Parallelizing band selection for hyperspectral imagery with many-threads

Abstract

It is time-consuming and expensive for band selection of hyperspectral imagery. In practice, band selection for hyperspectral imagery is a computing-intensive application, in which bands with less information are removed and the maximal information band should be preserved by quantifying information amount based on K-L divergence. Fortunately, it is suitable to parallelize band selection for hyperspectral imagery using accelerator with many-threads. Experimental results validate that band selection with China Accelerator would be much better than CPU and 1.25 X speedups than that of matched GPU.