An improved algorithm of hyperspectral image endmember extraction using projection pursuit

Abstract

Endmember extraction is one of the most important procedures in linear unmixing approach. In this paper, an improved projection pursuit-based endmember extraction algorithm is proposed to extract endmember through extracting non-Gussian structure of hyperspectral image data. Principal component analysis is used not only for removing correlation but also used to reduce dimension and noise in our approach. Procedure of removing "uninteresting" projections is developed to be more automatic. In order to evaluate the effectiveness of the improved approach, simulation data composed by spectrums from SPLIB04b mineral spectrum library offered by USGS is used in experiment. Simulation experiment result shows feasibility of its application in endmember extraction. And then, the algorithm is applied to mineral detection, which proves its effectiveness in automatic mineral endmember detection.