72
2: Raghuveer M. Rao, Manoj K. Arora
Landsat TM
Band I
Agriculture
Band 2
Built-up
Band 3
Forest
Band 4
Sand
Band 5
Water
Band 6
Fig. 2.11. A typical neural network architecture
gorithms such as neural network (Heermann and Khazenie 1992; Cipollini et
al. 2001), knowledge based (Kimes 1991), decision tree (Brodley et al. 1996),
genetic algorithms (Roberts 2000) and support vector machines (Melgani and
Bruzzone 2002) are in vogue. In fact, the neural network approach has been
viewed as a replacement of the most widely used MLC in the remote sensing
community. A brief description of a neural network classifier is provided here.
Details on other classifiers can be found in the references cited. Since support
vector machine is a recent development in image classification, focus has been
placed on this classifier. As a result two chapters (Chap. 5 and Chap. 10) have
been included in this book to discuss this classifier in greater detail.
Typically, a neural network consists of an input layer, a hidden layer and an
output layer (Fig. 2.11). The input layer is passive and merely receives the data
(e. g. the multispectral remote sensing data). Consequently, the units in the
input layer equal the number of bands used in classification. Unlike the input
layer, both, hidden and output layers actively process the data. The output
layer, as the name suggests, produces the neural network results. The number
of units in the output layer is generally kept equal to the number of classes
to be mapped. Hence, the number of units in the input and the output layers
are typically fixed by the application designed. Introducing hidden layer between input and output layer increases the network's ability to model complex
functions (Mueller and Hammerstorm 1992). Selection of appropriate number
of hidden layers and their units is critical for the successful operation of the
neural network. With too few hidden units, the network may not be powerful enough to process the data. On the other hand, with a large number of
hidden units, computation is expensive and the network may also memorize
the training samples instead of learning from the training samples. This may
2: Raghuveer M. Rao, Manoj K. Arora
Landsat TM
Band I
Agriculture
Band 2
Built-up
Band 3
Forest
Band 4
Sand
Band 5
Water
Band 6
Fig. 2.11. A typical neural network architecture
gorithms such as neural network (Heermann and Khazenie 1992; Cipollini et
al. 2001), knowledge based (Kimes 1991), decision tree (Brodley et al. 1996),
genetic algorithms (Roberts 2000) and support vector machines (Melgani and
Bruzzone 2002) are in vogue. In fact, the neural network approach has been
viewed as a replacement of the most widely used MLC in the remote sensing
community. A brief description of a neural network classifier is provided here.
Details on other classifiers can be found in the references cited. Since support
vector machine is a recent development in image classification, focus has been
placed on this classifier. As a result two chapters (Chap. 5 and Chap. 10) have
been included in this book to discuss this classifier in greater detail.
Typically, a neural network consists of an input layer, a hidden layer and an
output layer (Fig. 2.11). The input layer is passive and merely receives the data
(e. g. the multispectral remote sensing data). Consequently, the units in the
input layer equal the number of bands used in classification. Unlike the input
layer, both, hidden and output layers actively process the data. The output
layer, as the name suggests, produces the neural network results. The number
of units in the output layer is generally kept equal to the number of classes
to be mapped. Hence, the number of units in the input and the output layers
are typically fixed by the application designed. Introducing hidden layer between input and output layer increases the network's ability to model complex
functions (Mueller and Hammerstorm 1992). Selection of appropriate number
of hidden layers and their units is critical for the successful operation of the
neural network. With too few hidden units, the network may not be powerful enough to process the data. On the other hand, with a large number of
hidden units, computation is expensive and the network may also memorize
the training samples instead of learning from the training samples. This may