Multi-hop Graph and Synergistic Attention Convolutional Fusion Network for Hyperspectral Image Classification

Convolutional neural networks (CNNs) have powerful feature extraction capabilities, but they perform poorly in small sample scenarios and have difficulty extracting complex spectral-spatial features. To solve the above problems, this paper proposes a multi-hop graph and synergistic attention convolutional fusion network (MGSACFN) designed for hyperspectral image classification. Specifically, the proposed algorithm comprises two main branches. The first is the multi-scale spectral-spatial synergistic attention fusion subnetwork (MS3AFsN), which establishes multiple scales of features through pooling and unspooling operations, enhancing the representational capability of each scale's features via spectral-spatial synergistic attention (SSSA). The second is the multi-hop graph convolutional reinforcement subnetwork (MGCRsN), which utilizes superpixel blocks derived from superpixel segmentation as graph nodes, aggregates multi-level node contextual information through multi-hop graph convolution, and further enhances feature representation via convolutional reinforcement. Finally, the two sub-networks features are concatenated and fused to obtain discriminative spectral-spatial features. The experimental results on the three benchmark hyperspectral datasets show that the proposed algorithm combines the advantages of both CNNs and graph convolutional networks to achieve overall accuracies up to 96.86%, 98.92% and 99.84% on the Pavia University, Salinas Valley, and Kennedy Space Center datasets, respectively under extremely low training samples. These results outperform other comparative methods.