lon necessary
he image to
not). As a
tion could be
ollows. Two
chosen, and
n the image.
in the map,
g their zero-
| between the
azimuth time
1e across one
onds, which
ion threshold
slant range.
st, centre and
1024 lines in
-doppler, but
doppler. The
linate and the
ve, Was used
same sample
were used to
round range.
all projection
measured in
chose to do
rather than
the overall
IL correction.
ite an energy
be used as a
1ired tiepoint
ally using an
1 both the
input image,
und range to
proximations
from ground
is tiepointing
et the ground
projecting
. However,
rors between
ved a greater
usually very
features that
th published
he approach
from terrain
the azimuth
:d, the image
1e geocoding
50 generated:
servation; In
'onservation.
33 The geocoded result
In the range direction, the DEM slope is
approximately in the range [-20°, 27°]. Shadow would
be generated only at incidence angles greater than 70°,
and layover only at incidence angles less than 27°.
The scene centre incidence angle of JERS-1 is
nominally 35°, therefore no shadow or layover from
the DEM was to be expected. This is indeed the result
which was obtained.
An important internal check on the geocoding is to
compare the image space layover map against the
saturated regions of the input image, and also to
compare the map space shadow and layover maps
against the saturated regions of he geocoded image.
However, as these maps are (correctly) empty, all that
can be done are comparisons between the energy
conservation maps and the images. These show good
geometric and radiometric agreement.
34 Problems with JERS SAR
There were many problems involved in determining
from NASDA the exact definitions of the parameters
in the headers. Some of these problems were solved
late, some not at all. The consequence of this is that
our geocoding process is still not properly matched to
the standard JERS-1 SAR product. In this section, we
discuss the remaining problems.
The coordinate system for the orbit is stated to be
"ECR". "GSFC" is also mentioned in the same
context. We have used WGS84 with no conversion.
The actual relationship between these coordinate
systems is not known.
We have treated the supplied range samples as if they
were zero-doppler. It was subsequently confirmed that
this is not the case. There has not been time to make
appropriate modifications, but it is believed that the
necessary parameters do exist in the headers. This
does not invalidate the results above, as the tiepoint
correction of the overall projection corrects for this.
We need to be able to relate the azimuth coordinate of
the image to azimuth time. This is necessary so that
use can be made of the orbit. We used a workaround
involving two tiepoints and a linear approximation
(described above) as a placeholder, hoping that
NASDA would help us obtain the necessary
information. Eventually, NASDA confirmed to us that
the information does not exist in the standard product.
It may exist in lower level products.
Amongst the various parameters in the headers of the
standard product are corner coordinates of the image,
given in geographic and map coordinates. In the
product geocoded above, these were based upon the
ellipsoid GRS80 and the map projection UTM zone
31 northern hemisphere. Projection of the corners to
slant range would provide the necessary relation
between the azimuth coordinate of the image and the
azimuth time. However, investigation showed these
439
corner coordinates to apparently be in error,
preventing this approach. The error is large (1 - 4km),
and cannot be accounted for by datum errors. NASDA
could not help.
For the purpose of relating the azimuth coordinate of
the image to azimuth time, the supplied scene centre
time differs significantly (almost 600 pixels) from the
value derived by tiepointing. This would, anyway,
only have provided half of the necessary information.
The tiepoint correction of the overall projection from
the map to the image can be performed using points
measured in the map, or by using points measured in a
"simulated" image, as done above. Both approaches
have drawbacks. The use of published maps
significantly limits the number of measurable tiepoint
pairs, due to differences in content from that of the
image. The use of an image simulated from the terrain
gives poorer precision in the azimuth direction, due to
the nature of features in such an image. Neither of
these problems are specific to JERS-1 SAR; they are
general to all SAR.
4. PRINCIPLES OF STEREO SAR FOR DEM
PRODUCTION
The use of overlapping pairs of SAR images for the
production of 3-D data has been outlined in Dowman
et al (1992) and initial results presented in Dowman
et al (1993b) . The process requires 3 stages:
preprocessing of detected SAR images; stereo
matching to determine disparities between the two
images and transformation of the disparities to heights
in a ground reference system.
The preprocessing to remove the effects of speckle is
discussed in section 5 of this paper.
The method used for stereomatching is the
CASCADE algorithm (Denos 1991) which
automatically determines seed points in the top layer
of an image pyramid, these are then matched using the
Gruen adaptive least squares algorithms in the top
layers of the pyramid whilst in the finer layers the
Otto-Chau region matching algorithm, developed
from the Gruen approach, is used.
The final stage is the transformation of disparities to
3-D data. In the current work the method used is that
of Clark (1991) , which uses the range and Doppler
equations for the two images to determine the vector
P which is the co-ordinate vector of the ground point,
using known satellite position S, .and S, and the
satellite velocity. The principle is shown in Figure 1.
Clark has shown that the method is sensitive to errors
in timing and that the method will be more stable with
a longer base length. Work at UCL with ERS-1 data
(Chen 1993) has shown that the method works but
requires improvement to give satisfactory results.