163 
Table 1: Spatial and Spectral Attributes of Features 
(after Avery and Berlin, 1985) 
Attribute 
Comments 
Size/ area 
Both relative and absolute size are 
important. Relative size is strongly 
scale dependent. Absolute size can 
be measured by counting pixels. 
Shape 
Sometimes sufficient to identify a 
feature, but difficult to calculate. A 
measure of compactness could be 
used as a shape measure. 
Colour 
Typical of certain objects but 
depends strongly on season, angle 
of illumination etc. Straight-forward 
to extract from the image. 
Pattern 
Spatial arrangements of objects are 
especially useful in identifying 
man-made features. 
Texture 
Refers to the coarseness or 
smoothness of features, but is 
strongly scale dependent. 
Site/ 
association 
Important for contextual 
information, but can only be 
defined in conjunction with 
already identified features. 
Manual image interpretation usually proceeds 
systematically from feature category to category, as 
outlined above, and within these categories in a 
hierarchical fashion from more general 
descriptions to more specific details (Estes, 1977). 
Part of such a hierarchy is shown in Figure 2. In 
order to allow for a straightforward integration of 
interpreted features into a GIS, the automated 
system should solve its tasks in an equally 
systematic way. Maintaining a hierarchy is 
especially important, because spatial attributes, such 
as relative size, texture and pattern are strongly 
scale dependent. 
Urban 
Residential Commercial Industrial Transportation 
main road subsidiary road 
Figure 2: Image Interpretation Hierarchy 
Comprising Four Levels 
3.2 User Requirements 
In order to facilitate the interpretation of remote 
sensing data, an expert interpretation system 
should be able to extract accurately and consistently 
resource features belonging to the six categories (a) 
to (f) outlined in Section 3.1. When prompted by 
the user, it should be able to interpret an image 
according to: 
(a) Feature (e.g. identify all roads; is feature X a 
road), 
(b) Feature attribute (e.g. show all small features), 
(c) Feature location (e.g. what is the identity of the 
feature at location x, y), 
(d) Relationships between features (e.g. what 
features are next to feature X), 
(e) Combinations of these queries (e.g. identify all 
small features next to feature X). 
Quality control, or an assessment of the accuracy of 
information, is an important aspect whenever data 
are integrated into a GIS data base. It is therefore 
necessary to have a confidence value associated 
with an identification, which gives an indication of 
the reliability of an interpretation. In addition, data 
validation procedures should be available so that 
invalid data can be rejected by the system, before the 
interpretation process begins. 
One reason expert systems have become attractive 
for the solution of complex tasks is their ability to 
interact with the user by using a human-like 
interface to explain system reasoning. A successful 
image interpretation expert system should 
therefore possess at least an explanation facility and, 
if possible, a natural language user interface. 
Interactive screenwork is one of the most popular 
and user friendly methods of image manipulation 
and analysis. The user of an automated image 
interpretation system should be able to select 
regions by placing a cursor on the image. Similarly, 
results should be displayed on the screen. To 
facilitate the integration of these results into a GIS, 
the identity of the features, their coordinates, and 
any other relevant information should be stored in 
a compatible data structure, e.g. as map overlays, 
which correspond to the data categories outlined in 
Section 3.1 above. 
3.3 System Requirements 
Modularity to allow easy expansion of the system, 
and portability to permit transfer between 
mainframe and microcomputers were selected to be 
the most important system design requirements. 
Image interpretation is a complex and many faceted 
topic, and it was considered impractical to develop a 
multi-purpose interpretation system in one design 
phase. If a prototype system with a limited problem 
domain is to be of any future use, then it has to be 
expandable to a wide variety of applications.