392
G.M. Foody
although if training data are available it mayaiso be used to label cases (e.g. Ito
and Omato 1997; Ji 2000).
20.4
Current Status
The potential of neural networks for the provision of accurate environmental
information from remotely sensed data has been demonstrated in numerous
studies. With regard to both classification and regression type problems this has
been largely demonstrated with the MLP network.
Many studies have demonstrated that neural networks can generally be used to
classify remotely sensed data at least as, but usually more, accurately than
conventional statistical classifiers (e.g. Benediktsson et al. 1990). For example,
Peddle et al. (1994) provide a comparative evaluation of neural network,
maximum likelihood and evidential reasoning classifications of alpine land cover.
The results show that for a variety of classification scenarios based upon SPOT
HRV and ancillary topographic data that, in general, the highest accuracy was
obtainable when a neural network was used to classify the data. For example, in
one set of analyses it was apparent that the accuracy of a maximum likelihood
classification declined with the addition of ancillary information on topography
from 75% to 32% approximately while the corresponding accuracies derived from
a neural network showed and increase from 75% to 90% approximately. This
result is significant in that it shows the neural network was, unlike the
conventional statistical c1assification, able to constructively use the ancillary
information and derive a significantly higher c1assification accuracy.
The MLP has also been used successfully in regression applications. Many
studies have focussed on the estimation of biophysical properties from remotely
sensed data (e.g. Weiss and Baret 1999; Kimes et al. 2000). Studies that illustrate
the value of neural networks for this type of application include those focused on
the estimation of leaf area index (Smith 1993) and forest biomass (Foody et al.
2001). In terms of relative performance against standard regression analysis,
neural networks have been shown to be able to derive more accurate predictions.
For example, in the estimation of tropical forest biomass from Landsat TM data,
Foody et al. (2001) derived correlation coefficients between the biomass predicted
from a multiple regression analysis and a MLP neural network with that derived
from field measurement of r=0.50 and r=0.80 respectively.
Although the MLP is the most widely used neural network in remote sensing
data analysis attention is increasingly turning to alternative network types. For
supervised classification applications, perhaps the most common use of neural
networks in remote sensing, attention has recently turned increasingly to the RBF
(Fischer et al. 1997; Rollet et al. 1998; Bruzzone and Prieto 1999), PNN (Foody
G.M. Foody
although if training data are available it mayaiso be used to label cases (e.g. Ito
and Omato 1997; Ji 2000).
20.4
Current Status
The potential of neural networks for the provision of accurate environmental
information from remotely sensed data has been demonstrated in numerous
studies. With regard to both classification and regression type problems this has
been largely demonstrated with the MLP network.
Many studies have demonstrated that neural networks can generally be used to
classify remotely sensed data at least as, but usually more, accurately than
conventional statistical classifiers (e.g. Benediktsson et al. 1990). For example,
Peddle et al. (1994) provide a comparative evaluation of neural network,
maximum likelihood and evidential reasoning classifications of alpine land cover.
The results show that for a variety of classification scenarios based upon SPOT
HRV and ancillary topographic data that, in general, the highest accuracy was
obtainable when a neural network was used to classify the data. For example, in
one set of analyses it was apparent that the accuracy of a maximum likelihood
classification declined with the addition of ancillary information on topography
from 75% to 32% approximately while the corresponding accuracies derived from
a neural network showed and increase from 75% to 90% approximately. This
result is significant in that it shows the neural network was, unlike the
conventional statistical c1assification, able to constructively use the ancillary
information and derive a significantly higher c1assification accuracy.
The MLP has also been used successfully in regression applications. Many
studies have focussed on the estimation of biophysical properties from remotely
sensed data (e.g. Weiss and Baret 1999; Kimes et al. 2000). Studies that illustrate
the value of neural networks for this type of application include those focused on
the estimation of leaf area index (Smith 1993) and forest biomass (Foody et al.
2001). In terms of relative performance against standard regression analysis,
neural networks have been shown to be able to derive more accurate predictions.
For example, in the estimation of tropical forest biomass from Landsat TM data,
Foody et al. (2001) derived correlation coefficients between the biomass predicted
from a multiple regression analysis and a MLP neural network with that derived
from field measurement of r=0.50 and r=0.80 respectively.
Although the MLP is the most widely used neural network in remote sensing
data analysis attention is increasingly turning to alternative network types. For
supervised classification applications, perhaps the most common use of neural
networks in remote sensing, attention has recently turned increasingly to the RBF
(Fischer et al. 1997; Rollet et al. 1998; Bruzzone and Prieto 1999), PNN (Foody