The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
1352
sub-pixel accuracy in all directions. SAT-PP performs better
than ERDAS (differences up to 1/5 of the ground pixel size),
while the rigorous model of PCI needs more GCPs in order to
perform well. The residual distributions of Figure 4 after
orientation with PCI rigorous model show some inconsistencies.
There is a clear systematic deformation visible in the
planimetric residuals of the CPs, which is however not present
in the GCPs. Therefore the results of the rigorous model should
be analysed more in detail in terms of statistical behaviour of
each unknown parameter. In particular self-calibration and
external orientation modelling should be studied more in detail.
Unfortunately this is not possible with the PCI software. The
rigorous models developed at ETH Zurich (Poli, 2005,
Kocaman and Gruen, 2007) and Sapienza Université di Roma
(Crespi et ah, 2007) will be extended in order to process
Cartosat-1 scenes and complete the investigations in this respect.
In addition more tests are required for the analysis of GCPs
distribution, which seems to have an influence on the accuracy
of Cartosat-1 stereoscenes, especially in planimetry (Kocaman
et ah, 2008). During the investigations the accuracy assessment
will be performed using the Leave-One-Out method too, as
proposed in Brovelli et ah, (2008). In fact the leave-one-out
method is a well known statistical estimation technique
currently applied in different fields such as machine learning,
bioinformatics and generally in any other field requiring an
evaluation of the performance of a learning algorithm (e.g. in
geostatistics) and can give a good contribution for accuracy
assessment, avoiding to draw conclusions which are constrained
to a particular set of CPs.
5 DSM GENERATION
The image matching was carried out independently with the
software packages PCI-Orthoengine and SAT-PP, since they
contain different matching strategies. PCI-Orthoengine uses
crosscorrelation, whereas SAT-PP uses a multi-image matching
approach especially designed for Linear Array sensors, based
on the extraction of three kinds of features (feature points, grid
points and edges), as explained in (Baltsavias et al, 2006, Gruen
and Wolff, 2007)
Clouds are not part of the requested surface model. Therefore
they were masked out by manual measurements in the
images.The same was done for water surfaces, because they are
also disturbing elements for the image matcher.
Taking into account that the different radiometry between the
two images, due to the variations in the sensor view angles,
may influence the matching results, the Cartosat-1 images were
filtered to enhance the edges and then to reduce the radiometric
errors. The images used for DSM generation in PCI have been
radiometrically processed with a) a smoothing moving average
filter to reduce the high frequency noise, and then b) a Sobel
filter, that transformed the images into gradient maps. In case of
SAT-PP, the matching algorithm applies a Wallis filter
internally. After matching the images, from the 2D coordinates
of the homologous points the corresponding 3D point clouds
were generated using the corresponding orientation parameters
previously estimated, and interpolated to generate a 5m grid
surface model. In SAT-PP masks were manually measured in
order to exclude clouds and water surfaces (i.e. lakes) from the
DSM computation.
Figure 5. Zoom into the SAT-PP DSM, visualized in shading
mode. The white area is part of the mask.
6 ACCURACY ANALYSIS
At this stage, the corresponding DSMs were interpolated and
compared to the reference DSM over an area of only 2km x
2km with urban and open zones and free from clouds (Figure
6).
Figure 6. Area used for DSM comparison (BAND A).
The 3D differences were computed through a co-registration
with the LS3D (Least Squares 3D Surface Matching) software,
developed at ETH Zurich, which includes an algorithm for least
squares matching of overlapping 3D surfaces (Akca and Gruen,