4,2 Correlation results
test site points avg. oorr. low corr. failures
0.6 «r «0.7
Landshut 31200 0.84 7% 2%
Deggendorf 900 0.87 2% 0%
Vilsbiburg1 625 0.86 6% 0%
Vilsbiburg2 625 0.89 51 0%
Table 2 : Correlation results
The results of the correlation runs are listed in tab. 2, whereas
fig. 6-12 show the perspective presentation of the heights.Fig. 5 gives
a copy of the image content for test site Landshut.
Examinating the results in tab.2 it is obvious, that the image content
provided enough information for the correlation. The high level of the
average correlation coefficient as the small number of failures proves
the success of the calculations. The existence of failures for test site
Landshut might be caused by the errors in the orientation parameters,
which led to y-parallaxes in some parts of the covered area.
However, the value of the correlation coefficient doesn't say anything
about the quality of the resulting heights. Likewise the perspective
presentation in fig.6 shows some mostly isolated peaks, which doesn't
match the adjacent height level. These wrong calculated heights belong
to the problem of gross errors, which sometimes occur in the correlation
process. Due to the manifold reasons responsible for this effect the
phenomen can't be easily suppressed. On the other hand the quality of
the calculated heights is reduced at all.
To improve the results and to eliminate obviously wrong heights the
calculated data will be filtered. Using a non-linear, noise reducing
filter like a median working in a 3*3 window, the updated heights look
as fig. 7 shows. The improvement is impressing and the plot now clearly
repoduces the topography of the surface without disturbing influences.
To establish the amount of correction the calculated as the filtered
heights were compared with the heights of the reference DTM. Table 3
gives the results.
compared heights
test site correlated - refer. DTM filtered - ref. DIM
Deggendorf 40 31
Vilsbiburg1 34 19
Vilsbiburg2 25 15
Table 3 : Mean square height residuals in m
The improvement ranges from 21% to 46% and proves the validity of the
filter process. One important reason for the effectiveness of the median
filter in this context is, that the gross errors mostly occur isolated
what gives them the characteristics of noise.
As visual examples the correlated, reference and filtered heights are
given for test site Deggendorf in fig. 8 - 10. It is satisfying to
realise the correspondanoe in the relief structure for the filtered and
reference heights.
- 85 -