of the seven data sets. The results for the other X-Band images (data sets 710, 
711, 714, and 715) were of comparable accuracy while the L-Band images (data 
sets 712 and 713) were found to be somewhat worse due to the lower resolution 
and poorer contrast. 
These affine transformations were then used to register each of the radar images 
to the map. Beginning with the appropriate affine transformation, the digital 
resampling procedure employed was to step through the output image (i.e., each 
pixel of the map grid) and at each point calculate the corresponding radar image 
coordinates. The data sample nearest to this point in the input data (radar 
image) was then placed in the output image. (Note that although digital de- 
convolution is sometimes used to correct for the point-spread function of the 
sensor when resampling multispectral scanner data, no counterpart exists for 
overcoming coherent speckle in SAR data). The image map transformation 
operation was done afresh for each of the SAR data sets (i.e., a different set 
of coefficients was employed). To make optimal use of the SAR resolution and 
dynamic range, all digital analyses were conducted on the data sets in their 
original coordinate system. After processed images had been produced, the 
results were transformed to map coordinates for visual display. Here, in order 
to facilitate the computational burden, the data set was down-sampled by a 
factor of two in both directions leading to 3 m x 3 m grid system. This could 
be justified since the measured IPR for X-Band in slant range was 2.3 m 
(corresponding to a ground range resolution of approximaterly 2.8 m). Figures 
3a-c show the result of this resampling operation on X jy imagery from each of 
the three flight directions. 
According to the field-by-field ground truth which was collected on the same 
day as the SAR overflights by the SWISSAR Investigation Team (Dr.F.Fasler, 
Dr.K.Itten, Dr.P.Meylan and Dr.K.Stänz), the large dark fields near the center 
of these images contain wheat while the lighter colored large fields contain 
corn. The slender fields on the rigth range from 8 to 35 meters in width and 
provide examples of barley, grass, sugar beets, and potatoes. The geometrical 
effects of viewing this scene from three different directions is seen most 
dramatically in the appearance of the row of trees which runs from north-south 
in the right-center of the scene. In data set 709 with the view direction down 
the tree line the tree crowns are primarily responsible for the SAR returns 
(little shadow effect). On the other two images (data sets 710 and 711) 
acquired on passes perpendicular to 709, the tree line and its shadow alternate 
positions. 
DATA PREPROCESSING 
Because SAR imaging is done by a coherent sensor (Ref.3), images of relatively 
homogeneous areas, such as agricultural fields containing a single crop appear 
"speckled" (see Figure 3) although panchromatic aerial photographs of the same 
areas would appear relatively uniform in intensity. The coherent nature of these 
data makes it difficult to successfully adapt statistical pattern recognition 
methods for use on SAR data. The presence of coherent speckle in SAR imagery 
often precludes a pixel-by-pixel classification unless the data are highly 
smoothed. While previous researchers usually applied some sort of averaging 
procedure, I wished to avoid simply taking an average over a window of size 
2 x2, 3 x3, etc. While such averaging will reduce the effects of coherent 
speckle , it also visibly degrades edges (field boundaries in our case). I 
chose instead to apply a non-linear isotropic filtering.algorithm using 
neighborhood processing as implemented in ERIM's CytocomputertM (Refs.4-5). 
234 
EE REL S... 5 9 AA e -— s > DA 
Rs ICO 7 
  
Figu 
Data 
voly 
mari 
requ 
achi: 
adval 
form 
Revie 
supp 
repre 
I se 
Init: 
conte 
in ii 
set c 
freq 
in a 
gray 
The c 
the ' 
in or 
lay b 
oc=1 
the m 
a mov 
dz 2 
(scal 
each 
which 
avera 
four 
movin 
is ca 
finit