On the performance of endmember extraction algorithms for hyperspectral image analysis

Abstract

In this paper, we investigate the performance of an endmember extraction algorithm when it is implemented in different fashions. The implementation fashion is changed by the use of a dimensionality reduction process, parallel or sequential mode. This results in four different versions of a single algorithm. We take the Automatic Target Generation Process (ATGP) algorithm as a study case due to its excellent performance. The experimental results show that a dimensionality reduction process can not only reduce computational complexity but also improve performance by compacting useful information into a low-dimensional space; the parallel mode can provide better performance than the sequential mode with the increase of computational complexity. Instructive recommendations in the selection or implementation of endmember extraction algorithms for practical applications are provided.