Joint Supervised-Unsupervised Nonlinear Unmixing of Hyperspectral Images Using Kernel Method

In hyper spectral images pixels are mixtures of spectral components associated to pure materials. Nonlinear unmixing of observed pixels is a challenging task in hyper spectral imagery. In this paper, a joint supervised unsupervised nonlinear unmixing scheme is proposed based on the recent advance of kernel based regression and analysis techniques. The proposed scheme takes advantage of high quality training data from the unsupervised kernel algorithm and fast learning and inference speed of the supervised learning algorithm. Experiments on synthetic and real data show the effectiveness of the proposed method.