The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
Dataset
No. of used points
RMSE-E (m)
RMSE-X(m)
RMSE-Y(m)
RMSE-Z (m)
Mean / Min / Max - E (m)
Mean / Min / Max - X (m)
Mean /Min /Max - Y(m)
Mean / Min / Max - Z (m)
Tx/Ty /Tz (m)
Catalonia, DTM
2870270
3.14
0.02 / -56 / 63
1.33/-3.61 /-1.16
0.62
0.00 / -40 / 28
0.75
0.01 /-35/60
2.98
0.00/-51 /55
Sakurajima, DSM
394962
9.26
1.00 /-160/ 153
1.41 /-2.63/-1.09
4.68
-0.14/-66/ 156
3.77
-0.il/-65/51
7.04
0.61 /-66/ 135
Table 5. Statistical values of the Euclidean distance (E) differences between reference DSM (in Sakurajima) and DTM (in Catalonia)
and matching DSM and values of the shift parameters (T) between the reference data and the Cartosat-1 DSM.
Figure 8. Left: Orthophoto of the area used for the evaluation of the DSM generation with reference data. The black area is the
excluded cloud area. Right: Colour coded image of the Euclidian distances (DSM - reference). The color intervals correspond to 15
m. The white circle defines a critical area shown in detail in Figure 6. The other reddish areas around the cloud area are also due to
shadows and mainly very large perspective differences.
A color coded and shaded view of the generated DSM is
presented in Figure 7. Figure 8 shows an orthophoto of the area
with the typical structure of a volcano and the large shadow
areas. One of the two calderas had to be excluded because of
occlusion by clouds.
The results of the Cartosat-1 DSM valuation are summarized in
Table 5 and visualized in Figure 8. The big blunders and the
resulting worse accuracy are partly due to the steep volcano
surface with terrain cuts, partly also having shadows and/or
causing occlusions, low texture and large perspective
differences. Additionally, there are also blunders around the
excluded cloud area (see Figure 8). The larger sigma X and Y
values in Table 5, compared to the Catalonia dataset, are due to
the mountain slopes, with inclined Euclidean distance errors
projected more in planimetry.
6. CONCLUSIONS
Regarding image quality, these Cartosat-1 images were better
than other ones used in previous tests. The major remaining
problems are the interlacing noise and the blurring of the Fore
channel.
Regarding the necessary RPC refinement and the accuracy
potential for 3D point measurement, the following can be
concluded. RPCs should be corrected by an affine
transformation, shifts alone do not suffice. The shifts with or
without affine terms were much smaller than in previous tests,
showing an improved absolute geolocation accuracy. The GCP
distribution, although for many high resolution satellite sensors
(e.g. Ikonos), is not so important, for Cartosat-1 seems to have
an influence on the accuracy, especially in the planimetry. Thus,
to be on the safe side, a good GCP distribution is recommended.
The number of the GCPs is not so crucial. For an affine
correction of the RPCs and a certain redundancy, we
recommend the use of about 6 GCPs as minimum. Despite the
suboptimal B/H ratio, the height accuracy in pixels was
exceptionally good, even exceeding previous results achieved
with sensors like IKONOS. This indicates that the errors in the
planimetric positioning are rather due to the poor identification
of the GCPs.