Joint spectral unmixing and clustering for identifying homogeneous regions in hyperspectral images

Abstract

In this paper a joint spectral unmixing and clustering approach for the identification of homogeneous regions in hyperspectral images is proposed. The endmembers required in the unmixing stage are manually selected based on the most significant principal components of the image at hand. Each pixel is decomposed as a linear combination of the endmembers and is represented by the vector of the coefficients of its corresponding linear combination. The clustering stage utilizes the complete-link hierarchical agglomerative clustering algorithm in a layer-wise fashion in order to retrieve the optimal clusters, based on the latter pixels representation. Experiments conducted on real images support the high-quality performance of the proposed approach.