226
C. BERTOIA, J. FALKINGHAM, F. FETTERER
10.7
Improving Automated Classification
Based solely on expected backscatter, the single largest impediment to accurate ice classification is overlap of ice type backscatter distributions. The following methods might
be expected to improve classification of single-channel SAR data.
10.7.1
Dynamic Thresholding
Dynamic or adaptive thresholding divides an image into classes based on local, rather
than global, backscatter and a priori thresholds. (The ASF ERS-l algorithm is an example of a global threshold method.) Wackerman et al. (1988a,b) first suggested dynamic
thresholding for ice classification and pointed out that the method does not require
strictly calibrated imagery. Haverkamp et al. (1995) use dynamic thresholding to segment an ERS-l image into three ice types as a front end to an expert system. Thresholds are selected by examining the backscatter distributions for small areas within
images. When there is evidence that the local distribution is bimodal, a threshold is
selected that divides the populations represented by the two modes. This method
becomes complicated because metrics must be devised for choosing the right window
size, deciding if a distribution is bimodal, and determining how to interpolate between
thresholds in different parts of the image. The resulting classification is sensitive to the
scale of ice features relative to the scale of the local area used to determine the local
threshold. After the image is segmented, the classes must be labeled. In the implementation of Haverkamp et al., the number of classes is predefined, but, alternatively, a clustering method could be used to determine the number of classes supported by the data.
Theoretically, dynamic thresholding reduces error due to signature variability since
the class thresholds are determined based on local statistics. [This was remarked on,
but not quantified, by Haverkamp et al. (1995) J. Dynamic thresholding is promising for
ScanSAR RADARSAT data because it can potentially adjust for changes in backscatter
with incidence angle and across antenna beam patterns. It is also promising for segmenting imagery in the MIZ, where surface types (open water, grease ice, new ice)
exhibit varying but locally distinct backscatter distributions.
10.7.2
Texture
Ulaby et al. (1986) provide a precise definition of texture in SAR imagery and mathematically separate texture caused by intrinsic spatial variability of backscatter from texture caused by the presence of speckle. Adding parameters given by measures of intrinsic scene texture to measures of tone (Le., mean pixel backscatter) was shown to significantly improve (by 20% or more) classification of forest types in SeaS at imagery.
Results for sea ice have been less definitive. Since the initial use of texture for ice type
discrimination by Holmes et al. (1984), numerous papers have been published which
attempt to clarify the merits for classification of first-order statistics (measures such
as mean, variance, and skewness from an image or subimage) versus the discriminating ability of second-order, or texture, parameters. These are measures based on a
C. BERTOIA, J. FALKINGHAM, F. FETTERER
10.7
Improving Automated Classification
Based solely on expected backscatter, the single largest impediment to accurate ice classification is overlap of ice type backscatter distributions. The following methods might
be expected to improve classification of single-channel SAR data.
10.7.1
Dynamic Thresholding
Dynamic or adaptive thresholding divides an image into classes based on local, rather
than global, backscatter and a priori thresholds. (The ASF ERS-l algorithm is an example of a global threshold method.) Wackerman et al. (1988a,b) first suggested dynamic
thresholding for ice classification and pointed out that the method does not require
strictly calibrated imagery. Haverkamp et al. (1995) use dynamic thresholding to segment an ERS-l image into three ice types as a front end to an expert system. Thresholds are selected by examining the backscatter distributions for small areas within
images. When there is evidence that the local distribution is bimodal, a threshold is
selected that divides the populations represented by the two modes. This method
becomes complicated because metrics must be devised for choosing the right window
size, deciding if a distribution is bimodal, and determining how to interpolate between
thresholds in different parts of the image. The resulting classification is sensitive to the
scale of ice features relative to the scale of the local area used to determine the local
threshold. After the image is segmented, the classes must be labeled. In the implementation of Haverkamp et al., the number of classes is predefined, but, alternatively, a clustering method could be used to determine the number of classes supported by the data.
Theoretically, dynamic thresholding reduces error due to signature variability since
the class thresholds are determined based on local statistics. [This was remarked on,
but not quantified, by Haverkamp et al. (1995) J. Dynamic thresholding is promising for
ScanSAR RADARSAT data because it can potentially adjust for changes in backscatter
with incidence angle and across antenna beam patterns. It is also promising for segmenting imagery in the MIZ, where surface types (open water, grease ice, new ice)
exhibit varying but locally distinct backscatter distributions.
10.7.2
Texture
Ulaby et al. (1986) provide a precise definition of texture in SAR imagery and mathematically separate texture caused by intrinsic spatial variability of backscatter from texture caused by the presence of speckle. Adding parameters given by measures of intrinsic scene texture to measures of tone (Le., mean pixel backscatter) was shown to significantly improve (by 20% or more) classification of forest types in SeaS at imagery.
Results for sea ice have been less definitive. Since the initial use of texture for ice type
discrimination by Holmes et al. (1984), numerous papers have been published which
attempt to clarify the merits for classification of first-order statistics (measures such
as mean, variance, and skewness from an image or subimage) versus the discriminating ability of second-order, or texture, parameters. These are measures based on a