CHAPTER 10
Support Vector Machines for Classification
of Multi- and Hyperspectral Data
Pakorn Watanachaturaporn, Manoj K. Arora
10.1
Introduction
As discussed in Chap. 5, support vector machines (SVMs) have originated from
statistical learning theory for classification and regression problems. Unlike
the popular neural network classifiers, SVMs do not minimize the empirical
training error (Byun and Lee 2003). Instead, they aim to maximize the margin between two classes of interest by placing a linear separating hyperplane
between them. While doing so, the upper bound on the generalization error
is minimized. Thus, SVM based classifiers are expected to have more generalization capability than neural networks. Other advantages of SVMs are
their ability to adapt their learning characteristic via a kernel function and
to adequately classify data on a high-dimensional feature space with a limited
number of training data sets thereby overcoming the Hughes Phenomenon. The
theoretical background on SVMs has been presented in Chap. 5. This chapter
builds on this theoretical knowledge to apply SVMs for classification of multi
and hyperspectral remote sensing data.
SVMs are good candidates for remote sensing image classification for anumber of reasons. Firstly, an SVM can work well with a small training data set.
Selection of a sufficient number of pure training pixels (i. e. pixels belonging to
only one class) has always been a problem even though a remote sensing image
contains as many as hundreds of thousands of pixels. The problem becomes
severe in coarse resolution images where mixed pixels, i. e. pixels containing
more than one class, may be abundant. Secondly, SVMs have been found to
perform with high classification accuracy for data having hundreds of dimensions in applications such as the classification of personal income using 123
parameters in the data set (Keerthi 2002). In remote sensing these days, the
hyperspectral data are available in hundreds of bands. The processing of such
high-dimensional data to extract quality information will always remain a challenge. Conventional statistical classifiers fail to process high dimensional data
due to the requirement of large number of training samples. SVMs have the
potential to produce accurate classifications from high dimensional data with
limited number of training samples. Thirdly, unlike neural networks, SVMs are
robust to the overfitting problem as they rely on margin maximization rather
than finding a decision boundary directly from the training samples. Fourthly,
P. K. Varshney et al., Advanced Image Processing Techniques for Remotely Sensed Hyperspectral Data
© Springer-Verlag Berlin Heidelberg 2004
Support Vector Machines for Classification
of Multi- and Hyperspectral Data
Pakorn Watanachaturaporn, Manoj K. Arora
10.1
Introduction
As discussed in Chap. 5, support vector machines (SVMs) have originated from
statistical learning theory for classification and regression problems. Unlike
the popular neural network classifiers, SVMs do not minimize the empirical
training error (Byun and Lee 2003). Instead, they aim to maximize the margin between two classes of interest by placing a linear separating hyperplane
between them. While doing so, the upper bound on the generalization error
is minimized. Thus, SVM based classifiers are expected to have more generalization capability than neural networks. Other advantages of SVMs are
their ability to adapt their learning characteristic via a kernel function and
to adequately classify data on a high-dimensional feature space with a limited
number of training data sets thereby overcoming the Hughes Phenomenon. The
theoretical background on SVMs has been presented in Chap. 5. This chapter
builds on this theoretical knowledge to apply SVMs for classification of multi
and hyperspectral remote sensing data.
SVMs are good candidates for remote sensing image classification for anumber of reasons. Firstly, an SVM can work well with a small training data set.
Selection of a sufficient number of pure training pixels (i. e. pixels belonging to
only one class) has always been a problem even though a remote sensing image
contains as many as hundreds of thousands of pixels. The problem becomes
severe in coarse resolution images where mixed pixels, i. e. pixels containing
more than one class, may be abundant. Secondly, SVMs have been found to
perform with high classification accuracy for data having hundreds of dimensions in applications such as the classification of personal income using 123
parameters in the data set (Keerthi 2002). In remote sensing these days, the
hyperspectral data are available in hundreds of bands. The processing of such
high-dimensional data to extract quality information will always remain a challenge. Conventional statistical classifiers fail to process high dimensional data
due to the requirement of large number of training samples. SVMs have the
potential to produce accurate classifications from high dimensional data with
limited number of training samples. Thirdly, unlike neural networks, SVMs are
robust to the overfitting problem as they rely on margin maximization rather
than finding a decision boundary directly from the training samples. Fourthly,
P. K. Varshney et al., Advanced Image Processing Techniques for Remotely Sensed Hyperspectral Data
© Springer-Verlag Berlin Heidelberg 2004