Hyperspectral Data Processing: Algorithm Design and Analysis

27.1 Introduction

Hyperspectral data are collected by hundreds of contiguous and highly correlated spectral bands. Consequently, the same spectral bands used to acquire two different hyperspectral signatures may not provide the same level of signature information. Furthermore, recent advances in sensor technology have made it possible for sensor data to be acquired by more than hundreds or thousands of spectral channels, for example, hyperspectral or ultraspectral data. Of particular interest is chemical/biological (CB) defense for bioterrorism where CB data available for analysis are generally spectral data rather than image data. However, it also comes at a price that such wealthy spectral information is highly correlated. As a result, using all the hundreds or thousands of spectral channels might not be a good choice for preserving spectral information since a significant and crucial piece of information of interest may only be provided by a very narrow range of spectral coverage and could be overwhelmed by other dominated spectral channels. For example, the crucial information of chemical data is provided by the thermal range, and biological data are determined by their distinct protein spectral profiles, which are usually very small and can only be captured by very narrow diagnostic spectral channels. Therefore, the information provided by their spectral profiles in signature characterization becomes vital, and band selection (BS)-based spectral signature analysis and characterization ...