Superpixel-based local collaborative sparse unmixing for hyperspectral image

Abstract

Hyperspectral sparse unmixing aims to identify the best subset of endmembers provided in previously available libraries (which may be very large) and their corresponding proportions (fractional abundance). A sparse unmixing method, local collaborative sparse unmixing (LCSU), including spatial information in standard collaborative formulations, has produced better unmixing results. However, the local window size taken in this method can always have an effect on unmixing result. In detail, LCSU tends to be affected by the distribution of ground objects as well as collaborative sparse unmixing. To address this limitation, we introduce a robust strategy to enforce the local collaborative property in homogeneous areas. The proposed method first divides the denoised image into superpixels according to the quarternion theory, which involves spatial information, and then implements collaborative sparse unmixing in each superpixel. Compared with the most advanced sparse unmixing methods, experiments on two different synthetic datasets and a real hyperspectral dataset effectively verify the advantages of our proposed method.