A spatial compositional model for linear unmixing

Hyperspectral image unmixing into endmembers and abundances is important in many applications. Despite the plethora of linear mixing models, there is very little previous work on directly incorporating spatial information in the pixel likelihood while simultaneously modeling the uncertainty in the extracted endmembers. In this paper, we propose a spatial compositional model (SCM) for this purpose. In SCM, the model uncertainty of the endmembers can be estimated, while the spatial information in the abundances is directly incorporated in the likelihood via a random variable transformation. This results in a simple and efficient algorithm for both endmember extraction, abundance and uncertainty estimation. The results compared with current state-of-the-art algorithms on real datasets are promising.