KMNET for Hyperspectral Unmixing

Sparse Autoencoder Networks are recently used for Endmember Extraction and Hyperspectral Unmixing. Performing Spectral clustering before autoencoders provides meaningful information about the endmembers. In this direction, this article presents an end-to-end autoencoder framework, KMNET, that takes the initial end members by clustering the hyperspectral data in spectral dimension. The efficiency of the proposed architecture showed superior performance for unmixing with its loss function that consists of a Kullback-Leibler divergence term with spectral angle distance (SAD) and additional penalty terms. Experimental results on two publicly available datasets including Jasper Ridge and Samson support the superior performance of KMNET compared with other classical and recent approaches for spectral unmixing and endmember extraction.