AUTOMATIC HEIGHT EXTRACTION FROM ERS-1 SAR IMAGRY 
Zway-Gen Twu , IJ. Dowman 
University College London 
United Kingdom 
Commission II, Working Group 4 
KEY WORDS: SAR, DEM Accuracy, Range Error, Intersection Angles 
ABSTRACT 
Various aspects of stereo height determinations from ERS-1 SAR imagery are described in this paper. A pyramidal stereo 
matching algorithm is applied on an overlapping stereo pair of PRI and RTM imagery. The factors that influence the DEM 
accuracy are analysed on four different seed points sets. These factors include the ways to select the seed points and the 
geometric constraint conditions for SAR intersection. With this standard approach of stereo matching an accuracy of 78m 
can be achieved for a DEM. It is shown that the accuracy of the DEM is closely related to the range errors and hence the 
intersection angles of the SAR data and that if this error can be controlled a much better DEM accuracy can be obtained. 
Experimental results produce a rmse of about 17 m for four different data sets. 
1. INTRODUCTION 
It is of interest today to study the creation of the Digital 
Elevation Model (DEM) created from the Synthetic 
Aperture Radar (SAR) for it can provide DEMs in areas 
which are not easily accessible to other optical sensors. 
In this field of research, previous work is mainly focused 
on SIR-B. In particular, Leberl and his group have 
written many papers to discuss the subject of stereo 
matching (Leberl,1986a), (Leberl,1986b), and Mercer, 
(1995) has reported on the use of stereo airborne SAR. 
But not many papers have been published regarding the 
DEM derivation from ERS-1 SAR. Compared with SIR- 
B, ERS-1 gives more accurate orbit header information, 
thus it should create a better DEM. 
An alternative method of creating DEMs from SAR data 
insinterferometry. Stereo. SAR is seen as 
complememtary to IFSAR and can be used where 
interferometry can not be applied. Work has not yet 
been carried out to analyse the best condition for 
implementation of stereo SAR. 
In UCL, the work on the ERS-1 SAR has been undertaken 
for a number of years (Dowman et al.,1992a) (Dowman et 
al., 1993) and useful results have ben obtained . The 
purpose of this paper is to report on an investigation 
into the production of DEMs from stereoscopic ERS-1 
SAR data. In this paper, the pyramidal matching 
algorithm is introduced and a new strategy is proposed to 
increase the DEM accuracy tremendously. 
2 PYRAMIDAL STEREO MATCHING 
Compared with conventional optical imagery, SAR has 
poor image quality which is affected by layover, noise, 
and speckle. Thus to stereomatch SAR, there are many 
problems encountered. To overcome these problems in 
UCL, a new approach is proposed to implement a coarse 
to fine pyramidal method. This pyramidal matching is 
called CHEOPS [named after the Great Pyramid of Cheops 
at Gizza near Cairo] (Denos,1992). The CHEOPS 
algorithm is capable of automatically generating the 
shell scripts required to match both SAR and other forms 
of imagery. It does this by interpreting a script 
describing the topology of image pyramids written in a 
simple language called PDL (Pyramid Description 
Language) and then converting this script into an 
equivalent set of executable UNIX shell scripts. In this 
paper, the PDL file is defined to use the Otto-Chau stereo 
matcher for each tier. The Otto-Chau stereo matcher is an 
area-based patch correlation technique which 
incorporates the Gruen's Adaptive Least Squares 
Correlation and a sheet growing algorithm. This stereo 
matcher performed very well in the SPOT imagery (day 
and Muller , 1989). The detail of this stereo matcher can 
be found in (Otto and Chau,1989). For the seed points, 
the CHEOPS uses the random seed points generated in the 
first tier of image pyramid. Some research at UCL takes 
advantage of CHEOPS in dealing with the matching 
problems in SAR image. Dowman et al., (1992b) first 
applied CHEOPS on ERS-1 SAR data with different modes 
and different angles combination. (Denos, 1991) 
implement the CHEOPS on the NASA Seasat satellite 
images of Death Valley, and tried 9 tiers achieve 
coverage of 81% over 1024 by 1024 imagery. 
3. INTERSECTION 
The intersection of ERS-1 SAR in this paper is the 
analytic approach, proposed by Clark to geocode the 
SIR-B imagery in her Ph.D. thesis (Clark,1991). The 
analytic approach primarily utilises two Doppler 
equations (1) (2) and two range equations (3) (4) to 
obtain the solution. 
2(S1-P)(S1-P) 
f = 
DCI MIS1 N- (1) 
2($9 - PXS5 - P) 
f ss up ed 
DC2 A2|S2 -P| (2) 
RI - [Si -P| (3) 
R2 - |S5 - P| (4) 
where imagel and image2 are the stereo pair 
fpciis the Doppler value for imagel 
fpc2 is the Doppler value for image2 
RI is the range distance in imagel 
R2 is the range distance in image2 
S1is the velocity of the sensor for imagel 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996