6
Pramod K. Varshney, Manoj K. Arora
SVM has also been given due consideration and a separate section is written
to discuss the merits and demerits of existing optimization methods that have
been used in SVM classification.
Chapter 6 provides the theoretical setting of Markov random field (MRF)
models that have been used by statistical physicists to explain various phenomena occurring among neighboring particles because of their ability to
describe local interactions between them. The concept of MRF model suits
image analysis because many image properties, such as texture, depend highly
on the information obtained from the intensity values of neighboring pixels,
as these are known to be highly correlated. As a result of this, MRF models
have been found useful in image classification, fusion and change detection
applications. A section in this chapter is devoted to a detailed discussion of
MRF and its equivalent form (i.e. Gibbs fields). Some approaches for the use
of MRF modeling are explained. Several widely used optimization methods
including simulated annealing are also introduced and discussed.
Part III: Applications
The theoretical background, concepts and knowledge introduced in the second part of the book are exploited to develop processing techniques for different applications of hyperspectral data in this part. In Chap. 7, MI based
registration has been applied for automatic registration of a variety of multi
and hyperspectral images at different spatial resolutions. Two different cases
namely multi-sensor and multi-temporal registration, have been considered.
The performance of various interpolation algorithms has been evaluated using
registration consistency, which is a measure to assess the quality of registration
in the absence of ground control points.
Since, the hyperspectral data are obtained in hundreds of bands, for many
applications, it may be inefficient and undesirable to utilize the data from
all the bands thereby increasing the computational time and cost of analysis. Hence, it may be essential to choose the most effective features that are
sufficient for extracting the desired information and efficient in reducing the
computational time. Chapter 8 focuses on the use ofICA for feature extraction
from hyperspectral data. After clarifying the distinction between PCA and
ICA, the details of two ICA based algorithms for the extraction of features have
been given. The aim of these algorithms is to identify those features, which
allow us to discriminate different classes that are hidden in the hyperspectral
data, with high degree of accuracy. A method called spectral screening has also
been proposed to increase the computational speed of the ICA based feature
extraction algorithm.
In Chap. 9, the concept of ICA is further advanced to develop an ICA mixture model (ICAMM) for unsupervised classification of hyperspectral data.
The advantage of using ICAMM lies in the fact that it can accurately map
non-Gaussian classes unlike other conventional statistical unsupervised classification algorithms that require that the classes be Gaussian. The ICAMM
finds independent components and the mixing matrix for each class using
the extended infomax learning algorithm and computes the class member-
Pramod K. Varshney, Manoj K. Arora
SVM has also been given due consideration and a separate section is written
to discuss the merits and demerits of existing optimization methods that have
been used in SVM classification.
Chapter 6 provides the theoretical setting of Markov random field (MRF)
models that have been used by statistical physicists to explain various phenomena occurring among neighboring particles because of their ability to
describe local interactions between them. The concept of MRF model suits
image analysis because many image properties, such as texture, depend highly
on the information obtained from the intensity values of neighboring pixels,
as these are known to be highly correlated. As a result of this, MRF models
have been found useful in image classification, fusion and change detection
applications. A section in this chapter is devoted to a detailed discussion of
MRF and its equivalent form (i.e. Gibbs fields). Some approaches for the use
of MRF modeling are explained. Several widely used optimization methods
including simulated annealing are also introduced and discussed.
Part III: Applications
The theoretical background, concepts and knowledge introduced in the second part of the book are exploited to develop processing techniques for different applications of hyperspectral data in this part. In Chap. 7, MI based
registration has been applied for automatic registration of a variety of multi
and hyperspectral images at different spatial resolutions. Two different cases
namely multi-sensor and multi-temporal registration, have been considered.
The performance of various interpolation algorithms has been evaluated using
registration consistency, which is a measure to assess the quality of registration
in the absence of ground control points.
Since, the hyperspectral data are obtained in hundreds of bands, for many
applications, it may be inefficient and undesirable to utilize the data from
all the bands thereby increasing the computational time and cost of analysis. Hence, it may be essential to choose the most effective features that are
sufficient for extracting the desired information and efficient in reducing the
computational time. Chapter 8 focuses on the use ofICA for feature extraction
from hyperspectral data. After clarifying the distinction between PCA and
ICA, the details of two ICA based algorithms for the extraction of features have
been given. The aim of these algorithms is to identify those features, which
allow us to discriminate different classes that are hidden in the hyperspectral
data, with high degree of accuracy. A method called spectral screening has also
been proposed to increase the computational speed of the ICA based feature
extraction algorithm.
In Chap. 9, the concept of ICA is further advanced to develop an ICA mixture model (ICAMM) for unsupervised classification of hyperspectral data.
The advantage of using ICAMM lies in the fact that it can accurately map
non-Gaussian classes unlike other conventional statistical unsupervised classification algorithms that require that the classes be Gaussian. The ICAMM
finds independent components and the mixing matrix for each class using
the extended infomax learning algorithm and computes the class member-