Hyperspectral Data Processing: Algorithm Design and Analysis

9.3 Initialization-Driven EEAs

Thus far, we have described two types of EEAs, SM-EEAs in Chapter 7 and SQ-EEAs in Chapter 8, both of which make use of initial endmembers randomly selected from the data. As demonstrated in experiments, the use of random initial endmembers results in inconsistent final selections of endmembers. An ID-EEA for PPI was first developed by Chang and Plaza (2006) to cope with this particular issue and more general issues were later investigated in Plaza and Chang (2006). Due to the natural difference in implementation of an SM-EEA and an SQ-EEA, how to find a specific set of initial endmembers for an SM-EEA and an SQ-EEA is also different. For a given number of endmembers, p, an initial set of endmembers for an SM-EEA needs p endmembers altogether to initialize an SM-EEA. On the other hand, an SQ-EEA produces one endmember at a time sequentially. In this case, an initial set of endmembers used to initialize an SQ-EEA is generally a singleton set which consists of only one endmember instead of p endmembers. As a result, finding an appropriate set of initial endmembers for an SQ-EEA is reduced to looking for a specific data sample that not only can speed up the endmember-searching process but also can produce consistent final results. So, an EEA which only requires the first endmember to be generated for its initial condition is called IED-EEA. When an EEA is an SQ-EEA using a specific data sample as an initial endmember, it is called IED-SQ-EEA. On the ...