Hyperspectral Data Processing: Algorithm Design and Analysis

23.2 Band De-Corrleation

The PBDP in Chapter 21 is designed to prioritize all bands in accordance with their contained information measured by a custom-designed band prioritization criterion. It does not consider the case that bands with high priority scores may also be highly correlated. As a consequence, when PBDP is directly applied to BS, some of highly correlated bands may be also selected simply based on their priority scores. Such a dilemma can be avoided by BD that allows PBS to select bands not only with high priorities but also with interband correlation as least as possible. The idea of BD is to consider a band image with a size of pixels as an MN-dimensional band vector by concatenating pixels line by line from top to bottom. With this interpretation, two band images can be represented by two vectors with same dimensionality. The correlation between two band images can now be measured by discrepancy between their corresponding vectors. Two BD criteria, spectral measure-based BD and orthogonalization-based BD, are developed and described in the following sections.

23.2.1 Spectral Measure-Based BD

Since a band image can be represented as a band vector, the correlation of two band images can be interpreted as mutual discriminability of their representing band vectors. Accordingly, all the signature-based hyperspectral measures developed in Section 16.2 can be used for this ...