levi- 
) 22 
| of 
(fig. 
be 
CCU- 
fica- 
noto 
dif- 
con- 
n at 
idth 
om- 
arror 
with 
ı 20 
y is 
the 
ints 
nust 
9no- 
liffi- 
s of 
are 
reo- 
nigh 
and 
reo- 
SOr- 
d in 
| of 
ould 
lyti- 
st a 
in is 
ints 
lies 
‘hen 
Jata 
hen 
osi- 
| for 
has 
ing, 
Fa- 
iulp- 
esia 
peri- 
ereo 
ings 
AUTOMATIC FEATURE EXTRACTION FOR MAP REVISION 
Dr. Hiroshi MURAKAMI 
Deputy Director 
International Affairs Division 
Ministry of Construction, JAPAN 
and 
Dr. Roy WELCH 
Director 
Center for Remote Sensing and Mapping Science 
University of Georgia 
Commission IV 
ABSTRACT: 
Automatic feature extraction techniques were developed for use with digital images and map 
data to assess the feasibility of employing expert systems for map revision. The map and 
image data were placed in register to create a cartographic database suitable for use with 
a prototype expert system optimized for the extraction of building features. The expert 
system approach permitted control of image processing routines applied to the cartographic 
database for feature extraction. The accuracy of feature extraction increased as the image 
pixel resolution was improved. 
KEY WORDS: Cartographic Database, Change Detection, Expert System, Feature Extraction, 
Image Processing, Map Revision. 
1. INTRODUCTION 
Most developed countries have completed 
national mapping programs that provide 
topographic map coverage at scales of 
1:25,000 or smaller, and map revision is 
now the main task. Urban expansion, 
however, causes maps to become out-dated 
rapidly, while funds allocated to mapping 
have been reduced. Consequently, there is a 
need for more efficient and cost effective 
methods for labor-intensive map revision 
tasks, particularly for change detection, 
in which differences between newly acquired 
images and old maps are determined. 
Feature extraction studies have mainly 
focused on objects such as roads and 
buildings included in a digital image 
(Bajcsy and Tavakoli, 1976; Nagao and 
Matsuyama, 1980; Nevatia and Babu, 1980; 
Fischler et al., 1981; McKeown .et al., 
1985; Huertas and Nevatia, 1988; Wang and 
Newkirk, 1988). 
The objective of this study was to develop 
a method of detecting changes of buildings 
in SPOT images. Since change detection 
method needs the photo interpreters’ 
knowledge to identify each detected change, 
expert system approach was employed to deal 
with human expertise (Murakami, 1990). 
2. ISSUES IN FEATURE EXTRACTION 
This study focused on the following three 
important points out of the problems 
encountered in feature extraction (Nagao 
and Matsuyama, 1980; Hanson and Riseman, 
1988; Matsuyama and Hwang, 1990). 
2.1 Initial Parameter Value Selection in 
Image Segmentation 
Computers can not reliably extract specific 
objects directly from gray-scale images. 
Consequently, the original gray-scale image 
must be transformed [irst to an image in 
which each ground feature is independently 
569 
labeled. In image segmentation, initial 
parameters (e.g., threshold values) must be 
employed to distinguish ground features 
from their background. Appropriate 
threshold values, however, may differ from 
feature to feature - even in a single 
image. Consequently, developing a method to 
select appropriate threshold values in an a 
priori manner will be required. 
2.2 Extraction of Descriptor Values 
Interpretation of individual labeled 
regions requires descriptors of the 
characteristics of each ground feature. 
Most descriptors are related to the seven 
elements of photo interpretation, i.e., 
tone, shadow, pattern, size, texture, 
shape, and association (Paine, 1981; 
Lillesand and Kiefer, 1987). In theory, 
extraction and proper processing of all the 
information concerning these elements would 
provide the same understanding of the input 
image as human interpreters. Hence, 
selection of the most important elements 
for a particular kind of features, i.e., 
building, will be necessary. 
Consequently, there must be a procedure for 
establishing values for descriptors related 
to each of the interpretation elements. Of 
course, it must be understood that "human 
perception' does not necessarily correspond 
with "machine perception". 
2.3 Uncertainty Management and Inference 
Method 
Some uncertainty is associated with 
descriptor values derived from segmented 
regions. Thus, knowledge or guiding rules 
must be applied to establish the identity 
of each object. Unfortunately, these rules 
may also contain some uncertainty. For 
example, "A bright, elongated object (20 m 
x 40 m) in a satellite image is a 
building," may be true in most instances. 
However, a road or agricultural field may 
exhibit similar characteristics. Hence,