164
Modularity allows easy expansion of the system,
because the knowledge base of a modular system
can be changed without altering the structure of the
main program.
In recent years, personal computers have become
powerful enough for image processing and GIS data
base management. EOSAT (1988), for example,
makes TM floppy disk images covering 512 x 512
pixels commercially available to interested users.
Developing the system for use with
microcomputers, or allowing portability between
mainframe and microcomputers, is therefore
important to avoid redundancy in programming.
User and system design requirements are not
necessarily independent from each other and
meeting some requirements is more important
than others. For example, efficiency in terms of
time and cost was also selected as an important
requirement but has, in the initial design phase of
the system, a lower priority than accuracy of feature
identification.
4 IMPLEMENTATION
4.1 Study Area
A quadrant of a TM image centred east of Calgary
Alberta, Canada and acquired in late July, 1985
was chosen as the study area. This region,
predominantly agricultural landuse, is typical of a
North American prairie landscape. It is suitable as
a test area, because it consists of a fairly uniform
low variance data set. However, it shows a variety
of features commonly placed on 1:50,000 maps, such
as roads, urban areas, railways, fields, and
hydrographic features. The performance of the
system in identifying these features can therefore be
tested without interference from noise or highly
variable terrain.
4.2 Image Analysis
Before this image can be input to the interpretation
system, it has to be smoothed to remove excess
variance, and segmented into its characteristic
regions. Working with regions rather than
individual pixels during the interpretation is easier,
because it corresponds to the units the human eye
recognizes. In addition, the data volume the
system has to work with can be reduced (Perkins et
al., 1985).
All image bands except for the thermal band of the
256 x 256 image window of the study area were
enhanced using principal components analysis.
The three dominant principal components
accounted for 97.8 percent of the variation within
the scene, the first component (70.85 percent)
representing scene brightness, the second
component (17.28 percent) representing biomass,
and the third component (9.67 percent)
representing 'greenness' in the image. These three
components were used as input to a migrating
means clustering algorithm for image
segmentation. Twelve distinct clusters were
extracted from the imagery. Of these, one cluster
was found to represent water, four clusters various
non-vegetated surfaces and the remaining seven
vegetated surfaces.
A comparison with a manual segmentation of the
study area indicated that the machine segmentation
was of acceptable quality. Discrepancies occurred
because of the machine's better ability to
distinguish subtle spectral differences, e.g. due to
different moisture conditions within a feature. In
these situations, rather than representing such a
feature as a single logical unit within the image
database, it would be broken up into distinct
polygons with slightly different spectral properties.
One objective during testing of the image
understanding system therefore was to determine
if, and how well, the system could resolve such
conflicts.
4.3 Image Understanding
This part of the system was designed as a prototype
which should eventually be expandable for a wide
variety of image interpretation tasks. In order to
keep the problem domain (and therefore the
knowledge base) manageable in the initial design
phase, some simplifying assumptions were made:
(a) The system is aimed at a North American
prairie landscape.
(b) Only features recognizable on TM imagery
(green, red, infrared bands) were considered.
No height information is included.
(c) Some attributes of features (e.g. pattern) are
user supplied and not automatically extracted.
(d) Rules are aimed at imagery acquired during
the growing season.
The types of queries the system must be capable of
answering (see Section 3.2) require a flexible data
structure, which can incorporate values for all
attributes of interest (see Table 1). The system
should also be expandable for more complex
situations. This type of information can best be
encoded in a production rule environment. Such
rules are conditional statements which are easily
modifiable, because they are relatively independent
from each other (Bratko, 1987). The rules were
determined by visually interpreting the TM image
of the study area and by studying what kind of
attributes are most commonly associated with
individual features. Some textual examples of
rules encoded in the knowledge base are:
Rule: If the feature is circular and shows high
reflectance in the infrared region of the
E-M spectrum, then the feature is a pivot
irrigation field.
Rule: If the feature is linear, shows high
reflection in the green and red and forms
a dendritic pattern, then the feature is a
river.