UNSUPERVISED NONLINEAR UNMIXING OF HYPERSPECTRAL IMAGES USING SPARSITY CONSTRAINED PROBABILISTIC LATENT SEMANTIC ANALYSIS

Unsupervised spectral unmixing (i.e., endmember extraction and abundance estimation) of nonlinear mixture is a very challenging subject in hyperspectral image analysis. In this paper, we present a new interpretation of the reflectance mixture by normalizing the absolute reflectance value into a unit L-1 norm vector, such that the spectral reading can be treated as a probability distribution. The abundance can then be interpreted as the possibility that a spectral distribution belongs to an endmember distribution. Both endmember extraction and abundance estimation can be handled by the proposed sparsity constrained probabilistic latent semantic analysis (SC-pLSA). Experimental results using both synthetic and real data as compared to other unmixing algorithms show apparent advantage of the proposed method.