Chapter 20 . Classification of Remotely Sensed Images
385
assumptions about the data sets (Fischer et al. 1997; Fischer 1998). Comparative
analyses have also generally shown that neural networks may be used to c1assify
remotely sensed data more accurately than conventional statistical approaches,
particularly if the data are incompatible with the assumed model that underlies
statistical c1assifications (Benediktsson et al. 1990; Peddle et al. 1994; Paola and
Schowengerdt 1995).
20.2.2
Regression Applications
Commonly remotely sensed data have been used to derive estimates of
environmental variables. Thus, for example, remotely sensed data have been used
to estimate a range of environmental phenomena inc1uding the biom ass of forests,
moisture content of soils and the temperature and sediment content of water
bodies. This is often the only feasible way of deriving information on the
variables over large areas and the only practical source of spatial data to drive
some ecosystem simulation models (Lucas and Curran 1999). A variety of
methods may be used to derive the estimates of environmental variables but
regression techniques have proved popular (e.g. Lawrence and Ripple 1998
Curran et al. 2001). As with the c1assification of remotely sensed data, a
fundamental problem is that the assumptions that underlie regression analysis may
not be satisfied, leading to significant error and misinterpretation. Neural
networks, however, have considerable potential as non-parametric alternative to
regression analysis and in recent years have been used increasingly in the
extraction of environmental information from remotely sensed data (e.g. Baret
1995; Jin and Liu 1997; Foody et al. 2001). Similarly, neural networks have been
widely used in the related application of model inversion (e.g. Wang and Dong
1997).
20.3
The Neural Networks Used in Remote Sensing
A variety of neural networks have been used in remote sensing. This section aims
to briefly outline the key features of the most widely used networks. Emphasis
will be placed upon the feedforward networks, such as the multi-Iayer perceptron
(MLP), radial basis function (RBF) probabilistic neural network (PNN) and
generalised regression neural networks (GRNN), that have been used most widely
in remote sensing. Other network types that have been used less frequently but for
wh ich there is considerable scope for further application in remote sensing are
also discussed briefly.
385
assumptions about the data sets (Fischer et al. 1997; Fischer 1998). Comparative
analyses have also generally shown that neural networks may be used to c1assify
remotely sensed data more accurately than conventional statistical approaches,
particularly if the data are incompatible with the assumed model that underlies
statistical c1assifications (Benediktsson et al. 1990; Peddle et al. 1994; Paola and
Schowengerdt 1995).
20.2.2
Regression Applications
Commonly remotely sensed data have been used to derive estimates of
environmental variables. Thus, for example, remotely sensed data have been used
to estimate a range of environmental phenomena inc1uding the biom ass of forests,
moisture content of soils and the temperature and sediment content of water
bodies. This is often the only feasible way of deriving information on the
variables over large areas and the only practical source of spatial data to drive
some ecosystem simulation models (Lucas and Curran 1999). A variety of
methods may be used to derive the estimates of environmental variables but
regression techniques have proved popular (e.g. Lawrence and Ripple 1998
Curran et al. 2001). As with the c1assification of remotely sensed data, a
fundamental problem is that the assumptions that underlie regression analysis may
not be satisfied, leading to significant error and misinterpretation. Neural
networks, however, have considerable potential as non-parametric alternative to
regression analysis and in recent years have been used increasingly in the
extraction of environmental information from remotely sensed data (e.g. Baret
1995; Jin and Liu 1997; Foody et al. 2001). Similarly, neural networks have been
widely used in the related application of model inversion (e.g. Wang and Dong
1997).
20.3
The Neural Networks Used in Remote Sensing
A variety of neural networks have been used in remote sensing. This section aims
to briefly outline the key features of the most widely used networks. Emphasis
will be placed upon the feedforward networks, such as the multi-Iayer perceptron
(MLP), radial basis function (RBF) probabilistic neural network (PNN) and
generalised regression neural networks (GRNN), that have been used most widely
in remote sensing. Other network types that have been used less frequently but for
wh ich there is considerable scope for further application in remote sensing are
also discussed briefly.