The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part Bl. Beijing 2008 
1299 
3.2 Mausanne-les-Alpilles 
Two stereo scenes are available for this area with an overlap of 
nearly half a scene. Thus, there was a chance for separate and 
common evaluations. GCP from a GPS campaign were 
provided by the principal investigator from the Joint Research 
Centre of the European Commission (JRC, Institute for the 
Protection and Security of the Citizen, Agriculture and 
Fisheries Unit) for stereo pair Ml. Via several matching steps 9 
of these GCP could be transferred to stereo pair M2 (in the 
overlap region). The RPC correction summary is given in table 
3, the summary of the separate DSM processing in table 5. 
Figure 3: DSM in overlap area of stereo pairs Ml/2 (grid 
spacing 10 m, size of area 17.9 km x 23.5 km) 
The orthoimage check matching was made for the stereo pair 
Ml. The about 23400 shift vectors have a mean of (0.11, 0.21) 
and a standard deviation of (0.18, 0.32) for rows and columns, 
respectively. 
Special processing of the overlap area of M1/M2: 
In order to derive 4-ray tie points the matching for the overlap 
area was done through the following steps: 
1. hierarchical matching of overlap regions of the aft 
images (MIA and M2A) 
2. region growing for densification 
3. transfer of matched points to the corresponding fore 
images (MIA to M1F and M2 A to M2F) by a special 
transfer point matching feature of the region growing 
software 
The numbers of points accepted in multi-ray forward 
intersection is given in table 6 together with the number 
achieved when merging the object space points found in the 
separate processing of the stereo pairs Ml and M2 (last row of 
table 6). The latter number is much larger than the number of 4- 
ray points. This can be attributed to difficulties of multi 
temporal and multi-direction matching. Also the cloud region 
available in the overlap area is of course fully missing in the 4- 
ray tie point set which is a further disadvantage. The standard 
deviations of the height differences are comparable. 
Matching 
partners 
Ray 
s 
Nr. of 
accepted 
tie points 
(million) 
Height difference: 
reference DSM minus 
Cartosat-DSM (m) 
mean | o 
from forward intersection 
M1A/M2A 
2 
4.60 
-2.1 
5.51 
M1A/M2A 
/M1F 
3 
2.13 
-1.4 
3.64 
M1A/M2A 
/M1F/M2F 
4 
1.22 
-1.3 
3.36 
DSM/DSM shift estimation 
M1A/F 
M2A/F 
2 
2.16 
2.23 
-0.77 
3.2 
Table 6: Matching experiment in the overlap region of the 2 
Mausanne stereo pairs 
Thus, the pure merging of the separate matching and forward 
intersection results is more promising for practical purposes. 
The lateral shift vector of this 4.39 million point set versus the 
reference DSM is (3.2 m, -3.1 m). Figure 3 shows the DSM 
derived by merging of the point sets for the overlap area. 
3.3 Bavaria 
A laser DTM of the Bavarian Survey for the Taching area is 
available (5 km x 5 km, 5 m grid spacing). For GCP extraction 
an orthoimage of 2 m GSD could be retrieved from the Geoweb 
server of the Bavarian Survey (figure 5). 14 GCP are extracted 
from orthoimage and Bav-A and complemented to stereo tie 
points via LSM (as described for Catalonia). Evaluations 
reported here are restricted to this small Taching subset of the 
Bavarian stereo pair (see table 3 for RPC correction results). 
The results in table 5 show a large standard deviation of the 
height differences to the reference DTM. In order to prove the 
influence of the forest parts the laser DTM was subtracted from 
the (smoothed) Cartosat-DSM. In the difference image all 
pixels with values larger than 10 m are marked black. The result 
is shown in figure 4. A good correspondence of the black areas 
with the forest areas in figure 5 (orthoimage of Bavarian Survey) 
exists. This explains the high standard deviations in table 5. 
4. CONCLUSIONS 
Comparisons of the generated DSM with the JRC and ICC 
reference DSM/DTM result in standard deviations of the height 
differences of 3-4 meters after affine correction of RPC (for full 
scenes / case Bavaria excluded). This should be ranked as a 
very good result because no object classification in terms of 
DSM/DTM differences is applied. Thus, CARTOSAT-1 stereo 
imagery is well suited for the derivation of DSM and 
orthoimages with 1-2 pixel vertical accuracy (1 a) in terrain 
with good pattern matching characteristics and moderate slope 
angles. Prerequisite to achieve this accuracy is a set of well 
distributed GCP of high accuracy for RPC correction (about 
half pixel standard deviation at GCP). In mountainous terrain