Hyperspectral Data Processing: Algorithm Design and Analysis

23.1 Introduction

The PBDP (Chapter 21) coupled with the dynamic dimensionality allocation (DDA) (Chapter 22) paves the way for a new approach to band selection (BS), referred to as progressive band selection (PBS) to be studied in this chapter. The idea of PBS emerged from the issue of implementing PBDP for BS where high prioritized bands may also have high correlation. If PBDP is directly used for BS, it is very much likely that if a band is selected for its high priority, other bands highly correlated with this particular band such as its adjacent bands may also have high priorities. As a result, they will also be selected according to their high priority scores. In order to avoid such a situation, these bands must be removed prior to BS. In other words, once a band is selected, all other bands having high correlation with this particular band should not be selected. Using BD to resolve this issue for BS was investigated in Chang et al. (1999), where an information divergence-based band de-correlation approach was developed for this purpose. The PBS proposed in this chapter basically takes advantage of this approach by including BD as a preprocessing step prior to PBDP in which case the latter will not repeatedly select bands that have high interband correlation with bands already selected. Two approaches to BD are developed to effectively remove the interband correlation, spectral measure-based band de-correlation and orthogonalization-based band de-correlation. Since BP prioritizes ...