4 STEREO MAPPING
4.1 Accuracy Assessment
The stereo mapping accuracy was derived by an
image-to-map intersection of projection lines defined
by the homologue coordinates of the GCPs in image
geometry, so-called epipolar lines. 'This intersection
results in map projection coordinates in planimetry
as well as height, which can be compared to the mea-
sured coordinates of the GCPs. The resulting resid-
uals represent the stereo model set-up accuracy and,
consequently, a-priori estimates for the stereo map-
ping accuracy.
Statistical parameters of a-priori 3D stereo mapping
residuals are summarized in Table 3 for various stereo
models combined from the available image data. The
software package RSG allows to combine individual
image pairs to stereo models for further analysis with-
out extra investment.
Concerning the KFA-1000 stereo model, the achieved
values compare very well with results published by
Konecny et al. (1988, [4]) or Sirkià and Laiho (1989,
[11]). While the achieved accuracy in planimetry is
quite good and corresponding to the high resolution
input image data, the accuracy in height is compara-
bly poor due to the small base-to-height ratio (about
0.16) caused by the camera disposition.
As further documented, the three-line scanning mode
of the MEOSS scanner basically offers a good stereo
capability. Base-to-height ratios of about 0.4 and 0.8,
respectively, are determined by this particular geo-
metric imaging arrangement. For these first investi-
gations using airborne MEOSS stereo data, however,
the achieved values are worse than might be expected
from this arrangement, caused again by the geomet-
ric problems mentioned in section 3. More detailed
studies will be made to really exploit the stereoscopic
potential of these data. A high stereo mapping ac-
curacy of a few meters in planimetry and height is
proposed by the aerial stereo model having a base-to-
height ration of about 1.0. Here, general limitation is
given by the GCP measurement accuracy.
4.2 Relief Mapping
The aerial stereo model was used to automatically de-
rive a digital elevation model for a representative sub-
area of about 2.5 by 2 kilometers. Therefore, again
the software package RSG was used, the algorithms
implemented for stereo mapping being described in
general in Raggam et al. (1991, [9]). As shown in
other experiments, RSG offers also the possibility to
TABLE 3
Statistics of stereo model set-up accuracy on ground
(meters).
Stereo images East North Height
KFA-1 / RMS 10.8 9.3 91.2
KFA-2 MIN -23.4 -194 -1201
MAX 192 18.6 106.7
MEOSS-F / RMS 10.5 7.6 17.8
MEOSS-N MIN -130 -112 -333
MAX 15.0 11.0 19.5
MEOSS-N / | RMS 5.0 8.7 25.1
MEOSS-B MIN -6.4 -10.6 -27.3
MAX 8.2 11.7 38.9
MEOSS-F / RMS 5.9 5.5 9.1
MEOSS-B MIN -9.6 127 -16.3
MAX T.9 6.9 15.3
AIR-1 / RMS 4.8 4.1 8.2
AIR-2 MIN -6.0 -4.9 -8.0
MAX 6.3 7.4 11.9
combine images from different sensors for stereo map-
ping purposes (Raggam et al., 1991 [8], 1992 [10]).
The initial processing steps have been greylevel-based
image correlation and interactive measurement of ho-
mologue image points in forest areas, where no mean-
ingful correlation output can be expected. Here, a hy-
brid correlation method combining greylevel- as well
as feature-based approaches (Paar and Polzleitner,
1991 [6]), together with a proper quality control, e.g.
using forward-backward correlation, might reduce the
interactive work.
3D coordinates were determined from homologue im-
age point measurements through the intersection of
the respective projection lines. A DEM was gener-
ated by triangulation of the received irregular point
raster and subsequent interpolation of a regular ele-
vation raster. The resulting DEM frame is shown in
Figure 9 in an oblique view from South, whereas Fig-
ure 10 shows the equivalent frame of the map-derived
DEM. Figures 11 and 12 show the corresponding con-
tour lines (contour interval 20 meters) and Figure 13
shows an overlay of the stereo-derived DEM and a
geocoded aerial image.
Using RSG, statistical parameters for the height dif-
ferences have been determined, resulting in a standard
deviation of 9.7 meters. This corresponds well to the
a-priori height accuracy of 8.2 meters given in Table 3.
Also a visual comparison of Figures 9 and 10 shows a
good correspondence of the DEMs at least in those ar-
eas, where meaningful stereoscopic measurements can
be made.
189
