10.9 Conclusions

An EEA generally requires a set of initial endmembers for algorithm initialization. In doing so a random generator is usually used to produce a set of random initial endmembers for this purpose. Unfortunately, this practice affects the reproducibility of final results. Two approaches are explored to address this issue. One is considered in Chapter 9 where various custom-designed initialization algorithms are developed to produce specific sets of initial conditions that can lead to good results. The other adopts a completely opposite approach discussed in this chapter, by taking advantage of random initial conditions to convert an EEA into a random EEA. In order to materialize this idea, two most popular and widely used EEAs, PPI and N-FINDR, are extended as random PPI (RPPI) and random N-FINDR (RN-FINDR) and the the same treatment can be also applied to other EEAs to derive their random counterparts.

PPI has become a standard technique for endmember extraction due to its popularity and availability in the ENVI software. This chapter investigates several practical issues in implementing PPI. Among them two major issues are particularly severe. One is in-reproducibility caused by the use of randomly generated vectors, called skewers where different sets of the same number of skewers produce different results. Another is the requirement of a visualization tool for users to manually manipulate the final selection of endmembers. As a result, a novice and an experienced ...

Get Hyperspectral Data Processing: Algorithm Design and Analysis now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.