Introduction
5
chapter discuss the concepts of radiometric and geometric rectification, image
registration, feature extraction, image enhancement, classification, fusion and
change detection operations. The objective of these sections is to briefly familiarize the reader with the role of each image processing operation in deriving
useful information from the remote sensing data. Each operation has been
clearly explained with the help of illustrative examples based on real image
data, wherever necessary.
Part II: Theory
Having established the background on the basic concepts of data acquisition
and analysis systems, Part II of the book imparts theoretical knowledge on the
advanced techniques described later in this book. In this part, four chapters
are included. Chapter 3 deals with the description of the mutual information
(MI) similarity measure that has origins in information theory. MI is used for
automatic intensity based registration of multi and hyperspectral data. As the
computation of MI depends on the accurate estimation of joint histograms,
two methods of joint histogram estimation have been explained. The existing
joint histogram estimation methods suffer from the problem of interpolationinduced artifacts, which has been categorically highlighted and a new joint
histogram estimation algorithm called generalized partial volume estimation
has been proposed to reduce the effect of artifacts. Some optimization issues
to maximize the MI have also been addressed.
In Chap. 4, details of a technique called independent component analysis
(ICA), originally used in signal processing, have been provided. The theory
of ICA discussed in this chapter forms the basis of its utilization for feature
extraction and classification of hyperspectral images. ICA is a multivariate
data analysis technique that commences with a linear mixture of unknown
independent sources and proceeds to recover them from the original data.
The merit of ICA lies in the fact that it uses higher order statistics unlike
principal component analysis (PCA) that uses only second order statistics to
model the data. After introducing the concept of ICA, several algorithms for
determining the ICA solution have been presented, which is followed with its
implementation in a couple of applications of hyper spectral data processing.
Image classification is perhaps the key image processing operation to retrieve information from remote sensing data for a particular application. The
use of non-parametric classifiers has been advocated since these do not depend on data distribution assumptions. Recently, support vector machines
(SVM) have been proposed for the classification of hyper spectral data. These
are a relatively new generation of techniques for classification and regression
problems and are based on statistical learning theory having its origins in machine learning. Chapter 5 builds the theoretical background ofSVM. A section
is exclusively devoted to a brief description of statistical learning theory. SVM
formulations for three different cases; linearly separable and non-separable
cases, and the non-linear case, have been presented. Since SVM is essentially
a binary classification technique, a number of multi-class methods have also
been described. Optimization, being the key to an efficient implementation of
5
chapter discuss the concepts of radiometric and geometric rectification, image
registration, feature extraction, image enhancement, classification, fusion and
change detection operations. The objective of these sections is to briefly familiarize the reader with the role of each image processing operation in deriving
useful information from the remote sensing data. Each operation has been
clearly explained with the help of illustrative examples based on real image
data, wherever necessary.
Part II: Theory
Having established the background on the basic concepts of data acquisition
and analysis systems, Part II of the book imparts theoretical knowledge on the
advanced techniques described later in this book. In this part, four chapters
are included. Chapter 3 deals with the description of the mutual information
(MI) similarity measure that has origins in information theory. MI is used for
automatic intensity based registration of multi and hyperspectral data. As the
computation of MI depends on the accurate estimation of joint histograms,
two methods of joint histogram estimation have been explained. The existing
joint histogram estimation methods suffer from the problem of interpolationinduced artifacts, which has been categorically highlighted and a new joint
histogram estimation algorithm called generalized partial volume estimation
has been proposed to reduce the effect of artifacts. Some optimization issues
to maximize the MI have also been addressed.
In Chap. 4, details of a technique called independent component analysis
(ICA), originally used in signal processing, have been provided. The theory
of ICA discussed in this chapter forms the basis of its utilization for feature
extraction and classification of hyperspectral images. ICA is a multivariate
data analysis technique that commences with a linear mixture of unknown
independent sources and proceeds to recover them from the original data.
The merit of ICA lies in the fact that it uses higher order statistics unlike
principal component analysis (PCA) that uses only second order statistics to
model the data. After introducing the concept of ICA, several algorithms for
determining the ICA solution have been presented, which is followed with its
implementation in a couple of applications of hyper spectral data processing.
Image classification is perhaps the key image processing operation to retrieve information from remote sensing data for a particular application. The
use of non-parametric classifiers has been advocated since these do not depend on data distribution assumptions. Recently, support vector machines
(SVM) have been proposed for the classification of hyper spectral data. These
are a relatively new generation of techniques for classification and regression
problems and are based on statistical learning theory having its origins in machine learning. Chapter 5 builds the theoretical background ofSVM. A section
is exclusively devoted to a brief description of statistical learning theory. SVM
formulations for three different cases; linearly separable and non-separable
cases, and the non-linear case, have been presented. Since SVM is essentially
a binary classification technique, a number of multi-class methods have also
been described. Optimization, being the key to an efficient implementation of