Hyperspectral Data Processing: Algorithm Design and Analysis

22.1 Introduction

When DR and BS are performed, it assumes that for a given data set the values of q and are known and fixed during the process regardless of applications. However, in many practical applications this may not be true. Using linear spectral mixture analysis (LSMA) as an example, let be a set of p signatures used by LSMA to perform spectral unmixing. Due to their own spectral characteristics the spectral discriminatory powers of these p signatures should be determined by their spectral distinctions (see Chapter 2 in Chang (2003a)). Accordingly, each individual signature should also require different values of q and for its unmixing. This evidence has been witnessed in Figures 20.1, 20.2, 20.6, and 21.8–21.10. To address and resolve this issue, the parameters q and must be made variables to adapt dynamically instead of being fixed at constants. While such a thought is very desirable, how can it be really done? It is shown in source coding that to achieve optimal coding performance, variable-length coding instead of fixed-length coding must be used due to variable probabilities ...