International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
bias nor drift are considered. This modification of the stochastic
model allows to fit the HRSC object points to the MOLA DTM,
which is introduced with 00 prm = 100m for the terrain points.
In this part the absolute position of the image strip is adjusted in
an iterative way.
Both parts use blunder detection. In the first part blunders in the
tie points are detected iteratively by analyzing the residuals. In
the second part the difference between the MOLA surface and the
HRSC points is used to find non-fitting points (Spiegel, 2007b).
2.4 Evaluation
To evaluate the performance of the bundle adjustment, the quality
of the nominal and adjusted exterior orientation data is compared.
Because of the large amount of data, a single value describing the
geometric quality is easier to handle for systematic analysis of
the whole data set. The mean forward ray intersection error sum-
marizes the accuracy of all object points and is a good measure
for the internal consistency of the data (Gwinner et al., 2009).
The evaluation step calculates the forward intersection error for
all object points twice. Both times it uses the same set of tie point
coordinates and only switches between nominal and adjusted ori-
entation data, so that the resulting mean intersection error reveals
the gain in quality.
Additionally, both the nominal and the adjusted orientation in-
formation is used to generate dense 3D point clouds as used for
DTM generation (Gwinner et al., 2010). The intersection error
for all these points is used to plot two color coded error maps.
This allows a visual inspection of the spatial distribution and ac-
curacy of the HRSC points. Fig. 3 shows three examples of er-
ror maps. Strip h7295.0000 displays large black areas where the
dense matching algorithm was not able to find conjugate image
points. The color of the small number of points, shows a low
accuracy. There is no significant improvement through bundle
adjustment. This is also indicated by the values for the mean in-
tersection error (38.6m and 37.7m). In contrast, the error maps
and mean intersection error of h2091.0000 and h8500.0000 show
considerable improvement after bundle adjustment despite of the
high frequent oscillations visible in strip h8500.0000.
h7295 0000 h2091 0000 h8500 0000
before / after before / after before / after
intersection
error
higher
5 lower
mean values
38.6m 37.7m 46.3m 6.8m 443m 3.5m qu
Figure 3: Error maps with mean intersection error of the object
points (black: no points)
3 MULTI ORBIT BLOCKS
For the processing of photogrammetric blocks containing images
from more than one HRSC strip an adaption of the same work-
flow for single orbit strips (Fig. 2) can be used (Spiegel, 2007b).
To efficiently process large image blocks a suitable strategy has
to be set up. It is not reasonable to process all image strips of
the entire block together because only neighboring image strips
overlap. Therefore, the whole image block is broken down into
smaller sub-blocks which additionally opens up the advantage of
parallel processing. If a sufficient number of server nodes re-
spectively CPU cores is available, it is possible to run three of the
four processing parts (image pre-processing, tie point matching
and evaluation) for all sub-blocks simultaneously. This drasti-
cally saves computation time.
For the tie point matching each sub-block needs a reference im-
age which is matched against all other overlapping images. Every
strip of the entire block has to supply this reference image once.
If e.g. the block consists of two overlapping strips with ten sin-
gle images in total the nadir channel of the first strip is matched
against all other nine images. The second sub-block in this exam-
ple would simply consist of the second strip. At this point merely
the remaining area which was not captured by the first strip is
covered.
entire block 1. sub-block 2. sub-block 3. sub-block
Figure 4: Concept of sub-blocks using the example of three strips,
green: 5-ray points, red: 10-ray points
This strategy on the basis of three overlapping strips is depicted in
Fig. 4. Three sub-blocks are built up in the combinations 1 & 2, 2
& 3 and 3 separately. An advantage of this approach in particular
with strips of inhomogeneous geometry is that for each sub-block
an optimal average matching scale can be defined. In case of
processing all image strips simultaneously an average matching
scale for the complete block would have to be found.
Concerning computational limitations the number of tie points
per sub-block has to be limited. Tests have shown that if a suf-
ficient number of tie points in all overlapping areas is available,
a further increase of the number of points does not improve the
results. While it is clear that this strategy will fail for crossing
strips as to be found e.g. on the poles it is good compromise with
regard to computing power.
In the subsequent bundle block adjustment the results of the sub-
blocks are imported and processed simultaneously. As approxi-
mate values of the exterior orientation for the pre-rectification of
the images the results from single orbit strip processing are used.
The reason is that while images of one strip normally exhibit a
high relative accuracy, the discrepancy between neighboring im-
age strips can reach several hundred meters which would require
very large search spaces for matching. When employing adjusted
image strips the orbits fit much better among each other so that
smaller search spaces can be used.
The iterative blunder detection during bundle adjustment is con-
ducted initially for all strips independently as decribed above, and
286
QM m9 M Ph t t A WA mh At AS m0 RR Pla NY ss
