Chapter 20 . Classification of Remotely Sensed Images
391
this third layer than in the output layer. The fourth layer is the output layer that
integrates the outputs from the regression layer to provide the networks
predictions. Relative to MLP and RBF networks constructed for the same
problem, a GRNN typically has a much larger architecture (e.g. Foody et al.
2001). Because of the large size of this type of network it is typically slow to
apply to large data sets, although it does train rapidly.
20.3.1.5
Other Network Types
A range of other networks has been used in remote sensing. Three in particular are
worthy of brief discussion. These are the Hopfield, ART and Kohonen neural
networks. Each is very dissimilar to the feedforward networks that have been
described above.
The Hopfield neural network belongs to a set of recurrent networks. Unlike the
feedforward networks, in which information is essentially propagated in a single
direction, the outputs of units in a Hopfield network are fed back into the inputs.
The Hopfield network is, therefore, a fully connected network (Aleksander and
Morton 1990; Davalo and Naim 1991). It was used initially as a contents
addressable (associative) memory device but is also able to solve complex
combinatorial problems (Cote and Tatnall1997). The Hopfield neural network has
been used in a range of applications in remote sensing including the tracking of
clouds (Cote and Tatnall 1997) and in the refinement of soft/fuzzy image
classifications in order to effectively sharpen the resolution of thematic maps
(Tatem et al. 2001).
Adaptive resonance theory (ART) provides the foundation for real-time
networks that solve the stability-plasticity problem that handicaps the use of other
approaches such as the MLP (Carpenter et al. 1999a; Gopal et al. 1999; Gamba
and Houshmand 2001). That is, the ART networks are able to perform rapid,
stable on-line learning, recognition and prediction. ART networks may be adapted
to accommodate fuzzy logic and have been used in supervised image
classification applications (Gopal et al. 1999; Carpenter et al. 1999b).
The Kohonen network belongs to a set of self-organising competitive learning
systems. Sometimes referred to in the literature as a self-organising map (SOM) or
self-organising feature map (SOFM), the Kohonen network has been used
commonly for unsupervised classification. The network uses unsupervised
learning to provide a topologically ordered output that displays the relative
similarity of cases entered to it. The network consists of just two layers of units,
input and output. The output is a low, typically 2, dimensional array of units. Each
unit in the output layer is fully connected to all adjacent units in that layer as weil
as to all of the input units. The lateral connection of units in the output layer aids
competitive learning such that similar cases cluster together in the output array
and are associated with a different region of the output array than dissimilar cases.
This network has been used for unsupervised classification (Poth et al. 2001),
391
this third layer than in the output layer. The fourth layer is the output layer that
integrates the outputs from the regression layer to provide the networks
predictions. Relative to MLP and RBF networks constructed for the same
problem, a GRNN typically has a much larger architecture (e.g. Foody et al.
2001). Because of the large size of this type of network it is typically slow to
apply to large data sets, although it does train rapidly.
20.3.1.5
Other Network Types
A range of other networks has been used in remote sensing. Three in particular are
worthy of brief discussion. These are the Hopfield, ART and Kohonen neural
networks. Each is very dissimilar to the feedforward networks that have been
described above.
The Hopfield neural network belongs to a set of recurrent networks. Unlike the
feedforward networks, in which information is essentially propagated in a single
direction, the outputs of units in a Hopfield network are fed back into the inputs.
The Hopfield network is, therefore, a fully connected network (Aleksander and
Morton 1990; Davalo and Naim 1991). It was used initially as a contents
addressable (associative) memory device but is also able to solve complex
combinatorial problems (Cote and Tatnall1997). The Hopfield neural network has
been used in a range of applications in remote sensing including the tracking of
clouds (Cote and Tatnall 1997) and in the refinement of soft/fuzzy image
classifications in order to effectively sharpen the resolution of thematic maps
(Tatem et al. 2001).
Adaptive resonance theory (ART) provides the foundation for real-time
networks that solve the stability-plasticity problem that handicaps the use of other
approaches such as the MLP (Carpenter et al. 1999a; Gopal et al. 1999; Gamba
and Houshmand 2001). That is, the ART networks are able to perform rapid,
stable on-line learning, recognition and prediction. ART networks may be adapted
to accommodate fuzzy logic and have been used in supervised image
classification applications (Gopal et al. 1999; Carpenter et al. 1999b).
The Kohonen network belongs to a set of self-organising competitive learning
systems. Sometimes referred to in the literature as a self-organising map (SOM) or
self-organising feature map (SOFM), the Kohonen network has been used
commonly for unsupervised classification. The network uses unsupervised
learning to provide a topologically ordered output that displays the relative
similarity of cases entered to it. The network consists of just two layers of units,
input and output. The output is a low, typically 2, dimensional array of units. Each
unit in the output layer is fully connected to all adjacent units in that layer as weil
as to all of the input units. The lateral connection of units in the output layer aids
competitive learning such that similar cases cluster together in the output array
and are associated with a different region of the output array than dissimilar cases.
This network has been used for unsupervised classification (Poth et al. 2001),