bi
m.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
data itself might be faulty. Moreover strong corner reflectors
and their side-lobes are able to disrupt every interpretation and
unfortunately they can not precisely be forecasted. Normally the
SAR processor should reduce the effect of side-lobes.
If the SAR processor is not suppressing the side-lobes suffi-
ciently, it is also possible to suppress the side-lobes by a filter-
ing in the frequency domain of the SAR image. Applying a
wedge filter, according to the squint angle, in the Fourier image
and comparing the original image with the filtered image, re-
veals the areas affected by the side-lobes. These areas have to
be masked out in the future image interpretation steps. Unfortu-
nately not all the areas are detected by this simple method and
some areas may be wrongly masked out.
4. SIMULATION BASED GEO-REFERENCING
4.1 Model based geo-referencing using street vectors
The first step of a change detection is to combine different data-
sets. This is done by geo-referencing the datasets. SAR data is
normally, at least, roughly geo-referenced, but this initial spatial
reference is often not accurate enough for data fusion purposes.
The initial reference of the DOSAR image of Karlsruhe, for ex-
ample, has an offset of 150m.
In the approach described in this paper, one possibility to im-
prove the spatial reference of the SAR image is using GDF-
street data as ground-truth for automated geo-referencing. The
other possibility is to use 3D-models of landmark buildings, but
such 3D-models are often not available, whereas street network
data is commonly available. These standard datasets are, e.g.
provided for car-navigation systems. The street vectors are
transformed to the UTM coordinate system. Afterwards the
street vectors should be SAR simulated using a DEM of good
quality, if available.
Streets in SAR images are normally quite dark, because the
street surface is very smooth and reflects the SAR beam away
from the sensor. Cars and signs are strong reflecting objects on
or near the street, but they are not taken into consideration here.
Therefore, it is assumed that streets should appear dark in SAR
images and can be found by their structure.
The flight direction and the rough position of the image have to
be known for further processing. In this example, we used the
DOSAR flight over Karlsruhe mentioned above. The initial co-
ordinates of this flight have an offset of about 150m. This
spatial reference should ‘ improved using GIS data. For this
purpose GDF-street data, has been used. This data has an
accuracy of around £3m (Walter, 1997).
For geo-referencing the SAR image, corresponding points in the
SAR image and in the street data have to be found. The cor-
responding points should be evenly spread over the image, to
allow a stable referencing of the data. On the other hand, for the
automated search method described in this paper, it is necessary
to use points from areas containing many streets and junctions,
to avoid ambiguities. Depending on the content of the SAR
image this is a trade off, because streets and junctions are
normally not evenly distributed. In this approach, a huge
amount of distributed points are selected. From those, only the
points with the most junctions in their search subset are being
used.
In Figure 6, the footprint of the SAR image, the GDF-data and
the search areas are visible. Obviously the corresponding points
are not very well distributed. This is due to the concentration of
the algorithm on areas with many junctions and streets, mainly
found in the city area. This results in points, which are not
evenly spread. Another problem is the unfavourable distribution
of the points. The selected points reside mainly on one line.
This is due to the quite large search area used for the analysis.
Therefore, the small strip-width forces the algorithm to search
for corresponding points near mid-range.
To minimize the computational time, the resolution of the SAR
image is reduced, in this example by the factor 5 in both x- and
y-direction. Using this reduced image, the area around each
search point is extracted and the streets in the corresponding
areas are rotated in azimuth direction and are transformed to a
binary raster representation.
The binary representation is used as search mask to analyse the
reduced SAR image chips. The algorithm assumes, that streets
are dark areas in the image, while the surrounding areas are
bright. Therefore the sum of the pixel values of the real image
in areas where, according to the street data, streets reside are
divided by the sum of the pixel values where no streets reside.
The pixel with the lowest calculated value in the search area is
the point with the highest concurrence. This time consuming
search method is working well in urban areas. Promising search
results can be seen in Figure 7, showing the good automated
matching, between the GDF-street data and the SAR image.
MS
A
Figure 7. Search results
overlaying a DOSAR image
But the method has some problems in rural areas with less con-
trast between streets and surrounding areas, which can be seen
in Figure 8. Obviously there is some shift in the data on the left
side, but the position is approximately correct. Gross errors
exist in the second example on the right side. Apparently the
position is totally wrong. The problem here, is not only the
lower contrast. The small amounts of streets and junctions in
the area, are causing the method to fail.
Fate LAS SP A
: a i uA Pe
aying a DOSAR image
474
Inte
It is
cort
poit
seal
mar
bus
Aft
avai
min
the
COIT
redi
ing
anal
cort
Due
hou
ima,
due
ings
The
ang]
ings
can
ASSI
70°,
size
in tl
just
give
fail.
