image
scale
nts and
1).
eft part
height
) points
d were
original
. In the
] patch
in from
2 to the
aneous
lat and
he lack
m ; the
to the
ions of
fferent
nders a
orithm
nly ca.
of the
; 84m,
meters
lid not
m grid
Visual
a grey
Figure 6. Radiometric differences between fore and aft channels.
Figure 8. The triangular meshes of the 10,000 matched points
used in the 40 m DTM grid interpolation.
level image showed that some small bumps might have remained
in the data. The roving GPS data were not available to us, so for
a quantitative evaluation of the results we did the following.
Using the DTM an orthoimage with a 13.3 m pixel size from the
fore channel was generated and the GCPs that were included in it
were manually measured. Their pixel coordinates were
transformed to ground coordinates, their height was bilinearly
interpolated from the underlying DTM, and these values were
compared to the adjusted ground coordinates of the bundle. The
RMS difference in X, Y, and Z was 6, 5.7, and 5.1 m, while the
maximum absolute error was 11.5, 10.5 and 8.1 m respectively.
Due to the small number of these points (4), the results are not
conclusive. The second approach was based on orthorectified
stereo pairs (Baltsavias, 1996). A second orthoimage from the aft
channel was generated. If the DTM were correct, the
orthoimages of the fore and aft channels should not have any
parallax. We measured the parallax with LSM at different well-
defined points over a regular grid covering the whole image
format. In addition we subtracted the two orthoimages and
matched at places where the radiometric differences were high,
possibly due to remaining parallaxes. All in all, ca. 50 points
were matched and the maximum parallax was 0.6 pixels, i.e. 8 m.
115
Figure 9. The second 40 m DTM grid displayed as a grey level
image.
A visual control of the difference image showed that most of the
differences were due to radiometric differences in the original
images and not due to parallax. Only along creeks some
parallaxes existed, partly also due to terrain smoothing because
of the large patch size.
Some of DTM errors are due to interpolation. The original
measurements were ca. 10,000 while the DTM had ca. 152,000
nodes. These errors occur especially in the areas with little or no
texture (see the triangular meshes for the DTM interpolation in
Figure 8 and compare to image texture in Figure 1, e.g. at the top
right part). To reduce this problem we made a second selection of
points with the interest operator and relaxed criteria. Thus, 80%
more points were matched and a second 40 m grid DTM was
interpolated (see Figure 9). A comparison of the two DTMs
showed an average difference of 0 m, an RMS of 4.2 m, and a
maximum absolute difference of 26 m. 96.296 of the points had a
difference of less than 10 m, and 99.9% a difference of less than
20 m. These measures are an indication of the internal precision
and repeatability of the algorithm. The maximum difference in
particular is a good indicator of the maximum error in any of the
two DTMs. The second DTM, as expected, was modelling better
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996