Enhancing Endmember Extraction using K-means clustering and Pixel Purity Index

Hyperspectral images are of hundreds of bands and contain abundant information. The mixing of pixels in spatial domain makes differentiation of materials a critical task in hyperspectral images. The different materials are classified as endmembers and their area covered known as abundance maps. The existing unmixing techniques are dependent on random initialization of endmember locations and processed on full band data. This paper proposes a k-means clustering based purity index value on principal components to select the endmember candidates for initialization. The proposed strategy is tested on more efficient Vertex Component Analysis and NFINDR endmember extraction algorithms. The proposed strategy evaluated on Jasper Ridge and Urban dataset. The results were analyzed using root mean square error.