Hyperspectral Data Processing: Algorithm Design and Analysis

7.6 Conclusions

SM-EEAs are always desirable for endmember extraction. Unfortunately, a genuine SM-EEA is generally impractical because of its very high computational cost resulting from an exhaustive search, specifically when the number of endmembers, p, is large and data volume is huge. In addition, an SM EEA also suffers from several drawbacks: (1) requirement of precise knowledge about p, which is practically unknown; (2) assumption of endmembers present in the data, which is generally not true in many real applications; (3) necessity of dimensionality reduction (DR) due to enormous data volume, in which case selecting an effective DR transform is crucial; and (4) inconsistent results caused by the use of randomly generated initial endmembers. While some drawbacks such as p that can be estimated by the virtual dimensionality in Chapter 5, some other drawbacks such as computational complexity inheriting from the algorithm design. Table 7.1 summarizes design criteria of SM-EEAs discussed in this chapter and their drawbacks and disadvantages.

Table 7.1 Summary of design criteria of SM-EEAs and their drawbacks and disadvantages.

SM-EEAs	Design criteria	Drawbacks/disadvantages
MATLAB-PPI	Orthogonal projection	1. Determination of the number of skewers, K 2. Determination of a threshold for PPI counts, t 3. Determination of the number of dimensions, q, for DR 4. Inconsistent results due to the use of random conditions 5. Many falsely extracted endmembers
SM N-FINDR, IN-FINDR, ...