384
G.M. Foody
20.2
Neural Networks in Remote Sensing
Of the many pattern recognition applications, neural networks have been most
widely used in remote sensing for image classification and regression-type
analyses. They have been adopted increasingly as a result of their freedom from
restrictive assumptions as weIl as practical demonstrations of their ability to
provide more ace urate outputs than conventional methods.
20.2.1
Classification Applications
Remotely sensed data are extremely attractive for thematic mapping applications.
This is mainly because the data have a map-like format, provide complete and
continuous coverage of large areas inexpensively and are relatively consistent.
Consequently, remote sensing has become a major source of thematic maps such
as those depicting land cover.
Thematic mapping is typically based on some sort of classification analysis.
Two broad types of classification are used, supervised and unsupervised
(Schowengerdt 1997; Mather 1999). The unsupervised classifiers are essentially
clustering techniques. These are generally used to search for natural spectral
classes in the remotely sensed data. This type of approach can be useful in
exploratory analyses and where a basic generalisation of the data set is required.
Deriving meaningful labels for the derived clusters is, however, sometimes
difficult. More commonly, therefore, supervised classification techniques are
used. With this type of classification the classes of interested are known at the
outset. The classification algorithm is trained to recognise the remotely sensed
response of each class and then applied to the image in order to allocate each pixel
in the image to the class with which it has the greatest similarity (Campbell 1996;
Mather 1999). This type of approach has been widely used in studies of land
cover and land cover change. There are, however, many problems that often limit
the accuracy of classification analyses. Thus the considerable potential of remote
sensing as a source of land cover information is often not fully realised
(Townshend 1992; Wilkinson 1996; Estes et al. 1999). Many factors may be
responsible for this situation, ranging from the characteristics of the remote sensor
(e.g. its spectral, radiometrie, temporal and spatial resolutions), the nature of the
classes (e.g. level of class detail, degree of class heterogeneity) and the
classification methods used (Foody and Arora 1997; Scepan 1999; Estes et al.
1999; Loveland et al. 1999; Friedl et al. 2000). The latter issue has been the focus
of considerable attention. Recently attention has focused on neural network based
approaches for image classification which offer considerable advantages over
traditional methods (e.g. statistical approaches such as the maximum likelihood
classification), particularly with regard to the freedom from restrictive
G.M. Foody
20.2
Neural Networks in Remote Sensing
Of the many pattern recognition applications, neural networks have been most
widely used in remote sensing for image classification and regression-type
analyses. They have been adopted increasingly as a result of their freedom from
restrictive assumptions as weIl as practical demonstrations of their ability to
provide more ace urate outputs than conventional methods.
20.2.1
Classification Applications
Remotely sensed data are extremely attractive for thematic mapping applications.
This is mainly because the data have a map-like format, provide complete and
continuous coverage of large areas inexpensively and are relatively consistent.
Consequently, remote sensing has become a major source of thematic maps such
as those depicting land cover.
Thematic mapping is typically based on some sort of classification analysis.
Two broad types of classification are used, supervised and unsupervised
(Schowengerdt 1997; Mather 1999). The unsupervised classifiers are essentially
clustering techniques. These are generally used to search for natural spectral
classes in the remotely sensed data. This type of approach can be useful in
exploratory analyses and where a basic generalisation of the data set is required.
Deriving meaningful labels for the derived clusters is, however, sometimes
difficult. More commonly, therefore, supervised classification techniques are
used. With this type of classification the classes of interested are known at the
outset. The classification algorithm is trained to recognise the remotely sensed
response of each class and then applied to the image in order to allocate each pixel
in the image to the class with which it has the greatest similarity (Campbell 1996;
Mather 1999). This type of approach has been widely used in studies of land
cover and land cover change. There are, however, many problems that often limit
the accuracy of classification analyses. Thus the considerable potential of remote
sensing as a source of land cover information is often not fully realised
(Townshend 1992; Wilkinson 1996; Estes et al. 1999). Many factors may be
responsible for this situation, ranging from the characteristics of the remote sensor
(e.g. its spectral, radiometrie, temporal and spatial resolutions), the nature of the
classes (e.g. level of class detail, degree of class heterogeneity) and the
classification methods used (Foody and Arora 1997; Scepan 1999; Estes et al.
1999; Loveland et al. 1999; Friedl et al. 2000). The latter issue has been the focus
of considerable attention. Recently attention has focused on neural network based
approaches for image classification which offer considerable advantages over
traditional methods (e.g. statistical approaches such as the maximum likelihood
classification), particularly with regard to the freedom from restrictive