99 
Symposium on Remote Sensing for Resources Development and Environmental Management / Enschede / August 1986 
Spatial feature extraction from radar imagery 
G.Bellavia 
Computer Science, Aston University, Birmingham, UK 
J.Elgy 
Civil Engineering, Aston University, Birmingham, UK 
ABSTRACT: It is accepted that the major role of remote sensing as an 
information source will be in its contribution to geographical information 
systems. With the advances in remote sensing, images are being created at an 
increasing rate. The extraction of information from such data is traditionally 
done manually and is thus costly in both time and money. Therefore techniques 
need to be developed which automatically extract information from remotely 
sensed images. 
This paper considers the extraction of thin line features such as forest 
rides, dykes and streams from active microwave imagery. Because radar images 
are coherently created speckle is produced which renders traditional feature 
extraction methods virtually useless. It is assumed that global techniques 
such as generalised hough transforms or intelligent graph searching will be 
more successful than simple local methods. 
1 INTRODUCTION 
This paper considers the extraction of thin 
line features such as forest rides, dykes and 
streams from remotely sensed active microwave 
imagery. In the context of this paper remote 
sensing is the aquisition of digital images 
of the Earth from either airborne or 
satellite sensor systems. Although digital 
remote sensing started as late as 1972 many 
satellite systems have been successfully 
initiated with a rapidly increasing variety 
of sensor types and specifications. 
Several problems arise from this rapid 
advance in remote sensing technology. The 
efficient use of image data obtained from 
satellites or aircraft relies on the 
availability of human expertise and 
sophisticated computer systems. In particular 
radar imagery with its virtual continous 
monitoring capability and high resolution has 
shown itself to be superior to more 
conventional imagery for a variety of 
applications. But to fully utilise the 
potential information in this form of data, 
processing techniques which overcome the 
inherent speckle noise need to be developed. 
In general, the problem is one of storage 
and processing of the data, which is not 
currently met by computers or human effort. 
Hardy(1985) states that the population of 
Earth resource satellites will generate 
something approaching 10 16 bits in 1986. 
Assuming an average scene size of 4000x4000 
pixels and 8 bits/pixel radiometric 
resolution, a simple calculation shows this 
data rate to be equivalent to about 200,000 
single channel images/day. On common Computer 
Compatible Tapes (CCT's) which can store 
approximately 33 million bytes (MB), this 
data would require about 10^ CCT's/day. This 
rate signifies an increase of the data 
volume/sensor/unit land area of more than an 
order of magnitude over the last ten years 
and results in large quantities of unused 
data. To alleviate this waste of data there 
is a need for the automatic interpretation of 
remotely sensed images by computer. 
It is accepted that the major role of remote 
sensing as an information source will be in 
its contribution to geographical information 
systems. This role requires the extraction of 
semantic information from remotely sensed 
images and therefore automatic 
interpretation. 
The two arguments concerning data rates and 
GIS bring us to the conclusion that automatic 
interpretation is needed to fully realise the 
potential of remote sensing. 
The automatic interpretation process entails 
several other processes. One of the first 
steps is segmentation which aims to partition 
the image into separate distinct regions. 
Feature extraction techniques are also used 
to enable the segmentation. The segmented 
scene can then undergo a pattern recognition 
process using world knowledge giving an 
interpreted scene. Knowledge may also be used 
at different stages in the interpretive 
process. For example to assist or iteratively 
refine feature extraction, (see fia. 1). 
Real world images can be considered as 
comprising of three major high level spatial 
feature types. 
Feature extraction Pattern 
fig. 1 different schemes for the automatic 
interpretation process.