Spectral Unmixing Technique Based on Flexibly Selected Endmembers

In traditional spectral unmixing methods of hyperspectral imagery, all endmembers (EMs) take part in the process of each pixel unmixing, and each EM is invariable in the whole processing. This kind of ways discounts unmixing accuracy by introducing some EMs unrelated with mixed pixel to be analyzed, and by the inaccuracy representation of each class by a fixed EM. So the traditional way of spectral unmixing can not get satisfied result. In this case, the paper proposes spectral unmixng ways based on flexibly selected EMs to improve unmixing accuracy. In the first way, partly EMs corresponding to classes of interest are selected to be used in the whole process of spectral unmixing. In the second way, related EMs of a mixed pixel are concluded by utilizing spatial information, then spectral unmixing is done with only related classes considered. In the third way, each class EM is adjusted regionally in each locally divided area. And so, more accuracy EMs are used in spectral unmixing. Experimental results show that spectral unmixing based on flexibly selected EMs can get better performance.