228
C. BERTOIA, J. FALKINGHAM, F. FETTERER
as multiyear ice could be reclassified as open water if it was sufficiently elongated. The
system has two formalisms: a fact base, consisting of assertions about every feature;
and a knowledge base, consisting of inference rules that may add, change or delete facts.
Rules "fire" when a match is found in the rule base for the "if" portion. This in turn may
cause other rules to fire. For example, a rule that enforces the knowledge that in winter and away from the MIZ, all openings in the pack are linear, and first-year ice may
at times have very low backscatter, would be:
i f
and
then
(class feature 1 OpenWater)
(shape feature 1 not-elongated)
(retract (class feature 1 OpenWater))
(assert (class feature 1 FirstYear))
The system stops when a steady state is achieved, that is, when no more rules fire based
on existing facts. The authors noted a 10% change in the classification for two subscenes
of the ERS-l image with which the algorithm was developed, and visually evaluated it
to be a 10% improvement.
Emulating the wayan expert classifies an image is appealing because manual analysis of SAR is the standard by which other methods are (or should be) measured.Actual implementation of an expert system, though, requires unappealing complexity. Subjective guidelines (for instance, '''round' floes are probably MY") must be encoded as
rules, although measures of the degree of a characteristic such as roundness could
potentially be assigned and used in decision-making with fuzzy set theory. The expert
system of Haverkamp et al. (1995) can merge features, but cannot divide features that
were initially wrongly assigned as a single feature. The system is therefore dependent
upon a relatively good initial classification. To improve the initial classification, the
expert system is united with a dynamic threshold algorithm and a feature extraction
algorithm. The dynamic threshold algorithm has the advantages cited above, while the
feature extraction algorithm uses the image processing technique of dilation and erosion to improve the definition of floe boundaries when floes are touching. The system
was run on about 90 ERS-l images from both winter and summer. Although results were
not quantified, the authors noted many instances in which features were reclassified in
a manner that concurred with manual analysis procedures.
Expert systems have been used quite successfully for the analysis of geophysical data
[the Prospector system for geological data is an early and model example (Duda et al.
1979)]. One criterion for using an expert system is that the problem domain be wellstructured; i.e., a model for classification should already exist (Luger and Stubblefield
1993). There is no such rigid structure for sea ice classification. However, the difficulty
of correctly classifying single channel radar data procedurally demands continued
efforts to code the manual analysis process.
10.7.4
Summary
RADARSAT makes it possible to accomplish both Arctic-wide and regional analysis
with the same high level of detail. A successful operational algorithm, however, will
quickly process imagery from the central Arctic using averaged data for a less accurate
C. BERTOIA, J. FALKINGHAM, F. FETTERER
as multiyear ice could be reclassified as open water if it was sufficiently elongated. The
system has two formalisms: a fact base, consisting of assertions about every feature;
and a knowledge base, consisting of inference rules that may add, change or delete facts.
Rules "fire" when a match is found in the rule base for the "if" portion. This in turn may
cause other rules to fire. For example, a rule that enforces the knowledge that in winter and away from the MIZ, all openings in the pack are linear, and first-year ice may
at times have very low backscatter, would be:
i f
and
then
(class feature 1 OpenWater)
(shape feature 1 not-elongated)
(retract (class feature 1 OpenWater))
(assert (class feature 1 FirstYear))
The system stops when a steady state is achieved, that is, when no more rules fire based
on existing facts. The authors noted a 10% change in the classification for two subscenes
of the ERS-l image with which the algorithm was developed, and visually evaluated it
to be a 10% improvement.
Emulating the wayan expert classifies an image is appealing because manual analysis of SAR is the standard by which other methods are (or should be) measured.Actual implementation of an expert system, though, requires unappealing complexity. Subjective guidelines (for instance, '''round' floes are probably MY") must be encoded as
rules, although measures of the degree of a characteristic such as roundness could
potentially be assigned and used in decision-making with fuzzy set theory. The expert
system of Haverkamp et al. (1995) can merge features, but cannot divide features that
were initially wrongly assigned as a single feature. The system is therefore dependent
upon a relatively good initial classification. To improve the initial classification, the
expert system is united with a dynamic threshold algorithm and a feature extraction
algorithm. The dynamic threshold algorithm has the advantages cited above, while the
feature extraction algorithm uses the image processing technique of dilation and erosion to improve the definition of floe boundaries when floes are touching. The system
was run on about 90 ERS-l images from both winter and summer. Although results were
not quantified, the authors noted many instances in which features were reclassified in
a manner that concurred with manual analysis procedures.
Expert systems have been used quite successfully for the analysis of geophysical data
[the Prospector system for geological data is an early and model example (Duda et al.
1979)]. One criterion for using an expert system is that the problem domain be wellstructured; i.e., a model for classification should already exist (Luger and Stubblefield
1993). There is no such rigid structure for sea ice classification. However, the difficulty
of correctly classifying single channel radar data procedurally demands continued
efforts to code the manual analysis process.
10.7.4
Summary
RADARSAT makes it possible to accomplish both Arctic-wide and regional analysis
with the same high level of detail. A successful operational algorithm, however, will
quickly process imagery from the central Arctic using averaged data for a less accurate