The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
1366
5. DSM GENERATION AND ACCURACY ANALYSIS
The image matching process was realized by using the program
package SAT-PP developed at IGP, which has been used
extensively for various platforms, sensors and image scales with
very good results, published for several projects. More details
about the underlying algorithm can be found in Zhang (2005).
SAT-PP produced very densely matches. It is using multiple
match-point primitives (regular image grid points, interest
points and edgels). For both test areas, we chose a modified
cross-correlation matching and no LSM at the final stage.
A quantitative 3D evaluation was realized by co-registering the
generated DSM and the reference DSM/DTM by using the
method described in Gruen and Akca (2005). With the
respective program LS3D, three shift parameters (Tx, Ty and
Tz) were determined between the two dataset to remove
possible offsets. After this co-registration, the Euclidian
distances (E) between the two DSMs are computed point-wise,
and the error is also split into its X, Y and Z components.
5.1 Testfield Catalonia
For the whole area, a DTM with a grid size of 15 m was
provided. For the generated DSM, we chose a grid size of 10 m,
corresponding to 4 times the GSD of Cartosat-1. About half of
the testfield is a hilly region with forest, trees and bushes and
open area. The rest is a flat area with a part of the city of
Barcelona, some villages, trees and open areas. A lake and the
sea were excluded from the evaluation (see Figure 4).
Figure 4. Generated DSM of the testfield Catalonia (grid
spacing 10m). The black area is an excluded lake area. The size
of the evaluation area is 29 km x 25 km.
The results of the 3D evaluation are summarized in Table 5 and
visualized in Figure 5. The residuals smaller than -3 sigma and
larger than +3 sigma are very few. The larger errors are mainly
caused by shadow areas and differences between the reference
DTM and the matching DSM, like trees (see reddish areas in
Figure 5) and multitemporal differences, like surface mining
(see white circles in Figure 5), while the bluish arc at the
bottom left is due to a highway.
Figure 6. Blunder area resulting from different viewing angles
causing large perspective differences at terrain with large slope
and unfavorable aspect (see also Figure 8).
Figure 7. Generated DSM of the volcano Sakurajima, grid
spacing 5 m. The central red/blue part corresponds
approximately to the DSM evaluation area (4km x 4km). The
black area is excluded because of clouds in the images
Figure 5. Color coded image of the Euclidian distances (DSM -
reference). The color intervals correspond to 5m. The two white
circles show the blunders caused by surface mining.
5.2 Testfield Sakurajima
The main characteristics of the steep volcanic area results in the
following matching conditions: a large part of the area of
interest was covered by shadows, clouds obscure a part of the
image, the texture is generally weak, while the perspective
differences at steep slopes were very large (see Figure 6).