Hyperspectral Data Processing: Algorithm Design and Analysis

18.3 Algorithms Selected to Extract Pixel Information

While our selection of algorithms for performance evaluation may be subjective, it is our desire to make this selection as representative as possible. Nevertheless, such a selection by no means claims to be complete.

In the first category of algorithms, we evaluate three EEAs developed in Chapter 7 for endmember extraction: PPI available in the Research Systems ENVI software, N-FINDR that has widely been used for endmember extraction, and AMEE which is the only algorithm in this category that makes use of spatial information for endmember extraction. As noted earlier, the N-FINDR implemented here is actually IN-FINDR for its iterative advantages. The second category of algorithms consists of algorithms developed for unsupervised target detection, which have been used to generate a posteriori knowledge for applications in supervised target detection and classification. Interestingly, they can also be used for endmember extraction, as demonstrated in Chapter 8. Three algorithms, such as ATGP in Section 8.5.1, UFCLS in Section 8.5.3, and IEA in Section 8.5.4, are of interest. In the third category of algorithms, we look into algorithms developed for anomaly detection. This is due to the fact that occurrence of pure pixels is considered rare. In this case, endmembers behave like anomalies; thus, they can be extracted by anomaly detection algorithms. Since the RX algorithm has widely been used for this purpose and many anomaly detection ...