Dynamic Hypergraph Convolution and Recursive Gated Convolution Fusion Network for Hyperspectral Image Classification.

Recently, convolutional neural network (CNN) and graph convolutional network (GCN) have been used widely for hyperspectral image (HSI) classification which, respectively, specialize in characterizing the local receptive feature and structure feature. However, the existing CNN-based methods cannot learn the higher-order interactions of different spectral bands. The GCN-based methods mostly used the fixed or simple graph model for feature learning. To solve the problems, we propose the dynamic hypergraph convolution and recursive gated convolution fusion network (DHCRGCFN) for HSI classification. To learn the hidden and important relations represented in the HSI data, the dynamic hypergraph convolution network (DHCN) is designed which dynamically updates the hypergraph model and captures the global spatial information of HSI. To efficiently model the high-order interactions among the high spectral dimension, the recursive gated convolution network (RGCN) is developed for progressively capturing the interactions of spectral feature. The features extracted by the two branches are adaptively fused to achieve the complementary advantages. Extensive experiments are conducted on two public HSI datasets to demonstrate the effectiveness of the proposed DHCRGCFN.