Automated terrain corrected SAR geocoding

Urs Wegmiiller
Gamma Remote Sensing
Thunstrasse 130, CH-3074 Muri b. Bern, Switzerland
Tel: +41 319517005, Fax: +4131 9517008, email: wegmuller@ gamma-rs.ch

system parameters, and uncertainties in the map projection

ABSTRACT definition usually result in slight geocoding errors, which are
corrected, in a refinement step. Traditionally, this refinement
A technique for automated terrain corrected SAR
was done based on operator selected ground control points.
geocoding is presented. Instead of operator selected control
The key idea of the presented automation of the geocoding is
points the presented method uses cross-correlation with a
the simulation of a SAR image based on a Digital Elevation
SAR intensity image simulated from the Digital Elevation
Model (DEM). The requirement of a DEM is not very
Model to refine the geocoding transformation.
restrictive as it is used for terrain corrected geocoding
anyway. An automated cross correlation analysis between the
INTRODUCTION
simulated and real SAR images is used to determine the
Geocoding is the transformation between the coordinates geocoding refinement. The refinement applied to the initial
of an imaging system, in our case the range-Doppler lookup table leads to the refined lookup table which allows to
coordinates of the SAR, and orthonormal map coordinates. do the actual data resampling in a single step.
Geocoding is necessary to combine the information derived For a flow chart of the geocoding sequence see Figure 1.
from the SAR data with in-situ data or information of other
sensors. In addition, it is used for 2-pass differential SAR
interferometry and for the terrain slope corrected Digital elevation model (DEM)
1 SAR image

7
normalization of the SAR backscattering coefficient.
The most tedious and time consuming step in SAR
geocoding is the search for ground control points because of

~J
the difficulties associated with the precise identification of initial geocoding lookup table and
corresponding points in the SAR image and map. The very simulated SAR mtenshy Image
different appearance of the SAR image and the map as well
as the speckle noise present in the SAR image are the causes
registration of simulated with real SAR intensity image
for this. In addition, many of the targets which can be
identified should not be used because of the effect of the
target height on the target position in the SAR image.
Examples for this are houses, bridges, and trees. lookup table refinement
In this paper a way to automate this step is presented. After
an overview of the technique key steps as the parametric
geocoding transformation, the SAR image simulation based
on a DEM, and the automated fine registration between the resampling between SAR range-Doppler
simulated and the real SAR images will be discussed. Finally, and DEM map geometry
the accuracy and robustness of the presented geocoding
technique will be discussed.
Figure 1: Flow chart of automated terrain corrected SAR
AUTOMATED SAR GEOCODING METHODOLOGY geocoding.

The two main steps of the presented geocoding technique PARAMETRIC GEOCC)DING STEP
are the calculation of the transformation lookup table and the
resampling of data files using this lookup table. The lookup In the parametric geocoding step the DEM map geometry
table contains for each pixel of the map (i.e. the geocoded and the SAR range-Doppler imaging geometry are used to
images in the desired map projection) the corresponding determine the initial transformation lookup table. For each
coordinate in the SAR range-Doppler geometry. pixel of the DEM the DEM coordinate is first transformed
The calculation of the lookup table is done in several steps. from the map coordinates to Cartesian coordinates . Next the
In a first step a parametric description of the orbit, the SAR Datum shift between the Datum of the reference ellipsoid
imaging geometry, and the map projection, considering the used for the DEM map projection and the reference ellipsoid
local terrain height, is used to calculate an initial geocoding used for the description of the sensor orbit geometry, are
lookup table. Errors in the available orbit data, the SAR corrected for. Then the acquisition time and position

0-7803-5207-6/99/$10.00 (c) 1999 IEEE

corresponding to the selected map coordinate are determined. correlation analysis between the simulated and the real SAR
Finally, the look vector to the selected map coordinate is image, with the simulated SAR image representing the map
calculated leading to the desired azimuth and slant range geometry. For a large number of small image chips the
position. Image skew (for non-deskewed images) and the corresponding location in the other image is searched for.
local terrain height are taken into account in these Apart from the local offset estimates a quality measure is
transformations. determined for each estimate. Based on the reliable estimates,
A variety of map projections are supported. The modular i.e. the estimates with a quality measure above a certain
concept easily allows to include additional map projections threshold, the fine registration function is calculated using a
given the corresponding equations. The map projection, least squares fit.
reference ellipsoid, and Datum shift parameters are defined in
the DEMM4AP parameter file.
SAR
SAR IMAGE SIMULATION FROM DEM

After the above transformations the SAR look vector and

the terrain surface coordinates are available in the same
Cartesian coordinate system, allowing to directly calculate
geometric parameters such a the SAR look vector, the image
plane normal and the surface normal, the local incidence
angle, and the projection angle. For a detailed discussion of
the imaging geometry it is referred to [1]. The true local SAR
pixel size A [mz] for a tilted surface is calculated by
az”r
A=—
Cosy (1)
with
az: azimuth pixel spacing [m]
r: slant range pixel spacing [m]
v: projection angle, defined as the angle
1“
between the surface normal and the image I If
plane normal (see Fig. 2).

The backscattering of the simulated SAR image is found

by multiplication of the calculated pixel area with an .
empirical function of the local incidence angle which is used v
to account for the incidence angle dependence of the Figure 2: SAR imaging geometry with vectors: x (azimuth),
backscattering coefficient. The image simulation does not z (vertical), r (radar look vector), R (Earth center to pixel), S
consider the dependence of the backscattering on the surface (Earth center to SAR), n (surface normal), and m (image
type. As a result an almost constant value is obtained for flat plane normal) and angles: 6 (local incidence angle), v
areas. Figure 3 shows an example of a simulated SAR image (projection angle), and U,V(spherical angles of n).
for a part of Death Valley, USA. The valley floor, in the
center of the image, is flat. The simulated backscattering is In the case of the Death Valley e~ample (Figures 3,4), no
almost constant. The real SAR image (Figure 4) shows a lot reliable estimates are found for the image center where the
of backscatter variation for the same area because of the surface type clearly dominates the backscattering variations.
different surface types found (see [2]). For the sloped areas, A sufficient number of reliable offset estimates was found,
on the other hand, the local topography has a strong influence nevertheless, for the more rugged areas, allowing to
on the backscatter. determine the fine registration function.
The fine registration function is used to modify the
FINE REGISTRATION STEP geocoding lookup table resulting in what we call the refined
lookup table. The values of the refined lookup table indicate
The purpose of the fine registration step is the
for each pixel the corresponding location in the SAR image,
improvement of the geocoding geometry over what was
The reasons to apply the refinement to the lookup-table are
calculated with the parametric transformation equations. As
the intention to conduct the SAR ,geocoding in one single
fine registration function a hi-linear function (range and
resampling step to minimize the effects of data interpolation
azimuth offsets are linear functions of range and azimuth) is
and for efficiency reasons.
used. The registration function is determined in a cross-

0-7803-5207-6/99/$10.00 (c) 1999 IEEE

Figure 3: Simulated SAR intensity image corresponding to Figure 4: ERS- 1 SAR intensity image (5-look) over Death
SAR intensity image shown in Figure 4. Valley, USA.

ROBUSTNESS AND REGISTRATION ACCURACY In most cases the geocoding accuracy achieved with the
presented method is high because of the high number of
The presented method was successfully applied to the
locations used to determine the fine registration function.
geocoding of ERS and JERS SAR data. Not only high
Once the refined geocoding lookup table is determined the
resolution DEMs but also DEMs with moderate resolutions
geocoding is done in a single resampling step.
between 100m and 250m were used.
The presented method is efficient as no operator selected
For ERS the automated geocoding worked even in
control points are required. The technique is robust in the
relatively flat terrain with a maximum scene height variation
sense that quality measures for the success of the automatic
below 100trz. For JERS the robustness of the technique in
fine registration are determined, indicating that the fine
rather flat terrain is not as good as for ERS because of the
registration was not successful if no reliable offsets could be
much stronger influence of the surface class on the
determined, as it may be the case for flat terrain.
backscattering.
In most cases the geocoding accuracy achieved was higher
REFERENCE(S
than the pixel spacing of the DEM used. The standard
deviation of the local offset estimates from the fine [1] Ulander L., Radiometric slope correction of synthetic-
registration function serves as a measure for the reliability aperture radar images, IEEE TGR.S Vol. 34, No. 5, 1996.
and accuracy of the fine registration. [2] Wegmuller U., T. Strozzi, T. Farr, and C. Werner, Arid
land surface characterization with repeat-pass SAR
CONCLUSIONS interferometry, Proc. 2nd int. workshop on Retrieval of
Bio- and Gee-physical parameters from SAR data for land
A method to automate the fine registration step required in
applications, 21-23 Oct., ESTEC, Noordwijk, The
SAR geocoding was presented. Based on a DEM a SAR
Netherlands, ESA SP-441, pp. 565-570, 1998.
intensity image is simulated. For a large number of locations
the registration offset between the simulated and real SAR
images is determined using intensity cross-correlation.
Reliable estimates are used to determine the fine registration
function.

0-7803-5207-6/99/$10.00 (c) 1999 IEEE