Elastic constraints on split hierarchical abundances for blind hyperspectral unmixing

The applications of Hyperspectral Image (HI) are limited for the existence of the ”mixed” pixels. The Blind spectral unmixing (BSU) aims to capture the spectral signatures and extract the corresponding fractional abundance maps from the HI. The existing unmixing approaches do not well concurrently consider the structure of the local patches inside each abundance map and the diversity of different endmember signatures, which could deteriorate the performance of the subsequent unmixing. In this paper, we advocate an elastic constrained split abundances method for BSU. It does not need to know the statistical distribution of the HIs. To capture and seamlessly incorporate both the homogeneous information and distinguishable knowledge across different modalities, the divergence among the different endmembers is maximized, and each endmember signature is projected into a common semantic space, furthermore, each abundance map is differentiated into a consensus part and diverse local patches. Extensive experiments are implemented on synthetic and real HIs, and the vigorous experimental results validate the effectiveness of the proposed model and algorithm.