A linear hyperspectral unmixing method by means of autoencoder networks

Abstract

ABSTRACT Hyperspectral unmixing technique is a conventional approach addressing the mixed pixel issue. In this paper, we present an autoencoder (AE) network that deals with estimating the abundances in hyperspectral images (HSIs) given the endmembers. In the suggested network, the mixed pixel issue in a supervised scenario is investigated since the weights of the decoder are set equal to the endmembers. More importantly, the network is trained by a blend of two celebrated objective functions, mean squared error and spectral angle distance, in order to have both privileges of sensitivity to small errors and being scale invariant. To assure the convenience, the sparsity and physical constraints are imposed on the abundances, and the regularization techniques are employed. The abundances are initialized via the fully constrained least squares method thanks to the setting of the initial encoder weights. The superiority of the presented AE is demonstrated via conducting several experiments on synthetic and real HSIs and comparing the results quantitatively and visually with several existing methods.