A.J. SEPHTON AND K.C. PARTINGTON
While in principle almost any given level of edge detection can be achieved by
smoothing the image to increase the effective number of looks, it should of course be
noted that this will in turn produce a degradation in the spatial resolution. For an ERS-1
image, with a spatial resolution of approximately 24 m in range and azimuth for threelook data, the degradation in spatial resolution can be represented as in Fig. 2b.
The values of crill obtained in the edge detection process do not directly provide edge
positions in the image as they contain no edge direction information. Instead, the values of crill are used to "bond" neighboring pixels so that pixels are linked by bonds to
all other pixels in the same region, but not to pixels in any other region. A typical bonding method used is that described by Oddy and Rye (1983). As a consequence of artificially dividing some regions in the bonding process it may be necessary to include a
step of region merging, based for instance on the t-test (Sephton et al. 1994), before
attributes are derived for each region.
12.2.3
Classification and Concentration
The number of approaches to classification is large: this section describes the philosophy adopted within the IPAP system. We have already discounted a pixel-based
approach on a priori grounds. There is a further decision to be made regarding whether
supervised or nonsupervised methods are applicable. We have selected the supervised
approach, on the basis that we desire manual intervention in assigning the attribute
characteristics to a class before beginning classification. The unsupervised approach
finds its own class characteristics and members before requiring the user to assign
labels to each class, but suffers the disadvantage of requiring the number of ice classes
present to be known a priori. Supervised classification options include parametric classification, which requires assumptions about the distribution of samples in a class, and
is therefore considered less appropriate than nonparametric approaches, which make
no such assumptions. So we reject (for example) maximum likelihood classification in
favor of K -nearest neighbor (K -NN) and neural network, multi -layer perceptron (MLP)
methods.
The K-NN method is based on the assumption that vectors from a given class tend
to be close together, so that an unknown vector should be given the same class as some
K-weighted set of vectors which form the nearest class. For K == 1, the error rate is close
to twice the Bayes error (which defines the minimum possible error) and as K increases, the error rate approaches the Bayes limit asymptotically (Cover and Hart 1967).
The MLP method is based on the principle that sequential Von Neumann computers
are exceeded in performance by processors which mimic animal brains through structures which contain heavily interconnected nodes, and that this improved performance
extends even to the case where the number of processing units may be tens rather than
the many orders of magnitude more that exist in animal brains (Lippman 1987). An
MLP typically contains three layers, with nodes in each layer interconnected to nodes
in layers above and below. The weights associated with each node allow the predicted
class to be matched to the actual class by backwards adjustment of weights within the
network following poor matches.
Supervised classification is a two-stage process. In the first stage, selected regions are
classified manually. A judgment needs to be made on how many regions will be suffi-
While in principle almost any given level of edge detection can be achieved by
smoothing the image to increase the effective number of looks, it should of course be
noted that this will in turn produce a degradation in the spatial resolution. For an ERS-1
image, with a spatial resolution of approximately 24 m in range and azimuth for threelook data, the degradation in spatial resolution can be represented as in Fig. 2b.
The values of crill obtained in the edge detection process do not directly provide edge
positions in the image as they contain no edge direction information. Instead, the values of crill are used to "bond" neighboring pixels so that pixels are linked by bonds to
all other pixels in the same region, but not to pixels in any other region. A typical bonding method used is that described by Oddy and Rye (1983). As a consequence of artificially dividing some regions in the bonding process it may be necessary to include a
step of region merging, based for instance on the t-test (Sephton et al. 1994), before
attributes are derived for each region.
12.2.3
Classification and Concentration
The number of approaches to classification is large: this section describes the philosophy adopted within the IPAP system. We have already discounted a pixel-based
approach on a priori grounds. There is a further decision to be made regarding whether
supervised or nonsupervised methods are applicable. We have selected the supervised
approach, on the basis that we desire manual intervention in assigning the attribute
characteristics to a class before beginning classification. The unsupervised approach
finds its own class characteristics and members before requiring the user to assign
labels to each class, but suffers the disadvantage of requiring the number of ice classes
present to be known a priori. Supervised classification options include parametric classification, which requires assumptions about the distribution of samples in a class, and
is therefore considered less appropriate than nonparametric approaches, which make
no such assumptions. So we reject (for example) maximum likelihood classification in
favor of K -nearest neighbor (K -NN) and neural network, multi -layer perceptron (MLP)
methods.
The K-NN method is based on the assumption that vectors from a given class tend
to be close together, so that an unknown vector should be given the same class as some
K-weighted set of vectors which form the nearest class. For K == 1, the error rate is close
to twice the Bayes error (which defines the minimum possible error) and as K increases, the error rate approaches the Bayes limit asymptotically (Cover and Hart 1967).
The MLP method is based on the principle that sequential Von Neumann computers
are exceeded in performance by processors which mimic animal brains through structures which contain heavily interconnected nodes, and that this improved performance
extends even to the case where the number of processing units may be tens rather than
the many orders of magnitude more that exist in animal brains (Lippman 1987). An
MLP typically contains three layers, with nodes in each layer interconnected to nodes
in layers above and below. The weights associated with each node allow the predicted
class to be matched to the actual class by backwards adjustment of weights within the
network following poor matches.
Supervised classification is a two-stage process. In the first stage, selected regions are
classified manually. A judgment needs to be made on how many regions will be suffi-