sart( Z(DN42.5m)2/n ). [1]
Where n is the number of pixels in the window. The radar
texture, o72 , is computed in the same step as
O12 = [N(0400m/H100m)2-1]/(n-1)(1+1/SNR)2 [2]
where: woom and G400m are the mean and standard
deviation of the pixel values within the window, n is the
number of pixels within the window (64) and SNR is the
signal to noise ratio in the window. This results in a
texture image of the same size and coverage as the 100
meter low resolution scene computed in the same step.
2.4.3 Land cover classification: Next follows a
classification effort on the low resolution scenes by using
a modified Maximum-Likelihood classification algorithm
developed by JPL (Freeman et a/., 1995). The algorithm
utilizes the 100 meter resolution scenes and texture
scenes as input. Training areas are selected from scenes
where ground truth measurements were performed during
acquisition. The output will be a coarse land cover
classification with the following tentative classes:
1) Forest (e.g. primary forests, tree plantations and
secondary stands more than a few years of age)
2) Non-forest (e.g. clear cut areas, pastures, savannah
and other areas with low vegetation)
3) Flooded vegetation (e.g. flooded forests, flooded
grass, mangrove)
4] Urban areas
5] Open water
6] Unclassified
It is clear that a classification effort of this magnitude of
several thousands of scenes will be a generalization of
reality and certainly contain errors, but in order to
minimize large errors, random verification of the results
will be made by JPL, UCSB and NASDA.
2.4.4 Generation of 100 meter regional
mosaics: The 100 meter low resolution images
generated by JPL are then sent to NASDA Earth
Observation Research Center in downtown Tokyo for
mosaicking. Software for automatic mosaicking of JERS-
1 data is currently under development and expected to be
in operation around August 1996. The software utilizes
orbital information obtained from the SAR leader file to
compute the approximate position for the scene in the
mosaic, followed by automatic matching on overlapping
areas of neighboring scenes for fine adjustment. Given
as an input parameter, the program can generate mosaics
in any preferred projection.
Although the size of each image is small, less than 1 Mb,
making one large 100 meter mosaic canvas covering the
whole Amazon basin and surroundings would not be
feasible. The mosaics will therefore be generated in
regions, each mosaic covering an area of approximately
596
5° by 5°, or some 80 - 100 scenes. This will result in ap
estimated number of 30-35 regional mosaics per season,
2.4.5 Generation of 400 m semi-continental
mosaics: 400 meter resolution mosaics will be
generated by JPL by using the 100 meter resolution
regional mosaics above as input. The 100 meter mosaics
will be downsampled by block averaging to 400 meter
pixel size regional mosaics. These will then be mosaicked
together to form one large semi-continental scale canvas
per season, giving a unique overview of some 8 million
km2, including its inherent seasonal characteristics.
3. CENTRAL AND WEST AFRICA
3.1 Dual season study and comparison with
ERS-1 data
Given the importance of large flooded environments on all
from local to global scales, it is naturally of interest to
investigate the seasonal patterns of the second largest
tropical river basin on the Earth, the one of the Congo
river. In contrast to the relatively homogeneous changes
between dry and rainy season in the Amazon basin
however, the seasonal patterns of the Congo river basin
are more diverse. There is an invisible but distinct border
along the Equator which divides the climate into an
independent southern and northern hemisphere. The
Congo river intersects the Equator twice so one part of
the basin may experience dry season while the other may
be in the middle of the rainy season. For monitoring of
seasonal characteristics in the region, it is thus not
possible to adjust the satellite acquisitions so easily to hit
the dry and rainy seasons, respectively, as was the case
in Amazonas. The problem is yet less complicated than it
appears since what the JERS-1 L-band SAR is sensitive
to, and what we are interested in, is the temporal and
spatial distribution of flooding rather than the rain that is
causing it. Although perhaps less pronounced, the Congo
river also undergoes an annual cycle of high and low
water which we need to consider when choosing the time
periods for data takes.
Figure 2. GRFM coverage over central and western Africa
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
