Incorporation of spatial and spectral contents in mixed-pixel decomposition of hyperspectral images

Abstract

Spectral unmixing is often involved two important steps. One is the identification of unique constituent elements of hyperspectral image known as Endmembers (EMs) and the other is their abundance fractions estimation. Accurate spectral unmixing has great impact on interpretation of unknown hyperspectral images. Many algorithms were developed to recognize EMs. Most of them are emphasized on exploitation of spectral information with lack of spatial contents support. Hence, several preprocessing modules (PPs) prior EM extraction stages were offered in order to incorporate both spatial and spectral contents. In this paper, we propose a novel preprocessing algorithm by recognizing spatial homogenous areas utilizing both unsupervised k-means clustering and a novel over-segmentation technique. Afterwards, areas with greater spectral purity degree are found thorough homogenous regions as the best EM candidates for the next EE stages. Respect to experimental results done on AVIRIS Cuprite scene, our scheme can improve reconstruction of original image and extract EMs more precisely near their USGS library signatures. Moreover, it provides a significant reduction in computation time of EM identification stage.