238
10: Pakorn Watanachaturaporn, Manoj K. Arora
the structure of an SVM is less complex even with the high dimensional data.
Neural networks, for example, have a very complex structure for processing
high dimensional data. Finally, as compared to another recent nonparametric
classifier namely the decision tree classifier, an SVM does not require the generation of rules that heavily depend on the knowledge from experts. This is
crucial to achieve high classification accuracy.
Classification of remote sensing data using SVMs has been introduced recently. Gualtieri and Cromp (1998) and Gualtieri et al. (1999) used SVMs to
classify an AVIRIS image. The results were compared with those obtained
from 'the extraction and classification of homogeneous objects' (ECHO) classifier and the Euclidean classifier (Tadjudin and Landgrebe, 1998). Accuracy
of 87.3%,82.9%, and 48.2% were obtained from SVM, ECHO, and Euclidean
classifiers respectively. In 2002, Zhu and Blumberg (2002) classified an ASTER
image using SVMs. They reported the results of a classification experiment
with an SVM using polynomial kernels of different degrees and the radial basis function (RBF) kernel. The overall accuracies were obtained in the range of
87% to 91 %. Melgani and Bruzzone (2002) used an SVM to classify an AVIRIS
image and compared that with K-Nearest Neighbors (K-NN), and RBF neural network (RBF-NN) classifiers, and obtained accuracies of 93.42%, 83.94%,
and 86.99% from SVM, K-NN, and RBF classifiers respectively. Huang et al.
(2002) intensively investigated the accuracy obtained from SVM classifiers on
a Landsat TM image. They reported an accuracy of 75.62% from the SVM,
74.02% from the back propagation neural network, 73.31 % from the decision
tree, and 7l.76% from the maximum likelihood classifiers. Shah et al. (2003)
reported the accuracies attained by SVM, MLC, and back-propagation neural
network classifiers when applied on an AVIRIS image with and without the use
of supervised feature extraction techniques. They used the LSVM (see Chap. 5)
with a linear kernel, polynomial kernel of different degrees, a RBF kernel, and
a sigmoid kernel. They showed that accuracies of 90.9% to 97.2% were obtained
from SVM classifiers while the maximum likelihood and neural network classifiers could achieve accuracies of 59.4% and 6l.9% respectively on the full
dimensionality dataset. They also showed that the classification results from
the full dimensionality dataset were better than the classification of features
obtained using the discriminant analysis feature extraction (DAFE) and the
decision boundary feature extraction (DBFE) techniques (Lee and Landgrebe
1993).
From these limited studies, it can be seen that in all the cases SVM derived
classifications of both multi- or hyperspectral datasets produced the highest
accuracy. However, there are many issues, which need to be further investigated
before SVM classifiers can be implemented at the operational level.
The aim of this chapter is to understand the application of SVMs and the
behavior of associated parameters for the classification of multi and hyperspectral remote sensing data. In the next section, the details of parameters
considered are provided. Section 10.3 describes the remote sensing data used
to produce SVM classification. Experimental set up, results and their analyses
is presented in Sect. 10.4. A summary is provided in Sect. 10.5.
10: Pakorn Watanachaturaporn, Manoj K. Arora
the structure of an SVM is less complex even with the high dimensional data.
Neural networks, for example, have a very complex structure for processing
high dimensional data. Finally, as compared to another recent nonparametric
classifier namely the decision tree classifier, an SVM does not require the generation of rules that heavily depend on the knowledge from experts. This is
crucial to achieve high classification accuracy.
Classification of remote sensing data using SVMs has been introduced recently. Gualtieri and Cromp (1998) and Gualtieri et al. (1999) used SVMs to
classify an AVIRIS image. The results were compared with those obtained
from 'the extraction and classification of homogeneous objects' (ECHO) classifier and the Euclidean classifier (Tadjudin and Landgrebe, 1998). Accuracy
of 87.3%,82.9%, and 48.2% were obtained from SVM, ECHO, and Euclidean
classifiers respectively. In 2002, Zhu and Blumberg (2002) classified an ASTER
image using SVMs. They reported the results of a classification experiment
with an SVM using polynomial kernels of different degrees and the radial basis function (RBF) kernel. The overall accuracies were obtained in the range of
87% to 91 %. Melgani and Bruzzone (2002) used an SVM to classify an AVIRIS
image and compared that with K-Nearest Neighbors (K-NN), and RBF neural network (RBF-NN) classifiers, and obtained accuracies of 93.42%, 83.94%,
and 86.99% from SVM, K-NN, and RBF classifiers respectively. Huang et al.
(2002) intensively investigated the accuracy obtained from SVM classifiers on
a Landsat TM image. They reported an accuracy of 75.62% from the SVM,
74.02% from the back propagation neural network, 73.31 % from the decision
tree, and 7l.76% from the maximum likelihood classifiers. Shah et al. (2003)
reported the accuracies attained by SVM, MLC, and back-propagation neural
network classifiers when applied on an AVIRIS image with and without the use
of supervised feature extraction techniques. They used the LSVM (see Chap. 5)
with a linear kernel, polynomial kernel of different degrees, a RBF kernel, and
a sigmoid kernel. They showed that accuracies of 90.9% to 97.2% were obtained
from SVM classifiers while the maximum likelihood and neural network classifiers could achieve accuracies of 59.4% and 6l.9% respectively on the full
dimensionality dataset. They also showed that the classification results from
the full dimensionality dataset were better than the classification of features
obtained using the discriminant analysis feature extraction (DAFE) and the
decision boundary feature extraction (DBFE) techniques (Lee and Landgrebe
1993).
From these limited studies, it can be seen that in all the cases SVM derived
classifications of both multi- or hyperspectral datasets produced the highest
accuracy. However, there are many issues, which need to be further investigated
before SVM classifiers can be implemented at the operational level.
The aim of this chapter is to understand the application of SVMs and the
behavior of associated parameters for the classification of multi and hyperspectral remote sensing data. In the next section, the details of parameters
considered are provided. Section 10.3 describes the remote sensing data used
to produce SVM classification. Experimental set up, results and their analyses
is presented in Sect. 10.4. A summary is provided in Sect. 10.5.