Tertre
und with
lying the
es. Con.
zical and
Shadow
lermined
ct class.
pical for
1dicators
on phase
: Fischer
models,
overlap)
ir neigh-
single or
lers by a
mps like
basis of
ommend
sociated
pectrally
x (199)
e treated
h=0});
> further
nding on
4
Jochen Schiewe
To demonstrate a simple but typical example for the need of
the proposed combined semantical and geometrical ap-
proach, figure 8 shows a height profile of an inland water
within SPOT imagery (Schiewe and Isaac, 1999). The
matching algorithm produces an undesired smooth transi-
tion from the surrounding to the water area, so that no con-
stant height can be observed within the water body (in our
examples height deviations come up to 40 m).
image
inhomogenous
regions
homogeneous
regions
The identification of homogeneous areas can be done auto-
shadows waters others matically using a variance filter. A fast elimination of this
NS dc error would be to set all heights within the homogeneous
region to the lowest value in this area. In order to improve
Ah=const. | Ah # const. | the quality especially around the water boundary image pre-
v processing and topological post-processing steps can be
Ahz 0 Ah z 0 applied (Schiewe, 1997).
A similar height prediction model can be set up for other
critical image regions like repetitive structures.
h = undefined
Figure 7. Height prediction for
homogenous image regions.
200 I I
original DTM |
100+ I corrected DTM y ?
~~ |
04
~~]
1004 m
homogeneous
su region
-300 t ç ,
0 200 400 600 800
Figure 8. Detection of erroneous water regions within an elevation model (distance and heights in meters).
SUMMARY AND CONCLUSIONS
In order to achieve satisfying results by automatically processing remotely sensed imagery - for instance for the genera-
tion of reliable Digital Surface Models (DSMs) or the successful extraction of topographical objects - mutual benefits
are expected from methods combining elevation and image data. We have presented and empirically proved modified
and new procedures for different, but closely linked topics using digital airborne imagery:
For the automatical determination of an approximated Digital Terrain Model (DTM) from a given DSM (resp. for
computing a normalized DSM), we propose a new approach called compressing opening that in contrast to existing
methods adjusts the size of local operator windows to the that of topographical objects which shall be eliminated.
For detecting buildings and wooded areas in images we are using a combination of various elevation parameters
(like normalized DSM altitude or density of slope gradients) and spectral indicators (like NDVI or spectral texture)
by introducing probabilities for the membership of a pixel to a certain object class with respect to every indicator
which are combined with a maximum a-posteriori estimation rather than relying only on single parameters or sim-
ple thresholding operations.
For the detection of blunders within elevation models we recommend the integration of image or image-derived
information in order to check the height behaviour accordingly to the corresponding structures or object classes and
their neighbourhood (as a very simple example: water areas should show a constant height).
So far very promising results have been achieved by applying only rather simple image processing operations. Future
1
Mplementations will also consider topological as well as advanced geometrical and semantical methodologies. Fur-
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 813
