248
A close look at the Figure Of Merit (FOM) values,
overlapped to the image, indicates that poor matchings
are actually concentrated along breaklines showing that
the self-diagnosis tool of the system is (on visual
evidence) in good agreement with the error plot.
The same area (but with a mesh size of 5 m) was
surveyed with system B (which has as available tuning
parameters only the correlation window size and
correlation coefficient threshold). A low resolution grid
(we used a 50 and a 20 m mesh size, which amounts to
130 and 800 points respectivly) taken from the
reference data was fed to the program to support it in
the search for homologous points. The system output
around 13000 points in about Ih 30’ on an equivalent
machine. The statistics are less satisfactory than in the
proceeding case: the RMS of the discrepancies is 7 and
6 m for the 50 m and the 20 m grid respectively; 83%
and 87% of the errors are within 1 RMS in both cases.
The distribution of DEM errors is shown for the two
cases in Fig. 4 and 5. As it is apparent, compare to
System A not only errors are larger, but also they
extend further from the breaklines. As far as the
support given by the grid is concerned, the is a rather
small improvement only. This suggest that the system
may not be adequate to the task.
The procedure described in Section 2.3 for an
independent check based on two orthophotos has been
implemented in a software program and applied to the
stereo pair, in a smaller area. About 2000 points have
been matched and the corresponding ground
coordinates computed by intersection. The RMS of the
differerences with System A are in the same range as
those found from the reference values, but larger, a
comparison with the DEM from manual plotting
showed a small percentage of outliers, likely due to
insufficient robustness of our matching algorithm: no
further comparison were therefore performed with this
data set.
4. CONCLUSIONS AND PROSPECTS
The work is really still in progress, since the second
flight due next summer will provide new, more
redundant data, where multi-image techniques will be
applied. The preliminary results seems to point out that
in such a demanding environment only sophisticated
systems do provide acceptable results. It should be
stressed, nevertheless, that contour lines do not provide
an ideal reference set in breaklines areas, that is. where
the largest errors are concentrated: their nominal
accuracy along step edges cannot be maintained. Also
wooden areas, where it is known that results are
unreliable, should be perhaps taken out from the
comparison. Besides, a few buildings in the southern
area are not taken into account in the reference DEM,
so positive discrepancies there are indeed not errors. If
this is the case, these preliminary figures would
improve and. as far as System A is concerned, it may
be proved that it is suited to the taks. A definite answer
will be only available after the completion of a
topographic survey with a laser scanning theodolite.
5045700.0&-c_,
5045600.00-
5045500.00-
5O454OG.0G-K
5045300J
5045200.00- fecs
p
i'yj
5045100.00- riv
5045000.
5044S00.0Q->
JMfO,
5044800.00- p//// jI f{ \
mum vv'
gnji f -Tf'.rs fw
1602300.00 1602500.00
1602700.00
1603100.00 1603300.0
-50.00 to -6.00
-6.00 to 6.00
6.00 to 30.00
Fig. 4 - DEM error plot on system B (5 m grid. 20m support grid)
