to-5m 
o<5m 
9 15m 
  
to-5m 
o«5m 
)15m 
  
  
ınbul 2004 
  
    
  
  
   
  
  
  
  
  
  
  
  
  
  
   
   
   
  
  
  
   
  
   
   
   
   
   
   
  
  
   
   
  
    
    
   
   
   
  
  
  
  
  
   
   
   
   
  
  
  
  
  
   
   
     
   
   
   
    
       
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part Bl. Istanbul 2004 
  
  
  
  
Figure 4: Detail of multi-sensor stereo data (left: THR, right: 
HRS) for built-up area (top) and rural area (bottom). 
This is confirmed in Table 6, which shows the performance of 
image matching for the selected test areas. Now, more than 
10% of the pixels were not matched for all of the three test 
areas. In a relative sense, however, the matching performance 
of the mountainous area was not as drastically degraded as for 
the rural or the urban test site. This is due to the fact that the 
stereo images now are more similar even in the mountainous 
areas and image matching is facilitated, although on the other 
hand the stereo condition was significantly degraded by a factor 
of 7. 
  
  
  
  
  
Rural area 10,07 % 
Mountainous area 12,25 % 
Urban area 13,78 % 
  
  
Table 6: Matching performance for THR-HRS stereo pair. 
The digital surface models which have been generated from 
these stereo data are shown in Figure 5 together with the 
difference DEMs, which were determined with respect to the 
given reference DEM. Statistical parameters like mean, 
standard deviation, minimum and maximum of these elevation 
differences are summarized in Table 7. The following 
conclusions can be made: 
Rural area: The extension of unreliable areas, specifically 
represented by larger negative height differences, was 
significantly reduced, leading to an increase of the bias to 4 
meters, which may realistically be caused by the forest 
areas (yellow and red areas in difference model). 
Mountainous area: The maximum elevation errors are 
drastically reduced, although large elevation errors of some 
150 meters are still locally present. The standard deviation 
is reduced to about 9 meters, while the bias was increased 
to 5 meters. Although not really clear, this could again be 
due to vegetation and forests, the surface of which should 
have been tentatively reconstructed. 
Urban area: The surface model clearly shows the road network 
of the city of Barcelona. Elevation differences in the built- 
up areas are typically in yellow, i.e. in the order of 5 to 15 
meters and correspond well to the potential height of 
buildings. Hence, also a bias of 11.4 meters is achieved for 
this test area. 
For each of these test areas the standard deviation corresponds 
well to the 8.4 meters RMS error which has been achieved in 
the a-priori analysis based on control points (see Table 2). For 
visual quality control again stereo ortho photos were generated. 
These are shown in Figure 6 in an anaglyph presentation for the 
mountainous and the urban test area. Again, a significant 
improvement can be immediately notified for the mountainous 
area, although major elevation errors still cause geocoding 
errors, which are well visible in the stereo ortho photos overlay. 
  
Model Area Mean | Std.Dev. Min. Max. 
  
  
  
  
2 4.0 6.9 -31.0 44.4 
THR-HRSI 4 5.0 8.8| -172.0 156.0 
6 11.4 09 -29.0 61.4 
  
  
  
  
  
  
Table 7: Summary of elevation difference statistics for 
investigated test cases. 
5. SUMMARY AND OUTLOOK 
A Spot 5 image data set acquired over the city of Barcelona was 
used to investigate the accuracy of 3D data being extracted 
stereoscopically. Stereo modelling using high quality control 
points has shown a height accuracy of some 4 meters for the 
HRS stereo pair, while the planimetric accuracy was worse by a 
factor of 2. When using a multi-sensor THR/HRS image pair, 
the planimetric accuracy can be improved to less than 3 meters, 
but the height accuracy is degraded by a factor of 2. Surface 
models were extracted from HRS image pairs as well as from a 
THR/HRS image pair for different type of terrain. However, a 
comprehensive and thorough quality analysis is hardly possible 
for vegetated and built-up areas, because only a ground model 
but no surface reference data are available. Future work will 
focus on the utilization of an image triple comprised by the 
HRS stereo images as well as the THR scene. This promises a 
significant upgrade of achievable accuracies in the order of a 
few meters in planimetry as well as height for any type of 
terrain. 
REFERENCES 
J.P. Gleyzes, A. Meygret, C. Fratter, C. Panem, S. Baillarin, c. 
Valorge, 2003. SPOT 5: System Oveview and Image Ground 
Segment. [EEE International Geoscience and Remote Sensing 
Symposium , published on CD-ROM. 
JOANNEUM RESEARCH, 2003. RSG in Erdas Imagine. 
Software Documentation, Version 2.1, RSG-Release 4.60. 
H. Raggam, M. F. Buchroithner and R. Mansberger, 1989. 
Relief Mapping Using Non-Photographic Spaceborne Imagery. 
ISPRS Journal of Photogrammetry and Remote Sensing, Vol. 
44, No. 1, pp. 21-36. 
H. Raggam and A. Almer, 1990. Mathematical Aspects of 
Multi-Sensor Stereo Mapping. /£EE International Geoscience 
and Remote Sensing Symposium . Vol. 111, pp. 1963-1966. 
H. Raggam and A. Almer, 1996. Assessment of the Potential of 
JERS-1 for Relief Mapping Using Optical and SAR Data. 
International Archives of Photogrammetry and Remote Sensing. 
Vol. 31, Part B4, Commission IV, pp. 671-676. 
Spot Image, 2002. SPOT Satellite Geometry Handbook. Ref. 
S-NT-73-12-SI, Edition 1, Revision 0.