10 Polar SAR Data for Operational Sea Ice Mapping
227
gray-level co-occurrence matrix (Haralick et al. 1973), which describes the spatial
arrangement of gray tones in an image or subimage. Some conclude that first order statistics are sufficient (e.g. Nystuen and Garcia 1992), while others conclude that second
order statistics offer improved discrimination (e.g. Shokr 1991). The lack of consensus
arises because of differences in speckle treatment, radar parameters, gray-level
co-occurrence matrix derivation parameters, and because most studies draw conclusions based on the analysis of a single image.
These studies evaluated texture based on samples of pure types. Practically, as with
dynamic thresholding, there is a "windowing problem"with texture use in an algorithm;
texture must be measured within small areas of a larger image and results directly
depend on how these areas are selected. Smith et al. (1995) have a workable approach
to this problem; first they segment an image without assigning types by filtering the
image histogram to sharpen it, then they assign pixels to the same area based on closeness to histogram peaks. Once an image is segmented, texture samples can be extracted without danger of drawing from two distinct areas. Texture then becomes another
parameter, along with mean backscatter, with which to classify areas. However, Smith
et al. (1995) found that second-order parameters added no discrimination capability
beyond first order parameters.
10.7.3
Expert Systems
Increasing the dimensionality of data by using multipolarization SAR or by data blending will not be operationally practical in the near term. Algorithms that use texture are
slow, and they are not conclusively better than other algorithms. Using a dynamic
threshold to segment imagery is expected to reduce, but not eliminate, the problem of
backscatter class overlap. Reproducing manual SAR image interpretation techniques is
an alternative solution for achieving better automated classification results.
The first expert system for SAR sea ice imagery was built by Intera Technologies Ltd.
for Canadian ice analysts (McAvoy and Krakowski 1989). It was used as a training tool
rather than for operational analysis. A user supplied answers to questions regarding the
appearance of displayed ice, and these answers led to an ice classification based on how
human experts would classify the ice. In 1991, CCRS sponsored a pilot project by Norland Science and Engineering Ltd. to construct a knowledge base for the Canadian Arctic. This rule base was not intended to be comprehensive, and not all of the rules could
necessarily be implemented. The project anticipated the limitations that the initial
image segmentation would place on the expert system's skill. It also stressed the importance of including an ice climatology in the knowledge base. In 1993, Norland Science
introduced a package called IceXpert that also used knowledge-based techniques to
train analysts.
Haverkamp et al. (1995) created the first "stand-alone" expert system for ice analysis.
This system started with an initial classification map from the ASF algorithm. The subsequent classification was refined by applying classification rules to ice features. Features were defined as contiguous pixels identified by the initial classification as the same
ice type. Most rules concerned derived geometrical feature characteristics such as "circularity;' or position relationships such as "enclosed by" or "adjacent to:' A feature
might be reclassified or merged with another feature. For instance, a feature classified
227
gray-level co-occurrence matrix (Haralick et al. 1973), which describes the spatial
arrangement of gray tones in an image or subimage. Some conclude that first order statistics are sufficient (e.g. Nystuen and Garcia 1992), while others conclude that second
order statistics offer improved discrimination (e.g. Shokr 1991). The lack of consensus
arises because of differences in speckle treatment, radar parameters, gray-level
co-occurrence matrix derivation parameters, and because most studies draw conclusions based on the analysis of a single image.
These studies evaluated texture based on samples of pure types. Practically, as with
dynamic thresholding, there is a "windowing problem"with texture use in an algorithm;
texture must be measured within small areas of a larger image and results directly
depend on how these areas are selected. Smith et al. (1995) have a workable approach
to this problem; first they segment an image without assigning types by filtering the
image histogram to sharpen it, then they assign pixels to the same area based on closeness to histogram peaks. Once an image is segmented, texture samples can be extracted without danger of drawing from two distinct areas. Texture then becomes another
parameter, along with mean backscatter, with which to classify areas. However, Smith
et al. (1995) found that second-order parameters added no discrimination capability
beyond first order parameters.
10.7.3
Expert Systems
Increasing the dimensionality of data by using multipolarization SAR or by data blending will not be operationally practical in the near term. Algorithms that use texture are
slow, and they are not conclusively better than other algorithms. Using a dynamic
threshold to segment imagery is expected to reduce, but not eliminate, the problem of
backscatter class overlap. Reproducing manual SAR image interpretation techniques is
an alternative solution for achieving better automated classification results.
The first expert system for SAR sea ice imagery was built by Intera Technologies Ltd.
for Canadian ice analysts (McAvoy and Krakowski 1989). It was used as a training tool
rather than for operational analysis. A user supplied answers to questions regarding the
appearance of displayed ice, and these answers led to an ice classification based on how
human experts would classify the ice. In 1991, CCRS sponsored a pilot project by Norland Science and Engineering Ltd. to construct a knowledge base for the Canadian Arctic. This rule base was not intended to be comprehensive, and not all of the rules could
necessarily be implemented. The project anticipated the limitations that the initial
image segmentation would place on the expert system's skill. It also stressed the importance of including an ice climatology in the knowledge base. In 1993, Norland Science
introduced a package called IceXpert that also used knowledge-based techniques to
train analysts.
Haverkamp et al. (1995) created the first "stand-alone" expert system for ice analysis.
This system started with an initial classification map from the ASF algorithm. The subsequent classification was refined by applying classification rules to ice features. Features were defined as contiguous pixels identified by the initial classification as the same
ice type. Most rules concerned derived geometrical feature characteristics such as "circularity;' or position relationships such as "enclosed by" or "adjacent to:' A feature
might be reclassified or merged with another feature. For instance, a feature classified