3. EVALUATION OF DTM ACCURACY
First, a regular DTM grid with a 2.5 m spacing was interpolated
from the VirtuoZo matching results using an own programme.
VirtuoZo is unable to interpolate a DTM grid in the usual N-S/E-
W direction, although according to the documentation it should
be able to do so. Then, the manually measured points were
interpolated in the DTM grids in order to evaluate the accuracy of
the latter (see Tables 1 and 2; for the version of DPW in the whole
image and points with quality code > 32, i.e. reliable points, see
section 4.2). These results were influenced by several blunders of
up to ca. 100 m in different difficult regions (usually regions of
little or no texture, and abrupt terrain discontinuities, see small
rectangles in Figure 3). Thus, a second accuracy analysis was
performed in a region where no big blunders in the matching
results occurred (see big rectangle in Figure 3). The number of
manually measured points in the region was ca. 4,000. The DTM
accuracy in this region was much higher (see Table 1) and is an
indication of what accuracy can be achieved by automatic DTM
generation algorithms, if the results have no big blunders. The
errors in the second smaller region are shown as grey level image
in Figure 4. In all cases no editing of the image matching results
was performed. In a practical situation all very dark regions could
be excluded a priori and thus many blunders could be avoided
and processing time could be reduced. The accuracy of both
systems is very similar, as Tables 1 arfd 2 indicate. Compared to
VirtuoZo, DPW 770 has slightly better RMS and maximum error
in the whole image, but its error distribution is slightly worse. In
the region without big blunders the accuracy of both systems is
very similar and quite high (ca. 99% of the points have an error of
less than 3 m = 3 RMS). Contours from the VirtuoZo results are
a bit more noisy than the ones from DPW, but this is probably due
to the smaller patch size. The small average in all cases shows
that no large systematic errors occurred. The main problem is the
existence of blunders. There are not many, but they are big. The
DTM accuracy that was requested by the glaciologists (0.5 m)
was not achieved. However, for the low base/height ratio of the
stereo model and the poor control point configuration, the
pointing accuracy in height that could be achieved with good
manual measurements was one meter. This accuracy was
achieved by both systems but only after exclusion of the big
blunders. Thus, it can be stated that glaciers (or other type of
terrain with similar difficulties) can be measured by automated
DTM procedures, but efficient, fast and comfortable methods for
the automatic detection and exclusion of blunders are still
missing.
The performance of these digital systems was checked in a
second test, which will be presented in another paper. However,
we think it is useful to briefly present these results here for
comparison. The second test involved a stereo pair with image
scale 1:68,000 and an overlap of ca. 50%. The terrain was rolling
(height differences of ca. 400 m) but quite tricky and difficult
including many forested and small urban areas. The digital
images had a pixel size of 25 microns. The orientation of the
images had an accuracy of ca. 1 m ahd was used in both digital
systems. The reference DTM was the DHM25 of the Swiss
Federal Office of Topography with an accuracy (RMS) of 1.8 m.
With the DPW 770 the new IOR (Iterative Orthophoto
Refinement) strategy was used. The green layer of the scanned
1:25,000 topographic map was used to separate between forested
and nonforested areas. Thus, three accuracy tests were performed
(in forested, nonforested and whole area). The accuracy in
forested areas can not be checked because thè reference DTM
does not include the height of the trees. In the nonforested areas
the RMS error for both systems (ca. 130,000 points) was 2.6 m, a
result that is very good, if the small image scale is considered,
and very close to the accuracy of the reference DTM. In both
systems 99% of the points had an error of less than 8 m (3 RMS),
and the error distribution was very similar.
4. COMMENTS ON AUTOMATED DSM GENERATION
4.1. DTM Strategies
Both systems permit the selection of a matching strategy.
VirtuoZo just offers the possibility to choose one of five
strategies, mainly depending on the terrain slope and form,
without permitting the setting of any other parameters except the
patch size used in matching and the grid’ spacing of the image
grid where matching is performed. We used the strategies 5
(rugged terrain) and 3 (smooth hillocks) and the results were
completely identical! DPW offers three basic strategies (flat,
rolling and steep terrain) and four other strategies for tie point
measurement and mosaicking. When the terrain has mixed form
and slopes, the steep strategy should be preferred. Each of the
basic strategies comes in four versions depending on accuracy
versus speed, elimination of trees and buildings, and DTM thin-
out in flat areas when the DTM is dense. In the new software
release there are also seven additional strategies (still in
experimental stage) employing the IOR matching method. The
user can modify any strategy file or create its own file. Each
strategy file consists of several processing passes, each pass
working on images of a certain pyramid level. The number of
passes/pyramid levels is 6/4, 8/5, 8/6 for the flat, rolling and steep
strategies respectively. Each pass has 38 or 44 parameters for thé
old and new strategy file versions respectively. The strategy files
are hard to read and their documentation poor. Some of the
parameters are not used but the user does not know which. Others
have wrong values, but again it is unknown whether they are
used. There are some parameters that are used and are not
optimally set, e.g. the patch size is 15 x 15 pixels for all strategies
and processing passes; larger slopes and less smoothing is
permitted in the upper pyramid levels etc. The parameters for
spike deletion do not seem to work well, while those for deletión
of obstructions (houses and trees) are too coarse, do not take into
account the obstruction height and can lead to removal of tips of
hills etc. Other parameters are fixed, although they should be
adaptively derived from the data itself. For example in the steep
strategies the pyramid levels are 6. This value may not be
sufficient for large parallaxes. The optimal value could be given
by the user or estimated from the data by a coarse matching.
Concluding, the strategy files and the algorithms behind them,
although they need definite improvement, are generally positive.
However, a user with little knowledge will use these strategies as
a black box and might run into problems without being able to
understand why. The sophisticated user needs more information
on the effect of the parameters, the internal computations and the
use of the thresholds. In this respect the supplied documentation
is insufficient.
A new Automatic Terrain Extraction (ATE), provisionally called
Adaptive ATE (AATE), is under development for DPW. Using a
rule file and an inference engine it will adaptively tune matching
parameters depending on the terrain and image content, and
106
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
