bodies, and rangeland. It was found that the most
easily identifiable features were those that are by
definition regular in shape and/or spectrally very
homogeneous - in other words fields, water bodies,
and man-made structures. Fallow fields, which are
both homogeneous, spectrally speaking, and regular
in shape, were correctly identified in nearly all
cases, and were successfully distinguished from
other non-vegetated surfaces. Planted fields, which
often show subtle variations in their spectral
responses, resulting in several distinct polygons
making up one field, could often be successfully
identified by using spatial relations, such as
adjacency. This made it possible to determine
whether contiguous polygons of similar spectral
response could be updated into one homogeneous
feature, matching conditions in the knowledge base
required to label it as a field.
However, the system cannot at the moment
successfully extract those features which are neither
regular in shape nor exhibit homogeneity in their
spectral response. An example of such a feature is
rangeland, or vegetated areas in general, that are
not cultivated. What the system can do, is
tentatively label individual polygons in the image
database as rangeland, but it is not successful in
using adjacency and other spatial relations to extract
entire units. The implication of this is that, for
more complex features, e.g. units of terrain, the
simple spatial and spectral measures currently used
in the system are not sufficient for identification
and will have to be supplemented by ancillary
information, such as topography and soils. It also
implies that when designing a knowledge base for
more complex applications, such as terrain analysis,
ancillary information, e.g. in the form of a digital
elevation model, will be necessary for the system to
perform satisfactorily.
Since the structure of the system as it currently
stands is quite modular, experimenting with
different types of ancillary information, and with
different combinations of spatial and spectral
attributes, to improve the scope and the robustness
of the current system, should be a relatively straight
forward matter. Experimentation with additional
attributes would ideally include different
combinations of spatial and spectral attributes,
more complex shape measures, and more complex
descriptors of spatial and spectral knowledge in the
knowledge base itself.
6 CONCLUSIONS
Using a rule based system for automated feature
extraction is attractive, because of the ability of these
systems to structure information from diverse
sources flexibly and in a modular fashion. The
prototype system introduced in this research takes
advantage of these properties by using a variety of
spatial and spectral information extracted from
images to identify rural landuse features in Landsat
TM imagery.
The system exhibits success in extracting simple
landuse features by using only image-derived
spatial and spectral attributes. However for more
complex applications, such as terrain evaluation,
the use of ancillary information, as well as more
complex spatial attributes than those currently
employed, will be a necessity for successful
operation. In addition, a fully operational system
should provide the capability to fully integrate both
algorithmic and expert knowledge under one
control structure to facilitate any image
manipulation functions or numerically intensive
computations that might be required for the
extraction of more complex spatial features during
the image interpretation process.
While the system described here is able to identify
features from digital imagery, the interpreted
images are still a long way from being directly
integratable into a GIS. The main reason for this is
that no uncertainty model has been developed to
indicate the degree of confidence associated with an
interpretation, especially where a choice exists
between several alternatives. This type of
ambiguity could be dealt with by incorporating a
fuzzy measure within the system (Klir and Folger,
1988).
If the automated image interpretation system is to
become a true expert system, then the development
of an explanation facility and a natural language
user interface is a necessity. Such an interface could
deal with ambiguity in attribute values (Karimi and
Lodwick, 1987). The consistency of interpretations
could therefore be improved for those attribute
values which are user supplied. Finally, the
development of an explanation facility, especially
where it supplies answers about unexpected results,
could serve as a learning tool for an inexperienced
image analyst.
7 REFERENCES
Avery, T.E. and Berlin, G.L., 1985. Interpretation of
aerial photographs. Burgess Publishing
Company, Minnesota, 554 pp.
Bratko, I., 1987. Prolog programming for artificial
intelligence. Addison Wesley Publishing
Company, Reading, Massachusetts, 423 pp.
Burrough, P.A., 1986. Principles of geographical
information systems for land resources
assessment. Clarendon Press, Oxford, 193 pp.
EOSAT, 1988. Landsat data user notes. Earth
Observation Satellite Company, Lanham,
Maryland, Vol. 3, No. 1.
Estes, J.E., 1977. A perspective on the state of the art
of photographic interpretation. Proceedings of
Eleventh Symposium on Remote Sensing of
Environment, Ann Arbor, Michigan, October
1977, pp. 161-176.
Estes, J.E., Sailer, C. and Tinney, L.R., 1986.
Applications of artificial intelligence techniques
to remote sensing. The Professional
Geographer, Vol. 38, No. 2, pp. 133-141.
Genesereth, M.R. and Ginsberg, M.L., 1985. Logic
programming. Communications of the ACM,
Vol. 28, No. 9, pp. 933-941.
Goodenough, D.G., Goldberg, M., Plunkett, G. and
Zelek, J., 1987. An expert system for remote
166