164 
3 DIGITAL ELEVATION MODELING 
3.1 Automatic parallax measurement 
The measurement of differential parallaxes between 
the 53 and 29 degree incidence angle images, as 
shown in Figure 4, has been achieved through an 
iterative procedure leading to a digital parallax 
model (DPM) that was transformed subsequently into 
a final DEM. Our objective is generation of a DPM 
as a complete automated process. 
A first correlation on the stereo images 
generates an initial DPM file using iterative 
matches at multiple spatial resolutions. The 
search for homologous points is performed on many 
pairs of stereoscopic images, each pair being 
obtained by reducing successively the resolution 
of the original data. From low to high spatial 
resolutions, such an iteration procedure allows 
for a high degree of automation by limiting the 
correlator's search area. Indeed, the information 
obtained at a given level of correlation is used 
at the next level as a prediction for the location 
of the homologous point, thereby constraining the 
correlation by virtue of the successive approach 
effect. This way, the risk of correlator 
divergence, which would require operator's 
intervention, is reduced. The concept of 
iteration at multiple resolutions is related to 
approaches that take into account the image 
context because the global information obtained at 
a given iteration acts locally at the next 
iteration. At each iteration the parallax value 
grid is searched for anomalous values which are 
excluded, is interpolated to fill data gaps, and 
then is low-pass filtered. The grid step is 4 
times larger than the pixel size. 
When geometric distortions caused by relief are 
important, it is possible to optimize the 
correlation success by a process called "stereo 
reduction". In this case, the DPM obtained in the 
first step correlation process is used to reduce 
the stereoscopic effect by resampling one of the 
original images into the other image geometry. If 
the first parallax model was error free, then the 
reference image and the rectified image would be 
in registration. In practice some residual 
parallax remains and has to be estimated again by 
digital correlation. However, this search is 
restricted over a smaller area of the image. Thus 
the measure of the residual parallax between the 
reduced and reference images will refine the 
previous measurements to yield a more accurate 
DPM. 
Parallax measurement was achieved on a 50 metre 
grid. Matches at the smaller correlation window 
were done on the image intensity values of a 9 by 
9 pixel matrix. Correlation success reached 
approximately 33% of all the match points. 
In the central portion of the image (Figure 5), 
automatic correlations were not achieved. A 
visual three-dimensional analysis of the stereo 
images of Figure 4, explains partly why the 
correlation failed in this central portion of the 
image. One can notice that the image content 
exhibits substantially fewer edges and less 
texture detail than elsewhere. Also, the 
overlapping image swath is not sufficiently large 
to include all sides of this high altitude and 
very massive mountain (possibly limiting the 
multi-resolution search procedure). 
3.2 Modeling of terrain elevations 
Parallax measurements were converted to elevation 
values using line of sight intersection with 
refined shuttle positions as described earlier in 
the text. Orthographic SIR-B images were also 
generated using the same DEM (see Figure 5). 
Contour lines defined at 200 meter intervals were 
overlaid onto the orthographic 53° incidence angle 
SIR-B image (this image having less planimetrie 
errors than the 29° incidence angle one). 
Contours appear to match well with terrain 
features such as ridges and valleys. Quantitative 
evaluation for selected profiles allow for a more 
detailed analysis of the quality of the DEM. 
Figure 5. Orthographic SIR-B images (29° left, 53° 
right) and contour lines at 200 m intervals 
overlaid with 53° image. 
Figure 6 shows elevation plots for each profile 
as obtained from SIR-B data and from digitized 
topographic maps. One can notice similarities 
between these two sets of curves but also identify 
the following differences: 
- Even if the average of the RMS errors 
represents 105 m (see Table 2) for the 5 profiles, 
some portions of those profiles fit better 
with the actual terrain data; this is particularly 
true for the profile portion 5A for which the RMS 
difference is as low as 38 m (relative to an 
average difference of 17 m). 
- It is noticeable that the left part of the image 
provides better elevation data than the right one 
(illumination originates from right); limited 
overlapping swath can be envisaged as responsible 
for such effects. 
- The 9 by 9 window size of the image correlator 
together with the filtering process reduced the 
terrain high frequency component.