Hyperspectral Data Processing: Algorithm Design and Analysis

16.1 Introduction

A traditional approach to designing hyperspectral measures assumes that data samples are not necessarily collected from imaging sensors as image pixels and could be also from other types of nonimaging optical sensors. In this case, the data samples should be analyzed as one-dimensional signals on the basis of their spectral characteristics (see Chapter 2, Chang 2003a; Category 2: Hyperspectral Signal Processing, Chapters 24–29). Since there is no prior knowledge available regarding the data sample vectors to be analyzed, the target analysis must be performed by some sort of an unsupervised fashion. In certain applications such as chemical/biological (CB) warfare defense there is a spectral library or database that can be used to identify unknown CB agents for target discrimination and target identification where in the former case, unknown data samples can be only discriminated one from another, while in the latter case an unknown data sample can be identified by comparing its spectral profile against the spectral signatures in a database or spectral library. However, technically speaking, such a target identification is actually target verification because it does not perform identification but rather verifies data sample vectors of interest via an existing data base or spectral library. To do so, signature vector-based spectral measures such as spectral angle mapper (SAM) and spectral information divergence (SID) developed in Chapter 2 of Chang (2003a) are generally ...