Spatial-Spectral Contrastive Graph Neural Network for Few-Shot Hyperspectral Image Classification

Graph Neural Networks (GNNs) have emerged as a promising solution for few-shot hyperspectral image (HSI) classification. However, existing GNN-based approaches face critical limitations in three key aspects: 1) suboptimal graph topology construction due to fixed or heuristic-based edge definitions, 2) inefficient propagation of discriminative node features across heterogeneous regions, and 3) inadequate fusion of spatially correlated and spectrally discriminative patterns. To overcome these challenges, we propose a Spatial-Spectral Contrastive Graph Neural Network (SSCGNN), which introduces three novel components. First, we propose a directed graph reconstruction module that transforms conventional undirected graphs into diversity-aware topology through multi-head self-attention mechanisms. Second, we design a collaborative learning framework where the spatial graph network and spectral graph network mutually enhance feature extraction. Finally, we construct a graph contrastive learning module that augments edge representation learning under limited supervision. By constructing positive/negative edge pairs through spectral-spatial similarity metrics, it regularizes the graph structure to emphasize class-discriminative connections and suppress noisy inter-node relationships. Extensive experiments on four benchmark HSI datasets—University of Pavia (PU), Salinas Valley (SA), WHU-Hi-HanChuan (WHU-HH), and WHU-Hi-LongKou (WHU-HL)—demonstrate SSCGNN’s superiority over state-of-the-art few-shot learning methods. With only 5 labeled samples per class, our model achieves overall accuracy (OA) values of: 93.36% (PU, +6.22% vs. baseline GCN), 98.37% (SA, +4.91% vs. spectral-spatial CNN), 84.68% (WHU-HH, +8.03% vs. meta-learning approaches), 95.87% (WHU-HL, +7.15% vs. graph attention networks).