ul 2004
ata and
points
itton of
mainly
are not
ibution
1e line.
nalysis.
search
1c SAR
| x- and
id each
onding
ed to a
lyse the
. Streets
eas are
| image
side are
reside.
| area is
suming
' search
omated
ge.
2ss con-
be seen
the left
s errors
ntly the
nly the
tions in
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
It is therefore possible, that, as a result of a failed search, some
corresponding points are placed in wrong positions. More
points than needed for the transformation, should therefore be
searched for. The analysis should concentrate on areas with
many streets and junctions, because these areas offer a more ro-
bust determination.
After the analysis of all areas, the corresponding points are
available in the reduced resolution. The final position is deter-
mined in the high-resolution image, using the same method. As
the position of each corresponding point is roughly known, the
computational time for searching in the high-resolution image is
reduced considerably. After the required number of correspond-
ing points are known, the transformation can be calculated. By
analysing the transformation and the resulting errors, wrong
correspondences can be eliminated.
Figure 9. Supposed street position, shifted in near range
direction
Due to the SAR imaging properties, the streets surrounded by
houses, are displaced. The position of the dark areas in the
image, supposed to be streets, is shifted in near range direction,
due to the shadow and layover effects of the surrounding build-
ings.
The length of the shift in near range depends on the incident
angle and the height of the surrounding buildings. If the build-
ings on both sides of the street have the same height, the shift
can be calculated according to
height
shift =——>
tan(off — nadir)
Assuming a building height of 10m and an off-nadir angle of
70°, the shift in near range direction is ca. 3.6m. Using a pixel
size of 0.26m, this is equivalent to 14 pixel. If all the buildings
in the analysed area have nearly the same height, the streets are
just shifted in near range direction. If the building heights in a
given subset vary a lot, the search method described above may
fail.
1 a s
14
E hdd LE 3 S ^
aying a DOSAR image
In Figure.10, the shift can be seen in the results. Three neigh-
bouring search areas are presented and there is some shift in
range direction between these regions. This shift is mainly
caused by the effect shown schematically in Figure 9. Addi-
tionally the analysis does not provide a very high accuracy it-
self.
Due to the near range shift of streets in urban areas, the position
in range will be wrong for a lot of the subsets. Even worse, the
shift is changing according to the environment. In inner city
areas, the shift is very high, due to the high-rise buildings there.
On the other side, streets without neighbouring buildings will
have no such shift at all.
An accuracy of the overall spatial reference of around 6m is
achieved by the method described above. This error is caused
by the inaccurate street data, as well by the shifting of the
streets as shown in Figure 9. Erroneous corresponding points
are furthermore causing a lot of problems. Considering the high
errors caused by shiftings and by inaccurate data, it is not easy
to automatically find the erroneous corresponding points.
Therefore the overall accuracy is not high enough to use this
geo-referencing method for automatic change detection applica-
tions. But it can be useful for manual change detection and it
may be used as rough geo-referencing method for the whole im-
age, which should be refined during the change detection proc-
ess for a local area.
4.2 Model based geo-referencing using landmark buildings
The geo-referencing can be more accurate if a 3D-city model al-
ternatively to GDF-street data is being used. Special landmark
buildings of the 3D-city model are being used for the method
described in this paper. Landmark buildings are unique build-
ings with special properties, which are separated from the
neighbouring buildings, to avoid them from merging together in
the SAR image. The buildings have to be selected by an opera-
tor. It is also possible to find such buildings automatically, by
analysing the GIS data. After the building has been selected it
has to be simulated according to the sensor parameters of the
real SAR image. The SAR simulated image of the selected
model should look quite similar to the building in the real SAR
image.
The first analysis step is not performed based on a raster repre-
sentation of the data, but on the vector representation, because
the upcoming search can be calculated much faster using vector
data. For this purpose, the model of the landmark building is
simulated without any speckle and the resulting image is trans-
formed in a vector representation (see Figure 11).
Figure 11. SAR simulation of the St. Bernhardus church as
vector representation
