Robust Deep 3D Convolutional Autoencoder for Hyperspectral Unmixing with Hypergraph Learning

Hyperspectral unmixing aims to acquire pure spectra of distinct substances (endmembers) and fractional abundances from highly mixed pixels. In this paper, a deep unmixing network framework is designed to deal with the noise disturbance. It contains two parts: a three-dimensional convolutional autoencoder (denoising 3D CAE) which recovers data from noised input, and a restrictive non-negative sparse autoencoder (NNSAE) which incorporates a hypergraph regularizer as well as a l2, 1-norm sparsity constraint to improve the unmixing performance. The deep denoising 3D CAE network was constructed for noisy data retrieval, and had strong capacity of extracting the principle and robust local features in spatial and spectral domains efficiently by training with corrupted data. Furthermore, a part-based nonnegative sparse autoencoder with l2, 1-norm penalty was concatenated, and a hypergraph regularizer was designed elaborately to represent similarity of neighboring pixels in spatial dimensions. Comparative experiments were conducted on synthetic and real-world data, which both demonstrate the effectiveness and robustness of the proposed network.