Hyperspectral Data Processing: Algorithm Design and Analysis

9.1 Introduction

As demonstrated by the experiments in Chapters 7 and 08, one common issue in implementation of SM-EEA and SQ-EEA is their use of randomly selected data sample vectors as initial endmembers to initialize an EEA. As a consequence, the final set of selected endmembers by an EEA varies. More specifically, two different sets of random initial endmembers may produce different final sets of endmembers. Such inconsistency results from the nature of randomness caused by the use of random initial endmembers. Interestingly, very little attention was paid to this issue in the past until a recent work reported in Plaza and Chang (2006). Although Berman et al. (2004) also realized the problem of starting points (i.e., initial condition) when their iterated constrained endmember (ICE) was developed, they did not specifically address this issue. Nevertheless, the ICE mitigated this problem by using PPI-generated endmembers as candidate points for their algorithm initialization in which case PPI can be considered as an EIA.

It is known that the ultimate goal of an EEA is to find pure spectrally distinct signatures present in the data. Despite that much effort has been devoted to the design and development of EEAs, it seems that very little has been done in addressing the issue of algorithm initialization which in fact has a significant impact on the final endmembers selected by an EEA. Such an initialization issue is also encountered in vector quantization where an algorithm may ...