Low rank representation for bilinear abundance estimation problem

In the hyperspectral abundance estimation problem, how to effectively utilize spatial information in the data remains as a challenging task: the mixed pixels within a small neighbourhood are usually correlated and have similar pure endmembers. Recently, the total variation and the joint sparsity priors have been incorporated to exploit this structure for linear mixture model (LMM) and bilinear mixture model (BMM), respectively. However, these structured priors are either too complicated or do not fit well for the spatial data structure. In this paper, we propose a low rank representation model which can better capture the spatial data structure for the BMM. The experimental results demonstrate the effectiveness of the proposed algorithms.