,
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bl. Beijing 2008
1327
in any case manually. By matching the neighboured points the
geometric relations are improved before going to the next
neighboured points. DPCOR is always following the path with
the highest correlation coefficient up to the complete coverage
of the stereo pair with corresponding points having a correlation
coefficient above a chosen threshold. The threshold for
Cartosat-1 images may be the correlation value of 0.6. Usually
only a limited number of useful points is located below this
limit. Of course if the area has no variation of the grey values, a
matching is not possible (Fig. 7).
Fig. 7: frequency distribution of correlation coefficient
horizontal: frequency vertical: correlation coefficient - above
r=0.0 below r=l .0 (Mausanne-left, Warsaw-right)
Fig. 7 shows the frequency distribution of the correlation
coefficients. In the Mausanne model the object contrast was
limited because of the winter, so that the highest number of
correlation coefficients is in the group r=0.90 up to 0.95,
Warsaw in the class r=0.95 up to 1.0. In relation to other
satellites, the matching with Cartosat-1 models is extremely
successful. The overlay of the matched and accepted points to
one of the scenes in figure 8 demonstrates the successful
solution. In the Mausanne scene in some parts absolute no
contrast was on the ground. In Warsaw slight snow coverage
caused some problems (Fig.8).
Regarding the stereo pair of Castelgandolfo the matching was
really good (Fig. 9) and the not matched points are mainly due
to the lakes and to the clouds, which cover together a big part of
the images. In Fig. 9 the matched points distribution over the
after image is shown on the left, and on the right the trend of
the correlation coefficient (r) is represented using grey values 0
for the r=0 and grey values 255 for r=l.
Fig. 9: matched image points (left) and quality image (grey
values correspond to correlation coefficient) of
Castelgandolfo
Fig. 10: frequency distribution of correlation coefficient
(Castelgandolfo)
In the open areas of Mausanne, the height model was even more
precise than the accuracy estimated by means of the RMSE y-
parallax (2.87m). Of course the generated DSM showing the
height of the visible surface has to be filtered for objects not
belonging to the bare ground because the reference DEM is
related to this (Tab. 5).
For Castelgandolfo’s scene the RMSE y-parallax of 4 million
points was 1.79m, corresponding to a standard deviation of the
height of 2.86m.
• V A?
tU'C
■ Zf. j* '
gjpf ’
>
... •*
'■ O '
f ■
,\V
' V ~>'v. " . a".
V
* * * '
> * /•
—
ßi*
K
ß
f- -
Fig. 8: overlay of matched points (white) to after scenes
(Mausanne-left, Warsaw-right)
For the Castelgandolfo stereo pair, based on the scene
orientation obtained with the RPC generated from SISAR, and
after the automatic image matching, the digital surface model
has been generated using the Hannover software RPCDEM, and
compared with a precise reference DSM extracted by aerial
photos.
On the contrary Mausanne and Warsaw DSM have been
extracted using the RPC supplied with the image.
The reference aerial DSM covers an area of about 85km 2 in the
centre of the scene, including forest parts and both open and
urban areas. These result are satisfying considering that the area
of interest is full of elements that do not belong to the bare
ground (like trees or buildings).
Image
Area
SZ
bias
sz =
f(a=inclination)
Mausanne
no filter
4.02*
-0.51
3.91 + 1.64*tan a
yes filter
3.30*
0.48
3.17 + 3.14*tan a
Warsaw
no filter
3.23*
-0.54
3.16 + 1.19*tan a
yes filter
2.43*
0.44
2.39 + 8.80*tan a
Castel-
gandolfo
no filter
2.88*
-0.06
2.71+0.41 *tan a
yes filter
2.29*
0.30
2.26+0.17*tan a
no filter
4.67**
-0.58
3.95+1.64*tan a
yes filter
4.06**
-0.34
3.27+1.91 *tan a
Tab. 5: accuracy of Cartosat-1 height models checked by
precise reference DEMs[m] preferred to open area,
* deferred to urban area) (Jacobsen 2006)
