Chapter 20 . Classification of Remotely Sensed Images
393
2001) and ART (Carpenter et al. 1999a; Gopal et al. 1999) networks which may
sometimes be more appropriate than the MLP.
Despite the demonstrated potential of neural networks for classification and
regression applications a variety of concerns and problems have been noted. For
example, the difference in accuracy relative to that obtainable from a conventional
approach is not always large, it can be difficult to understand the results derived
and the analysis is based upon a set of subjective decisions (Wilkinson 1997;
Foody 1999a). Moreover, if the assumptions underlying a model based technique
are satisfied it has been suggested that that technique be used instead of the neural
network in order to exploit fuHy advantage conveyed by the underlying model.
Sometimes the problems encountered with neural networks relate to general
issues. For example, although a neural network classification may make no
assumptions about the data sets used the classification analysis is founded on a set
of assumptions. Key amongst these are the commonly made assumptions that the
classes have been defined exhaustively and are discrete and mutuaHy exclusive.
While these assumptions can be satisfied in demonstration projects that have
shown the potential of neural networks they are less satisfiable in 'real world'
studies. Thus the accuracy of neural network analyses may still be insufficient for
some users (Wilkinson 1996). Frequently, the problem here is that neural
networks have generally been used to derive conventional 'hard' classifications
(in wh ich each pixel is allocated firmly to a single class). This is often
inappropriate due to the presence of mixed pixels, that typically dominate
remotely sensed data sets used in macro scale studies (Campbell 1996; Foody et
al. 1997). Moreover, in many studies the classes are continuous rather than
discrete. Additionally, the set of classes may not have been defined exhaustively.
There are, however, means of reducing these problems. For example, the
problems associated with continuous classes and mixed pixels may be reduced by
using the network to derive a soft or fuzzy classification rather than the
conventional hard classification (Foody 1996). Fuzzy classifications have been
derived using a range of networks, notably the MLP (Foody 1999b) and
ARTMAP (Gopal et al. 1999). These have been found to provide accurate
estimates of sub-pixel class composition and provide arieher thematic
representation that the conventional hard classification. Additionally, using a
Hopfield neural network the sub-pixel fractions derived from a soft classification
may be located within the area represented by a pixel to derive an enhanced or
super resolution thematic representation (Tatem et al. 2001). The effect of a nonexhaustively defined set of classes mayaIso be reduced by post-classification
thresholding of the network outputs or use of a network such as the RBF that
partitions feature space locally and so is less prone to untenable extrapolation than
the widely used MLP (Vasconcelos et al. 1995; Foody 2001).
393
2001) and ART (Carpenter et al. 1999a; Gopal et al. 1999) networks which may
sometimes be more appropriate than the MLP.
Despite the demonstrated potential of neural networks for classification and
regression applications a variety of concerns and problems have been noted. For
example, the difference in accuracy relative to that obtainable from a conventional
approach is not always large, it can be difficult to understand the results derived
and the analysis is based upon a set of subjective decisions (Wilkinson 1997;
Foody 1999a). Moreover, if the assumptions underlying a model based technique
are satisfied it has been suggested that that technique be used instead of the neural
network in order to exploit fuHy advantage conveyed by the underlying model.
Sometimes the problems encountered with neural networks relate to general
issues. For example, although a neural network classification may make no
assumptions about the data sets used the classification analysis is founded on a set
of assumptions. Key amongst these are the commonly made assumptions that the
classes have been defined exhaustively and are discrete and mutuaHy exclusive.
While these assumptions can be satisfied in demonstration projects that have
shown the potential of neural networks they are less satisfiable in 'real world'
studies. Thus the accuracy of neural network analyses may still be insufficient for
some users (Wilkinson 1996). Frequently, the problem here is that neural
networks have generally been used to derive conventional 'hard' classifications
(in wh ich each pixel is allocated firmly to a single class). This is often
inappropriate due to the presence of mixed pixels, that typically dominate
remotely sensed data sets used in macro scale studies (Campbell 1996; Foody et
al. 1997). Moreover, in many studies the classes are continuous rather than
discrete. Additionally, the set of classes may not have been defined exhaustively.
There are, however, means of reducing these problems. For example, the
problems associated with continuous classes and mixed pixels may be reduced by
using the network to derive a soft or fuzzy classification rather than the
conventional hard classification (Foody 1996). Fuzzy classifications have been
derived using a range of networks, notably the MLP (Foody 1999b) and
ARTMAP (Gopal et al. 1999). These have been found to provide accurate
estimates of sub-pixel class composition and provide arieher thematic
representation that the conventional hard classification. Additionally, using a
Hopfield neural network the sub-pixel fractions derived from a soft classification
may be located within the area represented by a pixel to derive an enhanced or
super resolution thematic representation (Tatem et al. 2001). The effect of a nonexhaustively defined set of classes mayaIso be reduced by post-classification
thresholding of the network outputs or use of a network such as the RBF that
partitions feature space locally and so is less prone to untenable extrapolation than
the widely used MLP (Vasconcelos et al. 1995; Foody 2001).