262
A.T. SEPHTON AND K.C. PARTINGTON
classification in the (general) case where backscatter coefficient alone is an ambiguous
attribute for classification. Segmentation is described more fully in Sect. 12.2.2.
12.2.1.2
Inherent Variability in Backscatter
The second issue which needs to be addressed in any classification of ice concerns the
inherent variability in backscatter within ice types, between ice types, and within areas
of open water. A summary of early ERS-l measurements of backscatter coefficient provided by Partington and Hanna (1994) indicated that ranges of backscatter within a
single ice type could be as high as 12 dB (e.g., rough first-year ice), and that the range
of backscatter for multiyear ice is covered by that of rough first -year ice. Bright features
include ridges, old ice (which has both appreciable radar penetration depth and generally rough surface characteristics), the edges of ice floes (which can produce a multiple reflection geometry with the water surface), frost flowers, and wind-roughened
open water. Dark features include smooth water, new ice, frazil ice (millimeter-scale fragments of ice in open water), and areas of new ice frozen into sheets. However, there are
also variations in backscatter contrast between different ice types which depend on surface temperature (particularly on whether the ice is melting) and on large-scale geographic variations in the mode and rate of formation of sea ice (e.g. salinity variations,
typical compressive pressures, etc.). This problem, together with that of speckle, makes
automated classification difficult except under well-understood, restricted conditions.
It also makes a pixel-based classification system difficult to justify, despite significant
savings in computational time.
The alternative approach, alluded to earlier, is to make use of attributes other than
backscatter coefficient in any classification. This suggests a requirement for segmentation and for a classification scheme which makes use of region attributes including
backscatter coefficient and measures of image texture. In practice this is what happens
in a manual classification, with the human interpreter making use of texture and context as much as mean backscatter coefficient in any classification. In a hybrid classification, such as that described below, segmentation and classification are partially automated (with texture being provided as a set of attributes), but context can be included
in a classification through the manual classification of selected 'training' regions.
Region-based classification is described in Sect. 12.2.3.
12.2.2
Segmentation
Segmentation is the process of separating an image into physically distinct regions such
as ice floes (that is, regions of differing local statistics, though the regions themselves
may exhibit continuous variations in the local mean intensity). Local operators for edge
enhancement and detection exist in profusion. The optimum statistical test for detecting the presence of an edge between two speckled regions of uniform intensity is Student's t-test (Kendall and Stuart 1977). However, this assumes that the boundary
between the two samples of pixel values to be compared in the t-test coincides with the
edge position. In practice, of course, the edge positions of a SAR image are unknown a
priori and so this requirement is impossible to satisfy.
A.T. SEPHTON AND K.C. PARTINGTON
classification in the (general) case where backscatter coefficient alone is an ambiguous
attribute for classification. Segmentation is described more fully in Sect. 12.2.2.
12.2.1.2
Inherent Variability in Backscatter
The second issue which needs to be addressed in any classification of ice concerns the
inherent variability in backscatter within ice types, between ice types, and within areas
of open water. A summary of early ERS-l measurements of backscatter coefficient provided by Partington and Hanna (1994) indicated that ranges of backscatter within a
single ice type could be as high as 12 dB (e.g., rough first-year ice), and that the range
of backscatter for multiyear ice is covered by that of rough first -year ice. Bright features
include ridges, old ice (which has both appreciable radar penetration depth and generally rough surface characteristics), the edges of ice floes (which can produce a multiple reflection geometry with the water surface), frost flowers, and wind-roughened
open water. Dark features include smooth water, new ice, frazil ice (millimeter-scale fragments of ice in open water), and areas of new ice frozen into sheets. However, there are
also variations in backscatter contrast between different ice types which depend on surface temperature (particularly on whether the ice is melting) and on large-scale geographic variations in the mode and rate of formation of sea ice (e.g. salinity variations,
typical compressive pressures, etc.). This problem, together with that of speckle, makes
automated classification difficult except under well-understood, restricted conditions.
It also makes a pixel-based classification system difficult to justify, despite significant
savings in computational time.
The alternative approach, alluded to earlier, is to make use of attributes other than
backscatter coefficient in any classification. This suggests a requirement for segmentation and for a classification scheme which makes use of region attributes including
backscatter coefficient and measures of image texture. In practice this is what happens
in a manual classification, with the human interpreter making use of texture and context as much as mean backscatter coefficient in any classification. In a hybrid classification, such as that described below, segmentation and classification are partially automated (with texture being provided as a set of attributes), but context can be included
in a classification through the manual classification of selected 'training' regions.
Region-based classification is described in Sect. 12.2.3.
12.2.2
Segmentation
Segmentation is the process of separating an image into physically distinct regions such
as ice floes (that is, regions of differing local statistics, though the regions themselves
may exhibit continuous variations in the local mean intensity). Local operators for edge
enhancement and detection exist in profusion. The optimum statistical test for detecting the presence of an edge between two speckled regions of uniform intensity is Student's t-test (Kendall and Stuart 1977). However, this assumes that the boundary
between the two samples of pixel values to be compared in the t-test coincides with the
edge position. In practice, of course, the edge positions of a SAR image are unknown a
priori and so this requirement is impossible to satisfy.