go 
e 
  
with different 
thods, one is 
aximum and 
ows that the 
of the results 
this visual 
give the best 
ilter but the o 
ense of more 
ected, in the 
ie eigenvalue 
'ferent filters. 
nimum errors 
npared to the 
error is very 
, o linear, B 
1e most effect 
is small. It 
h to further 
ler detection 
nid. 
roduced after 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
FILTER Mean Standard Max. 
error error error 
None 2.7 29.6 171 
Lee 1.6 29.0 141 
Kuan 21 30.4 176 
Frost 1.1 29.7 202 
Mod. Frost 0.8 20.7 201 
o linear 22 29.7 135 
MAP 0.9 29.2 162 
B 19 29.0 143 
y 22 29.6 201 
Li 29 30.0 166 
LVN 1.2 28.6 122 
Crimmins 1:8 29.3 169 
log-linear 17 29.2 190 
  
  
  
  
  
Table 2. Statistics relating DEM with disparity model. 
6. RESULTS FROM DEM PRODUCTION 
The final stage of DEM production is the 
transformation of the image disparities into three 
dimensional co-ordinates. This requires a geometric 
model as discused in section 2. Chen (1993) has 
developed the model described by Clark (1991) and 
has produced 3D co-ordinates without the use of 
ground control points from opposite side and same 
side stereoscopic pairs using roll-tilt mode data, of he 
area of Provence, France, to the North of Marseilles. 
The accuracy of the results has been tested with 38 
ground check points. The method makes no use of 
ground control points but requires accurate orbit data. 
To date there have been difficulties in determining the 
accuracy of the orbit data and this may contribute to 
the errors in the results. Table 2 shows the statistics 
for the 38 points used to check the transformation. 
It is apparent that there are systematic errors in both 
sets of data but that they apply to X and Z co- 
ordinates derived from same side data and only in Z 
with opposite side data. An analysis of errors 
indicates that these results are consistent with theory 
and that improved timing data and use of ground 
control points would give better results. 
Leberl (1990) has reported results from stereoscopic 
SAR as shown in table 4. The current results indicate 
that the method with ERS-1 data needs improvement 
to give comparable accuracy. 
  
  
  
  
  
  
Sensor Easting Northing Ht 
(m) (m) (m) 
Seasat 59 34 28 
SIR A 48 84 93 
SIR A 56 77 72 
STAR* 13 26 28 
  
  
  
  
  
*These results are from an airbourne system with GPS 
Table 4. Results from stereo SAR systems quoted 
by Leberl (1990). 
7. CONCLUSIONS 
Geocoding of SAR is now carried out at a routine 
process. Experience has shown that processors 
developed for ERS-1 can also handle JERS SAR data 
but thet there can be problems in obtaining the correct 
timing and orbit data. These may be ovcercome with 
the use of tie points. It is important to include a 
system for validation into the geocoding process and 
this includes operator interaction. The production of 
auxilliaty products such as layover, shadow and 
energy maps is also important. 
The work at UCL indicates that the use of 
stereoscopic SAR data can produce DEMs which can 
be used for geocoding. A SAR post proessing system 
which had provision for stereoscopic height 
determination and segmentation of the image into 
  
  
  
  
  
  
  
Same side pair Opposite side pair 
X(m) Y (m) Z(m) X(m) Y (m) Z(m) 
Minimum -48 -91 -116 -85 -64 -103 
residual 
Maximum 137 93 8 72 45 28 
residual 
Mean 51 6 -42 4 -7 -51 
Standard 70 35 S2 37 23 64 
deviation 
  
  
  
  
  
  
  
  
Table 3. Statistics from 38 ground check points. 
441