161 
A RULE-BASED SYSTEM FOR THE EXTRACTION OF CARTOGRAPHIC 
FEATURES FROM LANDSAT TM IMAGERY 
M. Stadelmann 
G.D. Lodwick 
Department of Surveying Engineering 
The University of Calgary 
2500 University Drive N.W. 
Calgary, Alberta 
Canada T2N 1N4 
Prepared for ISPRS COMMISSION VII Mid-term Symposium, Victoria, B.C., Canada, September 1990. 
ABSTRACT 
This paper describes work that has led to the development of a basic rule based system for the automated 
recognition of cartographic features in digital satellite imagery. The system uses frames to represent 
knowledge and is written in Arity PROLOG. Input to the rule based system consists of a segmented image 
modelled as a polygonal database. Image segmentation is carried out on the multispectral data using a 
migrating-means clustering algorithm. Simple landuse features, such as fields and water bodies, are extracted 
using spatial information, such as object areas, and simple shape measures, such as compactness ratios, 
combined with spectral knowledge. For more complex features, contextual information, such as adjacency, 
and more complex shape measures, such as elongation, are extracted from the image database. Results have 
been compared with conventional image interpretation and show acceptable accuracy, provided image 
segmentation errors are moderate. These conditions are met in images depicting rural areas as they contain 
spectrally homogeneous features. In more variable environments, such as urban areas, more advanced 
measures of coping with segmentation problems and the use of ancillary information are expected to generally 
enhance the success of automated image interpretation. 
KEY WORDS: Automated, digital, interpretation, mapping, cartography, Landsat, remote-sensing, GIS. 
1 INTRODUCTION 
Earth observation satellites, such as the Landsat 
series or more recently SPOT, have the potential to 
facilitate map production and revision, as well as 
aid in the management of natural resources. They 
are particularly useful because their high 
resolutions and repetitive coverage permit the 
detection and monitoring of temporal changes on a 
regional and local scale. However, the increasing 
resolution poses the problem of increasingly large 
data volumes, which must be analyzed and 
managed (Maslanik and Smith, 1984). 
Electro-optical sensors capture data in digital 
format suitable for computer processing. However, 
automated feature extraction has mainly been 
restricted to statistical classification methods, which 
distinguish between surface cover classes on the 
basis of spectral signatures. The interpretation of 
the images, i.e. the assignment of meaningful 
names to cover classes, has remained a manual 
task, because image interpretation, in addition to 
spectral knowledge, requires a considerable amount 
of contextual information. Extraction of this type 
of information requires human expertise and 
decision making which cannot readily be 
accomplished using traditional algorithmic 
methods. 
In the conventional map production process, 
image interpreters typically analyze a satellite 
image and transfer any objects of interest manually 
onto a map. A GIS data base is then usually 
produced by digitizing these map products, which 
have already undergone abstraction and 
generalization compared to the original remote 
sensing data (McKeown, 1987). GIS products 
therefore usually result from one or more thematic 
map overlays. 
Image interpretation is a slow process which 
requires highly skilled personnel. Digitization and 
integration into the GIS data base is tedious and 
error prone. For example, if an image polygon is to 
be placed in the GIS, data base corruption will occur 
if there is no perfect juxtaposition of the image 
polygon and neighbouring map polygons 
(Goodenough et al., 1987). In addition, the accuracy 
of a finished map product decreases considerably if 
two or more digitized map overlays are utilized in 
its production (Newcomer and Szajgin, 1984). The 
time involved in image interpretation and quality 
control of map products in GIS has created a 
bottleneck in spatial data handling that has to be 
eliminated if resources managers and planners 
wish to make efficient use of the technology 
available to them.