Linear Feature Based Matching of 
Stereo SPOT Satellite Images 
Nigel Butler 
Research Student 
School of Surveying 
University of New South Wales 
PO Box 1, Kensington 
Sydney, NSW, 2033 
Australia 
nigel@spectrum.cs.unsw.oz.au 
Commission IV 
Abstract 
A new technique utilizing Feature-Based Matching of 
stereo SPOT satellite images to derive Digital Elevation 
Models (DEMs) is presented. DEMs were derived by ex- 
tracting and then matching linear features from stereo im- 
age pairs. The features were extracted by convolving the 
images with 2x2 pixel windows and grouping the image 
pixels upon similar gradient orientation. A pair of 8 mutu- 
ally exclusive binary images are produced and are labelled 
via a connected-components algorithm. Lines are fitted 
to regions and parameters calculated providing a rich set 
of attributes, which are then used to match corresponding 
features in the stereo images. Heights are derived from 
these matches and manually checked, with accuracies ap- 
proaching the pixel level. These points are also triangu- 
lated into a network of nearest-neighbors. The network is 
interpolated onto a regular grid and one of the images of 
the stereo pair may be draped over the generated DEM. 
Key Words: Feature Extraction, Image Matching, 
Stereoscopic, SPOT, DEM. 
1 Introduction 
Feature Based Matching of stereo satellite images was per- 
formed to obtain a Digital Elevation Model (DEM) by 
extracting and then matching linear features from stereo 
pairs. In recent years the focus on the solution to auto- 
mated stereo matching has shifted from gray-level corre- 
lation to feature-based matching (Greenfeld and Schenk, 
1989). Feature and area based image matching has been 
tested and compared with hybrid approaches (Brockel- 
bank and Tam, 1991). Very few projects attempting to as- 
sess the feasibility of using SPOT stereodata as a source of 
height information have been carried out and few results 
have been presented (Theodossiou and Dowman, 1990). 
The technique may be applied to images acquired both in 
the near and far range, however this paper is concerned 
with the far range. The stereo analysis problem as in pre- 
vious treatments (Barnard and Fishler, 1982, Medioni and 
Nevatia, 1985) may be broken into the following steps: 
® image acquisition, 
e camera modelling, 
908 
e feature acquisition, 
© image matching, 
* distance (depth) determination and interpolation. 
Extensive work has already been completed in the cam- 
era modelling of the SPOT stereo system and the Syd- 
ney images in particular. The collinearity equations, the 
epher1eris data from the header of the SPOT images and 
a set of known control points were used to specify the 
camera model. Precision of computation of object coor- 
dinates have been shown to be of the order of 5-10 me- 
ters in planimetry and height coordinates depending on 
the precision of the ground control points (Trinder et al., 
1988). 
The focus of this project is on the last three stages of the 
list above. Figure 1 shows the image matching process 
followed. 
2 Image Acquisition 
Since the launch of SPOT-1 on the 22nd February 1986, 
the production of topographic maps from space on an 
operational basis is now a possibility. SPOT acquires high 
resolution imagery of almost all of the earth’s surface and 
on the basis of coverage per single image, its imagery is 
cheaper than aerial photography (Trinder et al., 1988). 
The test stereo pair of SPOT images consists of scenes 
of Sydney and its metropolitan area. One was imaged 
on the 22nd November 1986 with a left incidence angle of 
30.1° and the other on the 12th October 1986 with a right 
incidence angle of 21.3° (i.e. six weeks between dates of 
acquisition). The Base to Height (B/H) ratio of the pair 
was therefore 0.97. 
Different stereo applications often involve different kinds 
of scenes. Perhaps the most significant and widely recog- 
nized difference in scene domains is between scenes con- 
taining cultural features such as buildings and roads, and 
those containing only natural objects and surfaces, such 
as mountains, flat or ‘rolling’ terrain, foliage, and water. 
Industrial applications, on the other hand, tend to involve 
artificial, cultural objects exclusively. Cultural features 
present special problems. For example, periodic struc- 
tures such as the windows of buildings and road grids can 
confuse a stereo system.