PMGMCN: A Parallel Dynamic Multihop Graph and Composite Multiscale Convolution Network for Hyperspectral Sparse Unmixing

Abstract

In recent years, sparse unmixing (SU) has garnered significant attention in hyperspectral images (HSI) because it does not require endmember estimation, relying instead on prior spectral libraries to represent observed HSI data, which avoids the influence of endmember extraction on unmixing. However, SU methods based on representation models have limited capability in learning nonlinear features, which results in poor abundances estimation performance in complex environments. Recently, inspired by deep learning, SU models based on neural networks have been proposed to more effectively extract and handle nonlinear features. Nevertheless, the convolution strategies employed in existing SU network models lead to insufficient attention to long-range pixel dependencies, consequently resulting in restricted utilization of spatial priors. In view of the abovementioned shortcomings, this article proposes a parallel dynamic multihop graph and composite multiscale convolution network for SU, referred to as PMGMCN. The network combines the advantages of convolutional neural network (CNN) and graph convolutional network (GCN), achieving a complementary and enhanced integration of their characteristics. Specifically, the network captures long-range spatial features through the designed dynamic multihop graph interaction attention module, which is based on GCN, while the composite multiscale convolution spatial–spectral attention module, which is based on CNN, is designed to extract multiscale spatial–spectral information within local regions. In addition, this article introduces an adaptive weighted total variation loss function based on Sobel edge operator and Gaussian function to encourage piecewise smoothness in abundances maps while preserving edge information. Extensive experiments on synthetic and real datasets demonstrate the effectiveness and superiority of the proposed method compared with the state-of-the-art methods.