A Modified Huber Nonnegative Matrix Factorization Algorithm for Hyperspectral Unmixing

Hyperspectral unmixing (HU) has been one of the most challenging tasks in hyperspectral image research. Recently, nonnegative matrix factorization (NMF) shown its superiority due to flexible modeling and little prior requirement. But NMF algorithms tend use least square function as objective, which is sensitive outliers different kinds noise. In this article, we propose a modified Huber (mHuber) model achieve robustness Under robust model, accelerate half-quadratic optimization algorithm by replacing multiplicative updating rule with projected nonlinear conjugated gradient rule, achieves much faster convergence rate. Moreover, new tuning parameter, rather than fixed one, given adapt mHuber loss function. Finally, perform analysis experiments synthetic real-world datasets, confirms effectiveness proposed method when compared several state-of-the-art methods HU.