1104 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008 
multi-image processing of Linear Array CCD images, the 
approach is also applicable to single frame aerial photos (analog 
and digital). 
To improve the matching results, the software pre-processes the 
images using a combination of an adaptive smoothing filter and 
the Wallis filter. The first filter reduces the noise level, while 
sharpening edges and preserves even fine details such as 
comers and line end-points. The Wallis filter is used to strongly 
enhance the already existing texture patterns. After pre 
processing and production of the image pyramids, the matches 
of three kinds of features, here feature points, grid points and 
edges, on the original resolution images are found progressively 
starting from the low-density features on the images with the 
low resolution. Matching with edges and feature points alone 
may provide under certain conditions very sparse matching 
results. Therefore, SAT-PP uses also uniformly distributed grid 
points for matching. This force for a “must match” and a 
possible resulting blunder can be avoided by using only the 
feature points and lines. To get an analysis for the whole area, 
we used both features and the grid points. 
After determine the 3D coordinates of the features and grid 
points, a TIN is reconstructed using the constrained Delauney 
triangulation method. Therefore the evaluation of the DSM 
quality contains also the interpolation error. 
Optionally a least squares matching can be used as a modified 
MultiPhoto Geometrically Constrained (MPGC) matching 
method for point matching while a Least Squares B-Spline 
Snakes (LSB-Snakes) is used for matching edges, which are 
represented by parametric linear B-Splines in object space. For 
ALOS7PRISM images with a GSD of 2.5 meters and the 
radiometric problems mentioned above, we got no significant 
improvement using the time consuming MPGC. Therefore we 
have chosen here a less complex modified cross-correlation 
matching. 
Water areas like wide rivers and lakes can be defined in SAT- 
PP as so-called dead areas. Also, disturbing objects like clouds 
or image artifacts can be defined likewise and excluded from 
matching. 
4. RESULTS 
4.1 Testfield Bern/Thun - comparison of the results of 
stereo pairs (FN, FB) and the image triplet (FNB) 
For the comparison of the use of all three available views FNB 
or just a combination of two of them, we came back to an 
already used testfield Bem/Thun which is the area between the 
two Swiss cities Bern and Thun. The area contains beside the 
two cities different terrain types like a mountainous region in 
the southern part, smooth hilly regions, open areas, forests, two 
rivers and the Lake of Thun. The testfield in its current form 
and the GCP field was set up by our group under a contract 
with JAXA. For the orientation of the image triplet we 
measured GCPs all over the covered area. For the validation of 
the DSM generation we generated three sub areas as reference 
data by using aerial images of scale 1:25 000 to 1:34 000 and 
our software package SAT-PP. We have the following three sub 
areas: 
• Bern: city and tree area, max. height difference 400m, 
• SW (south west): mountainous and tree area, max. 
height difference 1500m, 
• Thun: city, open and tree area, max. height difference 
1000m. 
We excluded rivers and significant lakes as dead areas without 
any given height and they were not used for the evaluation 
procedure. The expected accuracy of the reference DSMs is in 
the range of 0.5 m to 2.5 m and is therefore by a factor 5 better 
than the expected PRISM matching results (see Gruen et al., 
2006). 
For the generation of the DSM with SAT-PP we first 
determined the parameters of the rigorous sensor model and in a 
second step we used these parameters to generate Relational 
Polynom Coefficients (RPCs). The Root Mean Square Error 
(RMSE) of the triangulation is in planimetry and height below 
one pixel for the Direct Georeferencing Model (DGR) and by 
this sufficient for a DSM generation with a grid size of 5m. 
For the comparison of the results of stereo pairs combination 
FN and FB with the image triplet FNB we determined the 2.5D 
RMSE of the height for each sub area and for the three 
combinations. 
Table 1 gives an overview of the DSM accuracy evaluation 
results. The histogram of the results for the FNB DSM of Bern 
in Figure 1 shows the normal distribution of the residuals and 
that we do not have any systematic errors. The other residuals 
have the same characteristics. 
Each sub area contains more than 2.7 million points. The 
overall RMS height errors for all three test areas for the FNB 
combination are better than three pixels (5.5 m - 6.6 m). As we 
expected, we get a less good accuracy for the FN combination 
with 6.4m - 7.5m and for the BF combination even 6.6m - 
9.3m. The worst results were obtained for the sub area 
Southwest with an alpine area. 
PRISM View 
Combination 
Number of 
points 
2.5D 
RMSE-Z 
Mean / Min / Max 
Bern FNB 
4340836 
5.7 
-1.3/-60.6/50.0 
NB 
6.4 
-1.1 /-59.8/74.8 
FB 
6.6 
-1.5 /-62.5/77.2 
SW FNB 
2752822 
6.6 
0.6/-76.9/84.5 
NB 
7.5 
1.1 /-79.0/91.1 
FB 
9.3 
0.0/-103.2/220.0 
Thun FNB 
3508099 
5.5 
1.2/ -41.6 /63.4 
NB 
6.4 
2.8/-68.7/96.0 
FB 
8.1 
1.2/-109.6/ 135.2 
Table 1. DSM accuracy evaluation results of the three test areas 
for the three different view combination FNB, FN and FB. 
Bern: city and tree area, maximally height difference 400m, SW 
(south west): mountainous and tree area, maximally height 
difference 1500m, Thun: city, open and tree area, maximally 
height difference 1000m.