■ S
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
45 Test Orbits
♦ Before i After
Figure 8: Standard deviations [m] of the ray intersections for all
45 test orbits in Z (height component) before and after BA
Table 1 shows the standard deviations of the ray intersections
averaged over all 45 test orbits. The respective left column
shows the value before adjustment using the nominal exterior
orientation. The right column shows the result of the bundle
adjustment (labelled “BA”). Results for Cross Correlation only
and LSM are given.
blunders is about 10% including the points eliminated which do
not fit to MOLA and that 10000-60000 points per strip have
been found.
Next, in the second step the HRSC object points are tied to the
MOLA DTM. For this task, the a priori accuracy of the exterior
orientation has been introduced into the bundle adjustment with
a value of 1000 m for the position and 25mdeg for the attitude
at the orientation points. The MOLA DTM is introduced with
an accuracy of 100 m in order to cope with differences between
HRSC object points and MOLA track points due to the limited
spatial resolution of MOLA. As mentioned before, the
resolution on the ground of HRSC is up to 12 m compared to
the MOLA surface footprint of about 168 m. Regarding local
areas, the MOLA data describe the surface less detailed as
HRSC object points.
The root-mean-square (RMS) differences between the HRSC
and the MOLA DTM is reduced significantly in the bundle
adjustment. Hence, there is a high consistency between HRSC
points and the MOLA reference system after the bundle
adjustment. This is clearly visible in Figure 9 which shows the
height differences between the HRSC object points and the
MOLA DTM of orbit 2063 before and after bundle adjustment.
Method
X
X„ A
Y
Y ra
Z
ZßA
CC
15.6
7.3
12.2
5.5
55.5
25.6
LSM
14.5
4.5
11.3
3.4
50.9
15.8
-•
i
Table 1. Standard deviations [m] of the
averaged over all 45 test orbits
ray intersections
Due to a small convergence angle of the HRSC the height
component cannot be determined as precisely as planimetry.
Before bundle adjustment the points derived from LSM are
slightly more accurate. The points derived from CC can be
improved by a factor of 2.2 by bundle adjustment whereas the
points derived from LSM have been improved by a factor of 3.2
With respect to an average resolution of 30 m of the prerectified
images a very high accuracy has been achieved. In the next
table the a posteriori standard deviations of the image
coordinates and the percentage of rays of the tuples have been
verified:
> 80 m
Method
Accuracy
of image
coordinates
2-ray
[%]
3-ray
[%]
4-ray
[%]
5-ray
[%]
CC
0.32
8.0
9.6
8.7
73.7
LSM
0.19
8.5
9.6
8.4
73.5
Table 2. A posteriori standard deviations [pixel] of the image
coordinates and the percentage of rays of the tuples
averaged over all 45 test orbits
It can be seen that with CC a high accuracy of 1/3 pixel has
been attained. With LSM an increased accuracy of 1/5 pixel
has been achieved. For a stable bundle adjustment a high
percentage of 5-ray points is desirable and the percentage of 2-
ray points should be as small as possible because they are
omitted in the bundle adjustment. About 3/4 of all tuples
consist of all five rays and the remaining values are below 10%.
This means that matching delivers rather strong geometric
blocks. Finally, it should be noted that the percentage of