165 
Obviously some attributes of features are more 
important than others and not all attributes are 
required to identify a feature. Generally speaking, 
spectral and shape attributes are more important 
identifiers at the very coarsest levels of the 
interpretation hierarchy, while size, shape and 
texture information become more important at 
more detailed levels. The system was therefore 
designed to give the user the option to supply those 
attributes which are deemed most important for the 
situation at hand. Alternatively, the system is 
capable of informing the user if more specific 
information is required, or if the system's 
interpretation is based on incomplete information. 
Once a feature has been identified, the result will be 
displayed to the user and will be added to the data 
base, so that it can be retrieved for future use during 
the interpretation. In addition to storing the 
identity of a feature, the category it belongs to, its 
place in the interpretation hierarchy, and its 
coordinates are also stored to facilitate integration 
into a GIS. 
4.4 Choice of Programming Languages 
Fortran was chosen for the image analysis part of 
the system because of its algorithmic nature and the 
availability of a wide range of statistical/analytical 
software routines. In addition, software previously 
developed by Paine (1987) was incorporated into the 
smoothing and segmentation process, and this also 
was written in Fortran. (The overall structure of 
the automated image interpretation system is 
shown in Figure 3). 
Prolog, a traditional artificial intelligence language, 
was chosen for the image understanding part of the 
system. Prolog is well suited to problems that 
involve objects and the relationships between 
objects, because of its declarative rather than 
procedural approach. This enhances the 
modularity of the program, and makes it easier to 
add new information to the knowledge base, if the 
application area of the system is to be expanded or if 
new information about the field becomes available 
(Genesereth and Ginsberg, 1985). These properties 
make Prolog a language suitable for prototyping. 
For automated image interpretation, modularity, 
and especially the backtracking mechanism, which 
is a built-in feature of Prolog, is extremely 
important. In a truly modular system, relaxing 
some of the assumptions outlined in Section 4.3 
above merely involves the addition of new 
modules to the knowledge base. The structure of 
the original rules can, for the most part, remain 
unchanged. 
Although the construction of a natural language 
user interface was not the focus of this research, it 
should be noted that Prolog is suitable for natural 
language parsing. An expert-like facade, which 
explains system reasoning, can therefore be added 
to the system at a later point in time without the 
problem of interfacing different programming 
languages. The extraction of spatial attributes, such 
as area and shape, however, is a task for which 
Prolog is not well suited, because spatial attribute 
extraction requires numerical computations which 
are procedural in nature. Area calculations in a 
raster image consist of counting pixels, while the 
degree of compactness is calculated as a shape 
measure (Perkins et al., 1985). C was selected as the 
language to extract spatial attributes, because it can 
relatively easily be interfaced with Prolog. More 
complex attributes, such as pattern information, are 
supplied directly by the user if they are required for 
an identification. 
IMAGE ANALYSIS ► FORTRAN 
IMAGE UNDERSTANDING ► PROLOG 
(SOME C) 
Figure 3: Structure of Automated Image 
Interpretation System 
5 RESULTS AND DISCUSSION 
The image understanding system as it currently 
stands contains about thirty-five rules aimed at 
identifying fields (fallow and planted with various 
crops), man-made structures, road segments, water