The objective of this research was to design and 
implement a simple rule based image 
interpretation system capable of identifying features 
in Landsat TM imagery by using a variety of spatial 
and spectral attributes, such as the ones commonly 
used by human interpreters (e.g. size and shape). 
Implementing a suitable knowledge structure and 
knowledge representation scheme was of primary 
interest in this investigation. In order to determine 
the success of the implemented prototype and 
detect any problems and design flaws in the system, 
it was tested on a relatively narrow problem 
domain - on the identification of rural landuse 
features in principal components enhanced TM 
images (Stadelmann, 1990). 
2 EXPERT SYSTEMS FOR IMAGE 
INTERPRETATION 
Expert systems (knowledge based systems) are 
computer programs that use symbolic knowledge to 
simulate the behaviour of human experts (Stock, 
1987). Expert system design is a subdiscipline of 
Artificial Intelligence (Estes et al., 1986). Figure 1 
illustrates the basic structure of such a system. 
Figure 1: Structure of an Expert System 
(after Burrough, 1986) 
The knowledge base of the system contains facts 
and rules in a specific problem domain, here 
feature extraction and interpretation from digital 
satellite imagery, whereas the inference engine is 
the reasoning mechanism that activates specific 
rules in the knowledge base. Knowledge within the 
system is usually acquired from human experts in 
the discipline under investigation. The 
explanation facility and the user interface are those 
units which allow for smooth expert-like 
interaction between the user and the system by 
explaining 'system reasoning' and offering on-line 
help facilities. 
Expert systems differ from conventional computer 
programs in that they are heuristic rather than 
algorithmic in nature. More importantly, the 
system's knowledge is modular and autonomous 
and therefore not embedded in the control 
structure of the program (Estes et al., 1986). This 
means that the knowledge base of an expert system 
can be changed without affecting the reasoning 
mechanism of the system. These properties make 
expert systems a promising approach for image 
interpretation. 
Apart from spectral pattern recognition, image 
interpretation depends on spatial information, such 
as the size and shape of objects and how they are 
related to each other. Using spatial, spectral, and 
contextual information interchangeably depends on 
a good deal of human reasoning and experience. 
Also, human reasoning involves the ability to find 
solutions based on limited or incomplete 
information (Lodwick and Colijn, 1987). 
Traditional image analysis methods (spectral 
pattern classifiers) are not suitable for 
interpretation, because procedural methods have 
problems where the number of decision paths is 
large or not known beforehand, as is the case with 
spatial relationships between objects. Apart from 
being inflexible when it comes to changing or 
adding knowledge to a program, procedural 
methods also do not accommodate large amounts 
of knowledge, and time and storage requirements 
for an exhaustive search for an optimal solution are 
often prohibitive (Lodwick and Colijn, 1987). 
Expert systems are declarative rather than 
procedural in nature, are capable of using 
heuristics, and therefore do not depend on 
exhaustive search techniques. In addition, their 
knowledge base can easily be changed or expanded 
if the need arises because of the system's 
modularity. However, image interpretation expert 
systems will be most successful with narrow 
problem domains, manageable amounts of image 
data and where skilled interpreters are available for 
the construction of the knowledge base. If these 
conditions can be met, expert systems have the 
potential to produce consistent and reasonable 
image interpretations in an automated 
environment. 
3 BASIC DESIGN CONSIDERATIONS 
3.1 Identifying Features by Attribute Values 
Paine (1987) indicates that GIS usually include the 
following data: 
(a) Terrain 
(b) Landuse 
(c) Hydrography 
(d) Vegetation 
(e) Soils 
(f) Geology 
(g) Cadastral information 
(h) Tables 
(i) Reports 
Information about the first six categories in this list 
is usually assembled by interpreting aerial 
photographs or digital satellite imagery and is 
represented as digital map overlays. Image 
interpretation is based on the assumption that 
objects have a specific set of spatial and spectral 
properties. Expert image interpreters therefore 
quite systematically employ the attributes set out in 
Table 1 in visual image interpretation.