SUPERPIXEL BASED UNMIXING FOR ENHANCED HYPERSPECTRAL DENOISING

Unmixing based denoising for hyperspectral images is a recent addition to the literature, and aims to reconstruct the data using noise-free and pure spectral signatures and their abundances. Unmixing based denoising has the potential of providing enhanced denoising performance by excluding the noise effects in the endmember and abundance matrices, and also enables selective elimination of effects such as cloudiness, if specific endmembers can be extracted for such effects. This paper proposes a superpixel segmentation and spectral unmixing based denoising and destriping approach for hyperspectral images. Superpixel segmentation enables integrating spatial information into the unmixing process efficiently without complex optimization procedures or a multitude of parameters, as in most spatial preprocessing approaches. The proposed methodology uses a single parameter, and provides fast and enhanced denoising performance with respect to previously proposed unmixing based denoising approaches for hyperspectral images.