juired
5 deg
ht at
arrow
94].
action
ut by
rfaces
' geo-
in the
either
m re-
uishes
an be
areas
Then,
> used
r look
slope.
short-
nages.
ge ac-
quired during Cycle Il needs to be taken into account. Once
the corresponding feature is identified in the opposite-side
image, its across-track width serves to decide which of the
two candidate solutions is the correct one, thus resolving the
remaining foreshortening/layover ambiguity. The image pro-
cessing tasks involved in the procedure (i.e., the identification
of the areas of interest and the establishment of correspon-
dences between them) were done manually.
The purpose of the project described in this study is twofold.
On the one hand the goal is to automate the image pro-
cessing tasks described above, so that the whole process of
height estimation requires no, or at least very little, human
interaction. On the other hand, we intend to extend the ideas
originally developed for the special case of discretely dipping
surfaces to the more general problem of improving the ac-
curacy of stereo-derived DEMs in foreshortening and layover
areas. As already mentioned before, the different geometric
appearance of both foreshortening and layover regions in the
two partners of a stereo image pair poses special problems
when applying conventional stereo matching algorithms, nor-
mally based on gray value correlation. Therefore, a match-
ing algorithm which is able to cope with foreshortening and
layover needs to be robust to the variations in across-track
width produced by the change in ground range resolution as-
sociated with foreshortening and layover. Furthermore, in
those cases where the same feature is foreshortened in one
stereo image, but laid over in the other one, an appropriate
reversal of match points needs to be carried out, in order
to obtain correct height estimates. Possible interactions be-
tween different layover regions can lead to even more com-
plex situations: Two layover regions which appear as separate
features in one SAR image may have merged into one single
layover region in another image of the same scene, acquired
with a steeper look angle. A thorough discussion of possible
scenarios arising from fusion between different layover areas
is given by [Kropatsch, 1990]. At the current status of our
project, interactions between different layover areas are taken
into account by a simplified model which assumes that lay-
over regions which are separate in one image, but joint in the
corresponding stereo image, are just at the initial stage of
merging, where relationships between adjacent pixels are still
preserved. The validity of this assumption is currently being
tested on simulated imagery, and, if necessary, the model will
be refined.
3 IMPLEMENTATION AND RESULTS
The realization of the concept discussed above can be di-
vided into several steps, two of which are presented and il-
lustrated in the following. First, a simulation program was
implemented in order to provide a test environment for the
newly developed algorithms. The second issue we address is
the development of an image matching algorithm specially
designed for SAR foreshortening and layover areas.
3.1 Simulation
In our project, the need for simulation arises for several dif-
ferent reasons. First of all, simulation is an important tool
for verification when dealing with planetary data, due to the
lack of ground truth. Secondly, simulation provides inexpen-
sive and flexible test data for the development of new image
processing algorithms, which often cannot be obtained from
other sources. A third particular need for simulation comes
up when trying to combine SAR images taken from opposite
285
sides: Due to the strong geometric and radiometric differ-
ences caused by opposed illumination directions, the iden-
tification of corresponding features in such kind of imagery
requires knowledge about the radar imaging geometry. When
automating the identification process, this knowledge can be
incorporated by using simulation. Finally, we also employ
sequences of simulated images acquired with varying sensor
look angle in order to get a better understanding of how the
layover manifests itself in the image, and to observe the ef-
fects of interactions between different layover regions. An
example of such a simulation series can be seen from Fig.4.
Input to the simulator was a DEM of the Otztal, a rugged
terrain in the Austrian Alps, where layover occurs frequently.
(c) (d)
Figure 4: The simulated views (a) - (d) correspond to look
angles of 40 deg, 26 deg, 20 deg, and 15 deg, respectively.
They demonstrate the transition from extreme foreshorten-
ing in (a) to beginning layover in (b), and the growing and
overlapping of the layover regions in (c) and (d).
The implemented SAR simulation program is based on a co-
sine reflectance model and the assumption of a straight sensor
flight path. In addition to the simulated SAR image, a so-
called layover map is generated, which marks those parts of
the image affected by layover. Fig.5 shows again a SAR stereo
image pair generated by simulation, with the corresponding
layover maps given in Fig.6. These layover maps were em-
ployed as test data for the matching algorithm presented in
the following section.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
