Hyperspectral Data Processing: Algorithm Design and Analysis

6.1 Introduction

Hyperspectral images are collected by hundreds of contiguous spectral channels, and thus the data volume to be processed is considered to be huge. With such high spectral resolution, spectral correlation among bands is expected to be very high and the band-to-band spectral information may be overlapped or shared in some aspects. To address this issue, two general techniques are commonly used. One is DRT and the other is DRBS.

As for DRT, a general approach is to use component analysis (CA) transforms. It is known that one of most frequently used CA transforms is the principal components analysis (PCA) that makes use of eigenvalues to determine the significance of principal components (PCs) generated by PCA so that DR is accomplished by selecting PCs in accordance with the magnitudes of their associated eigenvalues. Unfortunately, such PCA-based DR (PCA-DR) may not be effective or appropriate for hyperspectral image analysis as demonstrated in Wang and Chang (2006a). A similar approach, called maximum noise fraction (MNF) (Green et al., 1988) or noise-adjusted principal components (NAPC) transform (Lee et al., 1990), which was developed based on signal-to-noise ratio (SNR), also suffers from the same drawbacks as PCA does. One major issue for both PCA and MNF is that many subtle material substances that are uncovered by hyperspectral imaging sensors with very high spectral resolution cannot be characterized by second-order statistics. This may be due to the fact ...