'ook* (SPOT 
hemeris data 
mmended to 
| line. After 
TS) with a 
'e scene, this 
r orthoimage 
in the Earth 
ume. 
image space 
described in 
e resolution 
two forms: 
ccuracy and 
LSQM) for 
s 0.1 to 0.3 
ited with the 
arched for in 
| and quality 
Jstment (see 
the usage as 
on growing 
). This local 
S of 13 x 13 
pixel in each 
is performed 
Müller et al. 
CURACY 
accuracy of 
iation of the 
djustment or 
an already 
the interior 
(ers of the 
ted from the 
se of Bavaria 
pal. ERS 172 
10 meter in 
inous terrain 
part of the 
ased on the 
on (pointing 
collinearity 
object space 
ntric system 
ection of the 
1 parameters 
0 m grid has 
any ground 
ig 20 of the 
the quality 
ied out In 
el accuracy. 
tion for the 
stanbul 2004 
  
Table 1: Mean values and standard deviations for the 
difference to the orthoimages of 20 ground control points in 
meter in Gauss-Kriiger coordinate system (Bavaria) 
xl, yl — Coordinates of ground control points 
x2. y2 — Coordinates in orthoimage from forward looking 
x3. y3 - Coordinates in orthoimage from backward looking 
  
  
  
  
  
  
  
x2-x1 |y2-y1 |x3-x1 |y3- y1 
MEAN -4,3 5.0 -14.3 11.5 
Std. dev 5.89 2.35 6.23 8.64 
  
  
The result shows that even without any ground control, the 
absolute georeferencing accuracy of the HRS sensor is better 
than 20 meter and standard deviation less than one pixel. This is 
expected, since the values for the absolute pointing accuracy is 
given by (Bouillon et al. 2003) to about 33 meters with 90% 
accuracy. More detailed analysis can be found in the conference 
paper by Müller et al 2004. 
6. DEM PRODUCTION FROM TWO RAY STEREO 
DATA 
Having the mass points from the matching process as well as 
the exterior and interior orientation of the camera system, the 
object space coordinates can be calculated using forward 
intersection. This is done by least squares adjustment for the 
intersection of the image rays. The irregular distribution of 
points in object space after the forward intersection is 
regularized into a equidistant grid of 15 to 50 meter spacing. 
The interpolation process is performed by a moving plane 
algorithm (Linder 1999). The resulting DEMs, which are 
surface models, are compared to the reference DEMs, which are 
terrain models. Therefore a distinct difference is expected e.g. 
in forest areas. 
In the test area of Bavaria six reference DEMS are available for 
testing the accuracy (see Fig. 1). Fig. 2 shows area #6 (size: 
50 km x 30 km) east of Munich with moderate terrain, which is 
the largest of the six test regions. The DEM calculated from 
HRS data for this area is shown in Fig. 3. 
  
  
  
  
Figure 2. Part of the test area showing the region of 
reference DEM #6 (50 x 30 km) 
The comparison of the derived DEMS to the reference DEMS is 
performed in several ways. At first only those points, which are 
found during the first matching process (Lehner 1992), and 
therefore are highly accurate homologous points, are 
investigated. They are compared for all the areas where a 
reference DEM is present (see fig. 1). 
  
    
    
   
  
   
   
    
     
   
    
   
   
     
   
   
   
   
    
   
     
   
   
   
   
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part Bl. Istanbul 2004 
  
  
  
Figure 3. DEM from SPOT-HRS stereo data (region in fig 
2). 
The result is shown in table 2. The mean height differences are 
due to absolute orientation errors, they seem to be very similar 
for all reference areas, and can be eliminated using bundle 
adjustment methods (see chapter 9). The low standard deviation 
shows a very good agreement with the reference DEM. A 
second comparison is performed in using the regularized SPOT- 
DEM to perform an area oriented analysis. 
Table 2: Comparison of height for high quality homologous 
points in SPOT-DEM and reference DEM 
  
  
  
  
  
  
  
  
Reference Size and Mean Height | Std. | Points 
area Accuracy of Ref- | Difference |Dev.| [#] 
DEM [m] [m] 
DEM-01, Prien | 5x5km, 0.5m 6.8 2.01 240 
DEM-02, Gars | 5x5km, 0.5m 6.2 2.2 184 
DEM-03, 5x5km, 0.5m 5.6 1.8 261 
Peterskirchen 
DEM-04, 8x5 km, 0.5m 4.9 2.0 254 
Taching 
DEM-05, Inzell | 10x 10 km, 5m 5.7 3:51 458 
DEM-06, 50 x 30 km, 2-3 m 6.1 3.6 / 15177 
Vilsbiburg 
  
  
  
  
  
  
The area oriented approach should distinguish between at least 
two types of classes (forest and non-forest areas) because of the 
anticipated discrepancy between terrain models and surface 
models. The matched objects inside a forest area are distributed 
among different height levels and therefore the standard 
deviation for their heights should be higher. Table 3 shows the 
result for two of the reference areas in Bavaria. 
The mean height differences are of the same order (around 6 
meter) as for the single points in table 2. The standard 
deviations are much higher in this area due to lower matching 
accuracy of the densely matched points and due to interpolation 
errors in areas where the region-growing matching algorithm 
could not find enough well correlated points (e.g. low contrast). 
In the forest areas the differences are about 12 meter higher, 
what is due to the surface/terrain model discrepancy. Also the 
standard deviation is much higher in forest areas as was 
expected. 
There are many filtering techniques which can be applied to the 
DEM data. For this paper two techniques have been applied: an 
analysis of the statistics of correlating points (kernels) such as 
variance and roundness (Fórstner operator), and a median filter 
with a window size of 3 x 3 pixel. Table 3 shows that in all 
cases the filtering leads to significantly lower standard 
deviations and only little change in absolute differences. The