4 2ì 
A Combined Shape-From-Shading and Feature Matching Technique 
for the Acquisition of Ground Control Points in SAR Images of Rugged Terrain 
M. Adair 
Intera Technologies Ltd. 
Ottawa, Ontario, Canada 
and 
B. Guindon 
Canada Centre for Remote Sensing 
Ottawa, Ontario, Canada 
ABSTRACT 
This paper describes an image feature matching 
technique which can be used as a tool in ground 
control point acquisition for SAR images of rugged 
terrain. Topographic features, such as valley 
networks, can be readily extracted from both 
digital elevation data and from SAR images using 
shape-from-shading. First, in each network over 
lay, individual valley pixels are linked into 
contiguous segments. Valley pixels which are 
logically linked to two or more segments are 
denoted as nodes. An iterative relaxation method 
is then employed to identify node correspondence 
in the overlay pairs. Matching experiments have 
been carried out with SEASAT coverage of the Adams 
Lake region of central British Columbia and an 
available 1:250000 scale DEM. Image coordinate 
transformation errors of approximately 50 meters 
RMS have been achieved. 
INTRODUCTION 
The utility of spaceborne SAR systems will be 
greatly enhanced if the data can be rectified to 
user specified map projections, generally through 
the use of ground control points. This is rendered 
difficult in the case of SAR imagery of rugged 
terrain because the radiometry is dominated by 
terrain effects, which effectively masks point 
like features, and due to geometric distortions of 
foreshortening and layover. Terrain can be viewed 
as 'rugged' from a SAR point of view if it 
exhibits local slopes whose steepness imply near 
layover conditions. For SEASAT this implies 
regions with extended slopes of 20 degrees. 
Control points can be acquired through the 
correlation of real and simulated image chips, 
however this requires that the DEM exhibit the 
same level of spatial detail as the image itself 
(Guindon and Maruyama, 1986). In the case of 
spaceborne SAR sensors with spatial resolution of 
approximately 25 meters, accurate 1:50000 scale 
terrain information is required. However, for most 
of the earth's surface, only lower quality DEM's 
are available. 
This paper describes an alternative approach to 
the image rectification problem using a feature- 
based image matching technique. Ground control 
results from the correspondence of spatial point 
patterns of nodes in the valley floor network in 
areas of rugged terrain. In one case the network 
is derived from a low resolution digital elevation 
model (DEM) and in the other case it is derived 
from a SAR image using a shape-from-shading 
technique which extracts the topographic structure 
of terrain from a single SAR image (Guindon, 
1989). The processing is carried out in three 
steps: (1) detect the valleys in a DEM and trans 
form this valley 'mask' to a SAR projection (the 
'DEM mask'); this mask can then be compared with 
the valley mask extracted using the shape-from- 
shading technique (called the 'SAR mask'), (2) 
linking and labelling each of the masks to detect 
points where two or more valley segments meet 
(called 'nodes') and, (3) matching of the spatial 
point pattern of the nodes using an iterative 
relaxation technique. This set of matched nodes 
can, in principle, be used as a source of control 
points for image rectification. Each of the steps 
are described further below. 
DEM VALLEY EXTRACTION 
The detection of valleys in the DEM is carried out 
in two passes over the DEM and is based on the 
method of Qian et al. (1990). The first pass 
labels pixels as candidate valleys by examining 
the elevation values of three pixels in each of 
four directions (horizontal, vertical and the two 
diagonals), looking for a local minimum at the 
centre pixel. The second pass over the DEM locates 
the valley floor from the candidate pixels by 
tracing the path of steepest descent from pixel to 
pixel. This gives a vector representation of the 
valleys which is then transformed to the nominal 
SAR projection to match the SAR mask. In the 
present case the SAR projection is a ground range- 
azimuth projection with terrain height related 
parallax. 
The DEM of the study area, near Adams Lake, B.C., 
was rasterized on a 50m grid using 1:250000 scale 
digital elevation contour data provided by the 
Canada Centre for Mapping. The region is charac 
terized by elevation variations of 1500 to 2000 
meters and terrain slopes of up to 20 degrees in 
steepness. 
Portions of the DEM and SAR valley masks are shown 
in Figures 1 and 2. The SAR mask is derived from 
a SEASAT SAR image with 50m pixel resolution. 
While there are differences in the local details 
of each valley segment, the main segments are 
visually similar. The SAR mask shows valley 
segments which run mainly in the azimuth direction 
(top to bottom) due to the radiometric sensitivity 
to the range component of terrain slope (Guindon, 
1989). Other differences are due to layover in the 
mask derived from the radar image. 
MASK LINKING AND LABELLING 
The next step of the process is to link and label 
both of the valley masks. Every pixel is labelled 
as part of a valley segment or as a node. The same 
linking algorithm is applied to both valley masks. 
811 
?, "■ ' ‘ ■' ' I'f- ■ : 
8*$ téià 8 èìÈ é 
Wmm