18.1 Introduction

In traditional two-dimensional (2D) image processing, an image pixel is specified by its intensity and represented by a single value of the gray scale. In hyperspectral image processing, a hyperspectral image is an image cube formed by stacking 2D spectral images acquired by a range of hundreds of spectral channels where a hyperspectral image pixel is actually a column vector, of which each vector component is an image pixel acquired by a specific wavelength. To simplify our discussion, the term “pixel” will be used instead of “pixel vector.” Therefore, one great challenge in hyperspectral data exploitation is analysis of information extracted from a hyperspectral image pixel specified by hundreds of spectral channels. However, how much pixel information can be extracted is also determined by what algorithm to be used for information extraction, such as algorithms in PART II (Chapters 7–11) developed for endmember extraction and algorithms in PART III (Chapters 12–17) developed for target detection and classification. In other words, what we are interested in is, “Does an algorithm really do what it is designed for?” For example, in endmember extraction, do pixel purity index (PPI) in Section 7.2.1 and N-finder algorithm (N-FINDR) in Section 7.2.4 really extract pure signatures as they are designed for? In unsupervised target detection, do algorithms, such as automatic target generation process (ATGP) developed by Ren and Chang (2003) in Section 8.5.1 and unsupervised fully constrained least-squares (UFCLS) algorithm developed by Heinz and Chang (2001) in Section 8.5.3, really perform what they are asked for? In anomaly detection, does the RX algorithm developed by Reed and Yu (1990) really find anomalies or something else? Specifically, “what pixel information does an exploitation algorithm really extract?” and “does it really do what it claims to do?” It turns out that answers to these questions are more complicated than what we had thought, as will be demonstrated by experiments in this chapter. In order to facilitate our discussions, four different types of pixels in accordance with their spatial or spectral properties, pure pixel, mixed pixel, homogenous pixel, and anomalous pixel, are introduced for pixel information analysis.