Hyperspectral Data Processing: Algorithm Design and Analysis

9.4 Experiments

The experiments presented in this section continue to repeat the same experiments as presented in Section 7.5 and Section 8.7 to ensure that a fair comparative analysis on performance evaluation can be conducted among SM-EEAs in Chapter 7, SQ-EEAs in Chapter 8, and ID-EEAs in this chapter where the same three sets of data, synthetic images, Cuprite image, and HYDICE image are also used for experiments. Two categories of ID-EEAs, IED-SQ-EEAS and EIA-EEAs, are evaluated for performance analysis. In the category of IED-SQ-EEAs, IED-VCA, IED-SGA, IED-UFCLS, IED-ATGP, IED-HOS-EEA, and IED-ICA-EEA are considered for evaluation. The category of EIA-EEAs is further divided into two classes, EIA-SQ-EEAs and EIA-SM-EEAs, where four EIAs, IED-ATGP (or ATGP), IED-UFCLS, Maxmin, and ISODATA, are used to initialize two major SM-EEAs, PPI and N-FINDR, and three SQ-EEAs of interest, ATGP-VCA, ATGP-HOS-EEA, and ATGP-ICA-EEA. Several remarks are noteworthy:

1. It should be noted that VCA, HOS-EEAs, and ICA-EEA can be considered as PP-EEAs where projection vectors specified by projection indexes that point to the direction of endmembers are generated successively. These EEAs are SQ-EEAs which start with a random initial projection vector every time they search for a new endmember. Therefore, there are two ways to initialize such PP-EEAs. One is IED-PP-EEA, which only initializes the initial projection vector one at a time. The other is EIA-PP-EAAs, which use an EIA to provide PP-EEAs ...