[Endmember selection algorithm based on linear least square support vector machines]

Endmember (EM) selection is an important prerequisite task for mixed spectral analysis of hyperspectral imagery. In all kinds of EM selection methods, N-FINDR has been a popular one for its full automation and efficient performance. Unfortunately, the implementation of the algorithm needs dimensional reduction in original data, and the algorithm includes innumerable volume calculation. This leads to a low speed of the algorithm and so becomes a limitation to its applications. In the present paper, an improved N-FINDR algorithm was proposed based on linear least square support vector machines (LLSSVM), which is free of dimensional reduction and makes use of distance measure instead of volume evaluation to speed up the algorithm. Additionally, it was also proposed to endow the algorithm with robustness by controlling outliers. Experiments show that the computational load for EM selection using the improved N-FINDR algorithm based on LLSSVM was decreased greatly, and the selection effectiveness and the speed of the proposed algorithm were further improved by outlier removal and the pixel pre-sorting method respectively.