20.1 Introduction

Dimensionality reduction (DR) is probably the most common and popular technique that has widely been used in multivariate data analysis to resolve so-called curse of dimensionality (Fukunaga, 1990; Bischop, 1995). Chapter 6 provides a comprehensive description of DRT techniques. Since hyperspectral imaging sensors utilize hundreds of contiguous spectral bands for data acquisition, processing such enormous data volumes becomes a formidable and challenging task for image analysts. To mitigate this dilemma, DR is a feasible solution to reduce the original data space to a relatively low data space to meet practical needs, such as removal of data redundancy, reduction of the expensive computational cost. However, a major issue in DR is determination of the number of dimensions, q, to be retained by DR, where the value of q can be estimated by VD, as shown in various applications by Chang (2006a, 2006b). Nevertheless, it has also been shown in Chapter 5 that VD can adapt and vary with different applications.

Due to the fact that hundreds of spectral dimensions are needed to be dealt with, selecting an appropriate q from such a wide range of values, that is, img, where L is the total number of spectral bands, can be very tricky. This chapter develops and coins a new concept, to be called dimensionality prioritization (DP), to resolve this issue. The motivation of DP arises ...

Get Hyperspectral Data Processing: Algorithm Design and Analysis now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.