Sparse filtering based hyperspectral unmixing

This work proposes a hyperspectral unmixing technique based on sparse filtering approach. The proposed method exploits the sparsity of feature distribution rather than modeling the data distribution. The proposed sparse filtering based unmixing procedure is essentially parameter-free, and the only parameter is to find the number of endmembers to be extracted. This approach is a blind unmixing approach because it does not require prior knowledge of endmember matrix. Experimental results on two real hyperspectral datasets demonstrate that the proposed sparse filtering procedure provide better abundance maps compared to nonnegative matrix factorization based approach.