Snapshot Multispectral Image Completion and Unmixing with Total Variation Regularization on Abundance Maps

Keisuke Ozawa, Shinichi Sumiyoshi, Yuki Tachioka

Abstract: Unmixing is an important application of spectral imaging, and snapshot sensors could enrich its applicability. How-ever, their spatio-spectral tradeoff decreases spatial resolution as the number of bands increases. While basis spectra can be estimated even on the downsampled multispectral image, it is difficult to retain high-resolution abundance maps. In this paper, we propose a high spatial resolution unmixing method from a single snapshot multispectral image. The proposed method simultaneously completes a snapshot data to restore the full sensor size multispectral image. In a simulation, we show a resolution-enhanced unmixing and better completion accuracy compared with state-of-the-art tensor completion methods. We also demonstrate against real data the best quality for completion and unmixing in the full sensor size.