Denoising in hyperspectral images by superpixel based unmixing

Integration of spatial information into spectral processing is well-known to provide significant performance enhancement for tasks such as classification or unmixing. Recently, superpixel segmentation approaches have gained attention in the literature because they integrate spatial information efficiently into the spectral tasks, and have been utilized for enhancing hyperspectral classification and unmixing performances. In this paper, a superpixel based unmixing approach for hyperspectral image denoising is proposed. Denoising by unmixing aims to reconstruct the hyperspectral images using noiseless spectral signatures and their abundances. The methodology proposed in this paper, in turn, utilizes spatial and spectral information together for denoising, by performing unmixing on the superpixels. The proposed approach has a single parameter, and provides enhanced denoising performance with respect to spectral unmixing. The approach is validated through experimental results on both synthetic and real datasets.