Figure 1. Principle of stereoscopic SAR. 
5. SPECKLE REDUCTION FILTERS 
For many applications of SAR data the presence of 
speckle causes severe problems. Because of the nature 
of stereomatching it was assumed that the same would 
be the case for that application, early tests by Denos 
(1991) with Seasat data showed this to be the case. 
Further ests were therefore carried out at UCL 
(Clochez 1993) to determine which filter, out of a 
number described in the literature, was most suited 
for this purpose. 
Some initial tests were carried out on 12 filters in 
order to ensure that there was no serious distortion of 
radiometric values or degradation of edges. In order 
to test the effect of these filters on stereo matching the 
matching algorithm was applied to two overlapping 
sub scenes of PRI data covering an area in southern 
France to the north-west of Marseille. All of the 
scenes were fitered prior to matching. In order to 
evaluate the performance a number of criteria should 
be used: 
coverage - this is assessed by the number of 
successful matches; 
number of blunders; 
accuracy of stereo matching - this is assessed by 
the eigenvalue of the matrix produced by the 
least squares matching; 
distribution of the matches - this can be assessed 
by a visual inspection of the matches; 
accuracy of resulting DEM - which can be 
evaluated by comparing the DEM which is 
derived from the disparities, with the DEM 
derived from other sources. 
The coverage achieved for the filters is shown in table 
1. These results were achieved with 5 tiers of the 
image pyramid, the use of a sixth tier slightly 
improved the coverage but a greater number of points 
were rejected. It can be seen that the best result is 
achieved without any filter but that there is little to 
choose between the filters as regards coverage. 
440 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
FILTER Rejected | Coverage 
points % 
None 239 73.4 
Lee 271 71.7 
Kuan 248 72.6 
Frost 241 67.7 
Mod. Frost 247 69.1 
o linear 255 71.0 
MAP 264 70.7 
B 250 72.4 
y 215 72.3 
Li 251 70.8 
LVN 258 71.0 
Crimmins 264 68.2 
log linear 250 70.8 
  
Table 1. Coverage and rejected points with different 
filters 
The blunders can be assessed by two methods, one is 
visual inspection and the other by maximum and 
minimum errors. Visual inspection shows that the 
number of blunders and the smoothness of the results 
varies considerably and that from this visual 
inspection the MAP filter appears to give the best 
result, however, detail is lost with this filter but the & 
linear filter retains the detail at the expense of more 
blunders. The worst results are, as expected, in the 
areas of highest relief. 
The quality of match, as indicated by the eigenvalue 
also showed very little variation with different filters. 
Table 2 shows the maximum and minimum errors 
when the scaled disparity model is compared to the 
DEM of the test area. The standard error is very 
uniform but it can be seen tht the LVN, & linear, B 
and Lee produce the smallest maxima. 
It can be seen then that the filters have the most effect 
on blunders but that the overall effect is small. It 
would seem that the best approach to further 
development is to improve the blunder detection 
routines within the stereomatching pyramid. 
Table 2 also show that the accuracy produced after 
use of the various filters varies very little. 
  
  
  
  
  
  
  
  
e 
s 
Table 2. 
6. RI 
The fir 
transforr 
dimensic 
model a 
develope 
has proc 
ground c 
side stere 
area of P 
The accu 
ground c 
ground c 
To date t 
accuracy 
the error: 
for the 36 
  
Minimun 
residual 
Maximur 
residual 
Mean 
Standard 
[deviation 
Table 3.