) after
point
space
ins of
of 3rd
object
ation,
ng all
ed by
mated
by the strict model. The PMFs are much faster and almost equally
accurate as rigorous transformations using the strict model
(difference « 0.1 pixel). We have used the PMFs with SPOT
images in matching for DTM generation, and orthoimage
generation with great success (Baltsavias and Stallmann, 1993).
More details on the characteristics of the PMFs and how they can
be employed for the above mentioned tasks can be found in
Baltsavias and Stallmann, 1992.
3.1. Point Measurement
The measurement of the GCPs proved to be a very difficult task.
Although their ground coordinates were accurate to 10 cm, their
actual accuracy is rather in the 1 - 5 m range due to problems in
their identification. For the image measurement we used existing
image coordinates, sketches on printed image chips, and all three
preprocessed images. The image coordinates were refined by the
following procedure. After some runs of Kratky's bundle all
blunders, like measurement of wrong fencelines etc., were
identified and corrected or removed, if their identification was
not possible. The use of the Wallis filter and the nadir image
helped a lot in this procedure. Then, 30 points with good image
and ground measurements and distributed over the whole image
format were selected and used as control points. Their image
coordinates were further refined by runs of the bundle and
analysis of the image residuals. Since the confidence in the
ground coordinates was much higher than the one in the image
coordinates, it is expected that when ground and image
coordinates do not fit, this is due to image measurement errors.
After having a strong and accurate geometry, the rest of the GCPs
were sequentially introduced as control and their image
coordinates were corrected by the same procedure. Even if a
point had an error (note that gross error were removed
beforehand), it could not have a large effect on the sensor
orientation due to the strong and accurate network. This
procedure was performed for the fore and aft channels. In
addition, the manual measurements in the fore image were used
as reference and transferred to the aft image by Least Squares
Matching (LSM). In the nadir channel the measurements were
much more difficult due to the higher noise, which was further
enhanced by the Wallis filter. Thus, the results of the bundle were
suboptimal. We plan to use the images, as preprocessed for the
DTM generation, to repeat this task. The whole procedure took
two whole days. The image coordinates were distributed to and
used by other colleagues with good results (Fraser and Shao,
1996).
32. Evaluation of the Point Positioning Accuracy
To evaluate the point positioning accuracy we made several runs
of the bundle for the fore/aft images using 3 control point
distributions (6, 10 and 20 points), linear and quadratic attitude
rates, and image measurements in the aft image done manually
and by LSM. The selection of the control points was based on
their image quality and distribution over the whole image. The
manual and matching measurements led to similar results.
Previous tests with SPOT images have always shown that
matching measurements lead to better accuracy in height,
because the points in left and right images correspond better. This
was not the case here, because (a) the manually measured image
coordinates were refined by the use of bundle and the
correspondence between left and right images was already very
good, and (b) due to poor point definition and high noise the
113
accuracy of matching was decreased. The results for the
matching image measurements are shown in Table 1.
Table 1. Geometric positioning accuracy. RMS errors of check
points (in m).
Model | GCPs | Check | 09 X Y Z
points
IL 20 45 6.0 0.9 7.4 9.4
L 10 55 7.7 9.0 8.7 10.1
L 6 39 10.1 10.9 8.1 12.2
Q 20 45 3.6 6.2 6.4 6.7
Q 10 55 2.0 6.7 59 7.4
Q 6 59 2.3 7.4 10.7 7.9
The quadratic version is clearly better than the linear one. With
SPOT the difference between the two versions was small. An
explanation for the clearly better performance of the quadratic
version with MOMS-02/D2 can be either a less stable orbit of the
space shuttle, or the larger image dimensions in flight direction
(110 versus 60 km for SPOT). We also tried additional control
point selections, keeping their number as above. While the 10
and 20 point versions were not sensitive to the selection of the
GCPs as long as their distribution was reasonable, the 6 point
version, due to its very weak redundancy, depends a lot on the
point quality. As Table 1 shows the difference between the 10 and
20 point version is minimal. As a conclusion we can state that for
the given sensor model, 10 control points with quadratic attitude
rates and image measurements by matching lead to good results.
For the fore and aft images of MOMS-02 an accuracy of 6 - 7 m,
i.e. ca. 0.5 pixel, for all three coordinates was achieved.
4. DTM GENERATION
DTM generation was performed automatically using a
modification of the MPGC algorithm (Baltsavias, 1991). MPGC
is based on LSM and extends it by use of geometric constraints to
reduce the search space and simultaneous use of any number of
images. The constraints lead to a 1D search space along a line,
thus to an increase of success rate, accuracy and especially
reliability, and permit a simultaneous determination of pixel and
object coordinates. The measurement points are selected along
edges that do not have a direction similar to the direction of the
geometric constraints line. The approximations are derived by
means of an image pyramid. The achieved accuracy is in the
subpixel range. The algorithm provides criteria for the detection
of observation errors (i.e. erroneous grey levels) and blunders,
and adaptation of the matching parameters to the image and
scene content. The modified MPGC makes use of the PMFS to
constrain the search along pseudo epipolar lines (Baltsavias and
Stallmann, 1992) and has been previously used for SPOT images
(Baltsavias and Stallmann, 1993). In its current implementation it
can be used only for two images.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996