International Archives of the Phot«
erammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
5.2 Results
4
re the method was able
first examples are taken
We start this section with examples whe
to generate a DTM of good quality. Th
from the Eggenburg data set. Fi; | shows a shading of a
DSM acquired by image matching with MATCH-T before
filtering with SCOP++ in an area in the east of Eggenburg.
Figure 5 shows the DTM that could be derived by hierarchical
robust filtering.
Figure 5. Shaded view of a DTM after including break lines
and filtering with SCOP++ (Eggenburg east). The
DSM is shown in figure 1.
The influence of the off-terrain points on houses could be
eliminated completely. Even large buildings such as a factory in
the left lower part and blocks of houses in the right upper part
of the test area could be eliminated successfully. This was
mainly achieved by choosing a rather coarse resolution of 15 m
for thinning out the data in the first iteration, at the cost of a
degree of smoothing that cut off some terrain features. If these
smoothing effects are too large to be tolerated for the
application of the DTM, break lines determined by interactive
measurement can be considered in the filtering process. Thus,
these smoothing effects can be avoided.
It was interesting to observe that using a high degree of
smoothing in image matching smoothed the DSM at houses and
trees without completely eliminating them. As a result, some
off-terrain points were classified as terrain points by robust
filtering because houses were not accentuated enough to be
distinguished from the terrain. We think that it would be easier
to select the appropriate parameters in each step of our filter
strategy if the smoothing parameter were set to ‘low’ in image
matching. Actually, it would be desirable not to filter or smooth
the original data at all, to achieve a point distribution closer to
the one delivered by ALS.
Figure 6 shows the result for another area in Eggenburg. The
terrain is more undulating than in the example in figure 5, with
some abrupt changes along a railway line and dense vegetation
in the left lower part of the scene. Off-terrain points on houses
could be eliminated again, but hierarchical robust filtering could
not remove the off-terrain points in the dense forest south of the
railway line. As predicted in section 4.1, this was caused by the
lack of terrain points delivered by image matching. This
problem could only be circumvented by manual measurement of
3D points on the terrain in the forest, which is, however, hardly
feasible.
We selected the town centre of Eggenburg to check the
performance of our method in a densely built-up area (figure 7).
In the densely built-up area in the left lower part of the scene
+ 418
only a few terrain points were delivered by MATCH-T. The
grid width for the first thinning of the data had to be chosen
rather wide (30 m) in order to cope with large areas without any
terrain points. Still, it was not possible to get as good a DTM as
in the more rural areas of the previous examples. The reason for
this is the lack of terrain points in inner courtyards and narrow
streets between the houses. However, the influence of the off-
terrain points on the houses could be eliminated and a DTM of
quite a good quality could be achieved. Again, unwanted
smoothing effects could be reduced by introducing break lines.
m.
Figure 6. Shaded view of a DSM from image matching (left)
and the resulting DTM after including break lines
and filtering (right) (Eggenburg west).
Figure 7. Shaded views of an elevation grid acquired by
MATCH-T (left) and of a DTM after filtering (right)
for the city centre of Eggenburg; contour lines in the
DTM are shown.
Finally, we want to present some examples where our method
failed to eliminate the influence of the off-terrain points. In the
example taken from the waste disposal site near Stockerau, the
influence of off-terrain points on buildings and vehicles should
have been eliminated. The terrain contained small features such
as heaps of sand and waste (figure 8). As the shapes and
dimensions of the buildings are nearly the same as those of the
terrain, no satisfying result could be achieved. The filtering
method either eliminated both the buildings and heaps of sand
and waste material, or it eliminated neither of them. It is,
however, no surprise that the method only works if there is à
distinction in appearance between the terrain and the objects
that should be eliminated.
Internatio
UH iiid
Figure 8.
Figure 9.
The Schn
filtering
wooded a
of the ele
10 m. Co
way than
very few
grid widt
The para:
restrictive
possible
coarsely :
represent:
iteration,
above th
below th
original |
using the
Monnet
parametet
(h =
above the
off-terrair
ing
few terra
alanrith
algorithm
—
e
c
Image m:
was obvic
