(2)
r is the trace of
1 0.4 and 0.6.
ection and their
interest value of
racted by Harris
cal maximum in
;e interest values
raw image. In
current pixel to
ffective operator
s are: (1) easy
is used and no
e distribution of
st value of each
; selected in its
rator because of
Hatching
should be found
simulated SAR
tR image can be
irse registration,
jseful method to
iation theory and
ty between two
ages by mutual
;en widely used,
lg the statistical
on in gray level
ion histogram of
ion to the image
nation:
(3)
?))
i B respectively
es.
; two images are
ation is searched
n two images is
points, we will
ose center is the
;, and then its
corresponding location in real SAR image is calculated using
the coefficient obtained by coarse registration. Each location in
the real SAR image within a scope is searched by matching the
window in the simulated image with that in the real SAR image
until the maximum mutual information is founded and the
corresponding location is the homologous point.
4. ORTHO-RECTIFYING BY TIN SMALL REGION
Lots of homologous feature point couples distributed in the
simulated SAR image and the real SAR image provide a good
control for fine image registration because these points are tie
points between the real SAR and simulated SAR and they can
be used to build the precise pixel-to-pixel relationship. It is an
ideal solution to construct many small image regions using
these homologous feature points and then rectify those small
regions one by one. It is an easy and very effective method to
construct many triangle regions using triangle irregular network
(TIN). The vertices of the triangles are feature points and TIN
can cover the whole image. The mapping relationship between
the vertices of two triangles is known, so the relationship of the
pixels within the triangle can be calculated by affine translation.
Therefore the simulated image can be registered with real SAR
image finely. Figure 3 shows the mapping relationship between
one triangle in real SAR image and that in simulated SAR
image as well as the mapping relationship between the
simulated SAR image and DEM.
Homologous feature points
Figure 3. Map relationship between small regions and DEM
The real SAR image can be rectified based on the mapping
relationship between simulated SAR and DEM if the simulated
SAR image is finely registered with the real SAR image.
Therefore a pixel in real SAR image is output into the
corresponding position of DEM by the help of the
corresponding simulated SAR pixel. Figure 4 demonstrates the
realization method to rectify the SAR image based on DEM in
detail.
Figure 4. Principle of ortho-rectifying SAR by DEM
Feature points detected in the simulated SAR image and TIN
constructed by these feature points are shown in figure 5, while
the homologous feature points extracted in the real SAR image
by mutual information matching and the constructed TIN are
also shown in figure 6. The top-right part in figure 5 and figure
6 is a magnified part respectively. We can see that the extracted
feature points and constructed TIN are correct. Figure 7 shows
a Radarsat-1 SAR image with a geometrical resolution of
12.5m by 12.5m and its most parts are mountains, and figure 8
is the actual relief map of the same area. The geometrical
distortion of the SAR image is very serious by comparing the
SAR image with the relief map, because SAR image is side
look. Figure 9 is the orthorectified image using DEM by the
method proposed in the paper, the distortion owing to terrain
undulation is eliminated and the rectified SAR image is very
similar to the relief map shown in figure 8. Figure 10 and figure
11 show the local magnified result of orthorectified SAR and
relief map in the same area.
We merge the orthorectified SAR image into the relief map in
order to evaluate the accuracy. Figure 12 gives the merged
result and it is very satisfactory. We also check the accuracy by
measuring the coordinates of 16 outstanding points in DEM and
the orthorectified SAR image. The actual coordinates,
orthorectified coordinates and their difference are listed in table
1. The rectified horizontal accuracy reaches 27.46m which is
about 2 times bigger than the resolution of the SAR image.