International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004 
  
  
RMSE | bias | RMSE RMSE 
[m] [m] | without F(slope) 
bias 
  
DSM: all points 10.4 -4.] 9.5 92+70%tana 
  
  
DSM: 9.0 -2.8 8.6 7.9:5:50* tana 
open areas 
DSM: forest 16.5 -11.8 11.5 17.7 
  
DEM: all points 6.5 -0.3 6.5 6.0 + 4.4 * tan a 
  
  
  
DEM: 5.9 0.2 5.9 53+2.4*tan a 
open areas 
DEM: forest 10.9 -5.1 9.7 94+33 *tana 
  
  
  
  
  
  
Table 4: analysis of test area Vilsbiburg located in the northern 
model 
5.3 Test area Inzell 
The test area Inzell is quite different from the other parts. With 
heights ranging from 610m up to 1681m it includes parts of the 
Alps which are covered by forest. Only 32% is belonging to the 
open areas. For this steep terrain the reference DEM is mainly 
based on the topographic map 1:10 000 with a vertical accuracy 
in the range of 5m. A small steep isolated mountain has not 
been matched, so it is automatically not in the analysis because 
of the gap in the achieved DEM. 
  
RMS | bias | RMS 
[m] [m] | -bias 
RMS F(slope) 
  
DSM: all points 172 | -03 | 138 | 107 36.1 xtena 
  
DSM: open area 9.8 -4.7 8.7 6.1 * 39.6x tana 
  
DSM: forest 199 | -130 | 150 17.9 + 8.5 x tan a 
  
DEM: all points 13.7 | -7.1 11.7 | 7,9+45.7xtana 
  
DEM: open area 6.7 -2.5 6.2 | 4.8- 26.1 x tana 
  
  
DEM: forest 17.4 | -10.8 13:7 16.2+74xtana 
  
  
  
  
  
  
Table 5: analysis of test area Inzell 
The larger root mean square values of this area can be 
explained by the landscape, but also by the reference DEM. If 
the influence of the reference DEM with SZ — 5m is respected, 
for the open area after filtering the accuracy value would be 
reduced from 6.7m to 4.5m, that means to the same range like 
before. This is realistic because the open areas are mainly not so 
steep. The limited accuracy of the reference DEM can also be 
seen at the morphologic character, the reference DEM is too 
smooth for this type of terrain. The matched DEM shows more 
realistic structures of the mountains. A stronger dependency 
upon the terrain inclination also may be caused by this and a 
limited horizontal accuracy of the reference data. 
6. CONCLUSION 
The orientation of the SPOT HRS stereo models has not caused 
problems. The orientation accuracy is sufficient, shown at the 
RMSE of 3.5m and 3.9m for the Z-control point coordinates. 
Because of the high number of control points, the influence of 
the orientation accuracy to the finally generated DEM is 
limited. 
The advantage of imaging corresponding scenes from the same 
orbit can be seen at the high quality of the automatic image 
matching. Only in dense forest areas where the grey value 
distribution is limited, the matching had difficulties. A problem 
444 
exists in the very steep parts of the Alps, but this is a general 
case for matching with all optical images. Also InSAR is not 
better in such areas. 
By automatic image matching the height of the visible surface 
and not the bare ground will be determined. This is again a 
problem for all optical images, but also for short wave InSAR. 
The points not located on the ground have to be removed by a 
special filter method like used in program RASCOR. This is 
eliminating approximately 45% of the points with a clear 
improvement of the final results visible at the statistics, the 
distribution of the discrepancies but also the 3D-views of the 
DEMs. 
For flat terrain in open areas after filtering a standard deviation 
for the height between 3.8m and 5.3m has been reached with 
the mean value of 4.5m. If the bias is respected, the mean value 
is reduced to 4.2m. In inclined parts, a factor up to 26.1m 
multiplied with the tangent of the terrain inclination has to be 
added, but with the exception of the very steep area, the 
dependency upon the slope is only in the range of up to 3.8m 
multiplied with the tangent of terrain inclination. In the forest 
areas after filtering several points are still located on top of the 
trees. 
The standard deviation of 4.2m up to 4.5m for the height 
corresponds to a standard deviation for the x-parallax of 0.7 
pixels. This final result is confirming the high accuracy level 
which can be reached with SPOT HRS-stereo models. Under 
operational conditions better results cannot be expected from 
optical sensors. The main limitation is vegetation covering the 
bare ground. Its influence can only be respected and removed if 
at least few ground points with a sufficient small distance are 
available. So the height information in dense forest can never 
reach the quality like in open areas. 
As advantage of the DEMs generated by automatic image 
matching against DEMs based on Interferometric Synthetic 
Aperture Radar (InSAR) we do have the quite higher details 
and also a smaller spacing. 
REFERENCES 
Jacobsen, K. 2001: New Developments in Digital Elevation 
Modelling, Geoinformatics, June 2001, pp 18 — 21 
Passini, R., Betzner, D., Jacobsen, K. 2002: Filtering of Digital 
Elevation Models, ASPRS annual convention, Washington 
2002 
SPOT Image 2002: SPOT Satellite Geometry Handbook, 
http://www.spotimage. fr 
  
  
  
    
   
    
  
   
    
   
      
   
      
   
     
   
    
     
    
     
   
    
   
    
      
     
   
       
     
   
    
      
    
     
    
   
   
     
   
    
   
   
    
    
   
    
   
     
   
   
  
  
KEY 
ABS 
This 
. Scop: 
Inter 
adju: 
Heig 
urbai 
trans 
supe 
regic 
appr 
matc 
Allr 
SPO 
track 
tiltec 
stere 
the 
cont 
with 
with 
HRC 
band 
band 
grou 
enab 
imag 
251 
char 
imag 
case 
The 
Barc 
CNE 
the 7 
orbit 
each 
data 
Lir 
and 
(lo 
test 
radic 
THR 
HRS