For the first test we simulate the imaging process. The DTM,
the orientation of two images and an orthophoto are given
and the two images are generated by resampling from the
orthophoto. For each of the image pixels the corresponding
point location in the orthophoto is found iteratively by a nu-
merical difference scheme. Consequently the only differences
between this two images are of geometric nature. Now an
error is introduced into the DTM and the stereo orthophotos
are derived using the imagery found by the simulated imag-
ing process. The differences between the stereo orthophotos
are directly (uniquely) the consequence of the error in the
DTM. The results applying the linear regression model (A),
the FE-procedure (B) and the DOP-procedure (C) are shown
in figures 2, 3 and 4, respectively.
That the error is very well indicated by procedure A (figure
2) could be expected, because the idealized assumptions of
the simulated imaging process fit well to the simple model.
Because of this the form of the introduced DTM error which
approximates a cross can be seen quite nice in figure 2. In
exactly five neighboring points of the DTM an error of 3
m in elevation is added. More surprising is that with the
FE-procedure just in this case difficulties arise. The residual
images €, and € show noise as expected. In the difference
surface (figure 3), for which in the case of this figure a grid
spacing of 5 times the orthophoto pixel size is chosen, the
error can be recognized but much less distinct than expected.
Coarsening the resolution of this difference surface amplifies
the visual impression of the error but up to a spacing distance
of 10 times the pixel size. For larger spacing distances the
response in the difference image is lost. Neglecting the fine
grained lowest level of the DOP, the next three bandpass
levels of the DOP (figure 4) all indicate the error. So with
procedure A and C the approximate position of the error is
simply located and roughly size and form of the erroneous
region could be found.
The second test then uses the two images without modifica-
tion, i. e. just as they have been digitized from the aerial
photos (figure 1). From these two images the correspond-
ing orthophotos are derived by differential resampling. The
same DTM with the identical error as in the first test with
the simulated idealized imaging process was used. The stereo
orthophotos are plotted in figure 5. If viewing this image pair
stereoscopically the DTM error as well some trees which are
not represented in the DTM can be observed. The results
applying the linear regression model, the FE-procedure and
the DOP-procedure again are shown in figures 6, 7/8 and 9,
respectively.
The residual image which results from linear regression gives
some indications about the D'TM errors as can be seen from
figure 6. Plotted in this figure is the thresholded residual
image, i. e. only those residuals which are larger than two
times the standard deviation. The black blobs in the lower
left and right of this image correspond to trees. Figure 7
gives an impression of a residual image estimated by the
FE-approach. Typical for this FE-residuals is that, as can be
seen in figure 7, only noise is present. The difference surface
(figure 8) shows large systematic differences between the two
orthophotos. Hardly to recognize are some indications to
trees, and practically nothing is to see from the D'TM error
in this data. The last images (figure 9) show the first four
bandpass levels of the DOP. On the coarse levels 2 and 3 the
trees can be recognized. The DTM error shows up slightly
on level 2.
236
in total, this real example reveals considerable problems
which arise when the radiometric distortions between the im-
ages are large. If the radiometric transformation is very flex-
ible as in the case of the FE-model, the resulting error terms
indicate that the transformation works satisfactory. Unfor-
tunately the situation for localizing DTM errors in this case
is just inverse, i. e. the errors are also taken away with the
transformation. More promising are the more simple proce-
dures based on regression as well as on image pyramids.
4 CONCLUSION
In this paper we followed a pragmatic line of investigation:
at first we have explored some experiments to gain insight
into problems we meet on the way to automatic DTM ver-
ification. To eliminate differences between stereo orthopho-
tos arising from distinct physical processes we used a simple
model based on linear regression and a more complex pro-
cedure based on modelling differences by a finite element
surface. The idea behind the third procedure is multiscale
(bandpass) analysis, used to separate events arising at dif-
ferent scales.
The example chosen for this experiments are real images with
considerable radiometric differences. In summary, with this
image data, all three procedures have problems in detect-
ing inconsistencies, which result from differences between
real world geometry and the DTM. The more complex FE-
approach seems to eliminate too much, so that nearly no
hints to the DTM errors remain. The more simple and
computational more efficient approaches based on regres-
sion analysis as well as on bandpass DOPs seem to be
more promising. This especially holds true for imagery in
which systematic differences between the image intensities
are small. The results found based on the simulated imaging
process support this observation.
One direction of future work is to explore the information in-
herent in the symbolic level of image representation. Though
our first experiments with edge images have not been very
encouraging, we expect that the less sensitivity to physical
effects will lead to an improvement for the verification task.
REFERENCES
Ackermann, F., Hahn, M. (1991): Image Pyramids for Dig-
ital Photogrammetry. In: Ebner, H. et al.: Digital Pho-
togrammetric Systems, Wichmann-Verlag, Karlsruhe.
Ackermann, F., Krzystek, P. (1991): MATCH-T: Auto-
matic Mensuration of Digital Elevation Models. Pre-
sented paper to the 3rd Technical Seminar of Sociedad
Espanola de Cartografia, Fotogrametria y teledeccion,
Barcelona.
Blachut, T.J. (1971): Stereo-Orthophoto Systems.
Bildmessung und Luftbildwesen 39, S. 25-28
Coleman, E.N., Jain, R. (1982): Obtaining 3-dimensional
Shape of Textured and Specular Surface using four-
source Photometry. Computer Graphics and Image Pro-
cessing, 18(4).
Collins, S.H. (1968): Stereoscopic Orthophoto Maps. Cana-
dian Surveyor 22, pp. 167-176.
Finster
Bild
Hahn, ?
ture
DE!
Ikeuchi,
ram
Con
Jansa, .
pass
talei
Fern
Klinker
sure
Jour
Nayar,
Sha]
brid
Mell