An improved N-FINDR algorithm for endmember extraction in hyperspectral imagery

Since recent advances of remote sensing instruments have significantly improved sensor's spectral resolution, pixel resolution is larger than the object size, such that the hyperspectral signal collected by the sensor at each pixel is formed by an integration of signals, which can be considered macroscopically pure, are usually named “endmembers” in hyperspectral image. It is top of all in the hyperspectral analysis that the endmembers should be extracted from the image. N-FINDR algorithm, one of the most popular and effective endmember extraction algorithms, implements with the random initialization of the procedure which brings about the blindfold replacement of the endmembers, and the innumerable volume calculation causes a low speed of the algorithm. However, many published research on N-FINDR algorithm missed the comprehensive consideration of improvement in the two aspects. In this paper, two very typical improvements were applied to integrate the performance using automatic target generation process algorithm (ATGP) algorithm initialized endmember set and the distance calculation to replace the volume calculation in N-FINDR algorithm. The simulate experiment was finally implemented to demonstrate better performance of the hybrid improved N-FINDR algorithm by comparison with original N-FINDR algorithm (ONF), N-FINDR algorithm with initialized endmember set (INF) and N-FINDR algorithm with distance calculation other than volume calculation (DNF) using synthesis hyperspectral image. By comparing experiment results, it is indicated that in contrast to the other three algorithms, the hybrid improved algorithm in this paper shows the best performance that it needs a small amount of the spectrum set replacement and const the least of the procedure time.