A modified Pixel Purity Index method for hyperspectral images

This paper discusses issues with the Pixel Purity Index (PPI) method, which is a currently popular way to find endmembers in hyperspectral images. Due to randomness of PPI, it does not produce an entirely uniform set of directions. Consequently, some directions are favored in the space of pixel vectors, resulting in biased endmember identification. To overcome this difficulty, we propose a new method of construction with non-random uniform directions, which results in a more balanced identification of endmembers. Using a family of artificial examples, we show conditions under which the new method outperforms the classic PPI. In all scenarios, the new method is at least as good as the classic PPI.