The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
1310
the FastStatic method. The resulting horizontal accuracy of 0.10
m and the vertical accuracy was found at 0.15 m. Each GCP
was well documented with photographs and drawings to
provide additional data for C-SAP programme members.
Additional field survey: During the field survey 3 elevation
profiles were measured using GPS RTK method. Each profile
was 3.5 km long with reference points every 20 m.
3. SOFTWARE USED
The first phase included: the stereopair orientation, generation
of a DSM and orthorectification. All image processing for TS-9
was conducted with Leica Photogrammetry Suite (LPS) 9.0
software. The results and first insights were presented in
September 2006 during the ISPRS symposium held in Goa,
India (Dabrowski et. al 2006).
In 2008 some calculations were repeated with the newer version
of the LPS software (v.9.2 Service Pack 1). Both software
versions used the RPC method. According to ERDAS, Inc. (the
manufacturer of LPS), LPS 9.2 SP1 introduced changes in the
triangulation algorithm for the RPC method as well as sensor
model enhancements for Cartosat-1. To investigate the
differences between the results from the two versions of LPS
software, some calculations done originally with LPS 9.0 were
repeated using LPS 9.2 SP1.
4. ORIENTATION
The orientation of the stereopair was established using 4 sets of
GCPs: 4, 9, 16 and 25, spread evenly across the stereopair (Fig.
2). The GCPs left out of the sets were selected as independent
Check Points. The RMSEs (Root Mean Square Errors) obtained
during triangulation for the Check Points were used as the
measure of accuracy of the orientation. Table 1 includes detailed
results.
# of GCPs
RMSEX
RMSE Y
RMSE Z
4
1,49
1,32
1,48
9
1,46
1,22
1,16
16
1,38
1,18
1,11
25
1,21
1,22
1,08
Table 1. The RMSEs calculated for independent Check Points.
Clearly, each set provides sub-pixel accuracy of scene
orientation. Similar results were obtained with the same data for
TS-9 using different software tools (Jacobson 2006, Titarov
2007). The results suggest 4 to 9 well defined GCPs for scene
orientation were enough for imagery from the Cartosat-1
satellite. In the following sections the variant with 9 GCPs is
referenced
5. SETTING UP THE STEREO MODEL
After the orientation phase, the accuracy of the reconstruction
of the stereo model was tested. The independent test was based
on manual elevation measurements made at Check Points. The
results are summarized in Table 2.
Figure 3. The GCP and Check Point distribution in each 4 sets
(GCPs = triangles).
Point ID
H GPS
H Model
dH [ml
4
163,55
163,07
0,48
8
162,63
163,56
-0,93
10
165,79
165,10
0,69
14
138,05
138,30
-0,25
16
169,65
171,23
-1,58
26
163,82
165,20
-1,38
Table 2. The differences in elevation at GCPs between the
stereo model and the field survey.
The small differences between elevation values obtained from
the satellite imagery and the ones recorded during the field
survey clearly imply the stereo model’s accuracy as well as the
eligibility of the model for DSM extraction. The following
sections analyze quality assessment of the DSM in more detail.
6. DSM EXTRACTION
The next phase of the experiment was DSM extraction using the
autocorrelation method and assessing the resulting model for
accuracy. The DSM was based on the scene set with 9 GCPs
which had been found optimal in the previous phase. The DSM
was calculated with LPS 9.2 SP1, using the exact same
configuration as in 2006 with LPS 9.0 (Search Area: 21x3,
Correlation Window 7x7, Correlation Threshold 0.85)
(Dabrowski 2006/1).
The results from LPS 9.0 (lesser than LPS 9.2 SP1) were found
incorrect to a large degree as far as elevation was concerned.
The discrepancies were in hundreds of meters and covered
relatively large area. After visually inspecting both images of
the stereopair two types of errors were found, both of which
were independent of the satellite sensor, but rather resulting
from software.
The first type of errors resulted from the presence of clouds, or
rather their shadows on the ground as captured on the stereopair.