Total-variation-regularized local spectral unmixing for hyperspectral image super-resolution

Fusing a low-resolution Hyperspectral Image (HSI)with its corresponding high-resolution Multispectral Image (MSI) to obtain a high-resolution HSI is amajortechnique for capturing comprehensive scene information in both spatial and spectral domains. To exploit fully the spectral and spatial information of an image, an algorithm based on total-variation-regularized local spectral unmixing for HSI super-resolution was proposed in this study. Spectral features and corresponding spatial information were extracted from both HSIs and MSIs through coupled encode-decode networks, respectively. The decoder of the coupled network could effectively preserve spectral features, and regular terms integrating local low-rank and vector total variation constraints could make full use of spatial structure information in MSIs to extract a stable abundance matrix. Finally, the angular differences between representations were minimized to reduce the spectral distortion. Experimental results reveal that the reconstruction errors in CAVE and Harvard datasets reach 3.78 and 1.66, respectively, and the spectral angle maps are 6.57 and 3.03, respectively, thus outperforming the state-of-the-art methods. The proposed algorithm can make full use of the spatial properties and thus produces a better HIS super-resolution effect. © 2019, Science Press. All right reserved.