1022
this expert system simply reports the results and
does not try to automatically fix the misregistra
tions. For our forestry application, the forest map
is read into LDIAS and converted to grid format.
The selected elements are converted to a symbolic
location file. The remote sensing image is segmen
ted and the segments are classified. These are used
to produce a corresponding image symbolic location
file and statistics file. The simplified MICE
agenda from this point on is given in Table 7.
Table 7. Simplified mice agenda
1) Select map segment
2) Find all image segments in
map segment window (focus)
3) Compare class value
4) Compare segment sizes
5) Compare segment shapes
6) Compare segment locations
7) Output results
8) If not done, loop to (1).
Experiments with MSS imagery, federal maps, and
provincial maps indicate that shape and class values
are the most important heuristics for correctly
identifying areas of misregistration. The knowledge
of how to fix these areas has yet to be developed.
One might think that the map data should be tied to
the geocoded remote sensing image. For the resource
managers, this may be acceptable as a long-term
goal. However, they have a customer base familiar
with and using the existing resource maps. There
fore, rules will need to be developed for the major
types of misregistrations. A second problem with
the MICE expert system is that it is very slow. The
analysis of the 1:20,000 hydrography level of a
BCMOFL map and a MSS image requires approximately 3
hours on the AI VAXstation. Further experiments are
now in progress.
6 KEY ISSUES FOR THE FUTURE
Expert systems are required to integrate remote
sensing and geographic information systems for
resource management with automatic methods. The
complexities of the image analysis and geographic
information systems are such that one should use
cooperating, distributed expert systems. Much
research and experimentation remain to be done.
Many resource managers in Canada have been
reluctant in the past to use satellite remote sens
ing for resource management because of two issues.
Firstly, their need to be assured that there will be
a continuity of data. This need has been satisfied
by CCRS arranging to receive data from the LANDSAT
and SPOT satellites. Secondly, they desire that the
same resource information can be derived from each
data source. Their concerns include cost, accuracy,
and timeliness. Expert systems can aid in the pro
duction of the required information, but such sys
tems cannot compensate for the physical limitations
of the sensors.
For environmental monitoring, it is necessary to
access resource information stored usually in dif
ferent organizations. Artificial intelligence
methods are essential to achieve distributed, coop
erating systems which do not require large numbers
of highly skilled individuals.
In the future, the users will not need as much
computer expertise to achieve their resource manage
ment goals. Their systems, though, will be more
complex and the scientists and engineers which
develop them will be more specialized. We have
found that it takes approximately two years to make
a knowledge engineer.
Two other issues directly related to expert
systems technology are: who owns the knowledge? and
what if the expert systems are wrong? Experts are
going to be reluctant to provide their knowledge if
they think that their organization will eliminate
their job as a result. They will require commit
ments from their organization that this will not
happen. If expert systems succeed in reproducing
human performance in some of these limited domains,
then such systems will likely make errors at times.
The question of liability for such errors and their
subsequent results must also be addressed by organ
izations employing this technology.
ACKNOWLEDGEMENTS
The author is grateful for the cooperation of Mr.
Frank Hegyi of the British Columbia Ministry of
Forests and Lands. Messrs. John Zelek, Mike Robson,
and Syd Dubrofsky aided in the preparation of the
figures for this publication.
REFERENCES:
Gane, Chris and Trish Sarson 1977, Structured
Systems Analysis: Tools and Techniques, (Saint
Louis: McAuto).
Goldberg, M. and D.G. Goodenough 1976, Analysis of a
Spatial Filter for LANDSAT Imagery, SPSE Confer
ence Proceedings, pp. 276-282.
Goldberg, M. , D.G. Goodenough, M. Alvo, and G. Karam
1985, A Hierarchical Expert System for Updating
Forestry Maps with LANDSAT Data, Journal of IEEE
Geosciences and Remote Sensing, Vol. 73, No. 6,
June 1985, pp. 1054-1063
Goodenough, David G. 1979, The Image Analysis
System (CIAS) at the Canada Centre for Remote
Sensing, Cdn. J. of Rem. Sens., 5, pp. 3-17.
Goodenough, D.G., M. Goldberg, G.W. Plunkett, and
J. Zelek, 1987, An Expert System for Remote
Sensing, IEEE Journal of Geosciences and Remote
Sensing, (in press).
Goodenough, D.G., J.J. Palimaka, K. Dickinson, and
J. Murphy, 1984, Standard Format for the Transfer
of Geocoded Information in Spatial Polygon Files,
CCRS Research Report.
Goodenough, D.G., G.W. Plunkett, J.J. Palimaka,
1983, On the Transfer of Remote Sensing
Classifications into Polygon Geocoded Data Bases
in Canada, Proceedings of the Sixth International
Symposium on Automated Cartography, National
Capital Region of Canada, pp. 598-606.
Hegyi, F. and R.V. Quinet 1983, Integration of
Remote Sensing and Computer Assisted Mapping
Technology in Forestry, Canadian Journal of Remote
Sensing, 9, pp. 92-98.
Link, B.D. R. Kwok, D.L. Coutts, and N. Minielly
1985, TM Precision Geocoded Products from the
MOSAICS System, Proceedings ASPRS Fall Convention,
pp. 843-850.
Marble, Duane F. , Hugh W. Calkins, and Donna J.
Peuquet 1984, Basic Readings in Geographic
Information Systems, Spad Systems Ltd.
Plunkett, G.E., David G. Goodenough, and M. Goldberg
1986, Map Image Congruency Evaluation Knowledge-
Based System, Proc. Graphics/Vision Interface,
pp. 273-278.