Hyperspectral Data Processing: Algorithm Design and Analysis

9

Initialization-Driven Endmember Extraction Algorithms (ID-EEAs)

One major issue arising in all the endmember extraction algorithms (EEAs) developed in Chapters 7 and 08 is the use of randomly generated initial endmembers to initialize algorithms. Accordingly the final set of selected endmembers are generally not the same and the results are not repeatable. Such inconsistency certainly causes discrepancies during data analysis. Another issue is that EEAs are also sensitive to how initial endmembers are chosen. Unfortunately, very little effort has been devoted to deal with these problems until recent works by Chang and Plaza (2006) and Plaza and Chang (2006). Interestingly, a good initial condition can not only reduce computing time significantly but also produce consistent final results. An appropriately selected set of initial endmembers can make tremendous improvement in performance and computational complexity on the endmember extraction process. To address such initialization issues, this chapter introduces initialization-driven (ID) EEAs (ID-EEAs) where their initial conditions can be selected by a custom-designed process. Two procedures are developed for such a selection of initial endmembers. One procedure is called the initial endmember-driven (IED) initialization that is generally used by SQ-EEAs to produce the first initial endmember. The other procedure introduces a new concept of endmember initialization algorithm (EIA) that is particularly designed to generate an ...