Deep Contrastive Graph Learning Framework for Hyperspectral Image Classification

Hyperspectral Imaging (HSI) captures rich spectral information, but its classification remains challenging due to high dimensionality, scarcity of labeled samples, and class imbalance. These factors exacerbate overfitting, inter-class similarity, and intra-class variability, especially in low-label scenarios. This paper introduces a novel framework that integrates contrastive learning into a Graph Convolutional Network (GCN) within a self-representation learning paradigm. A spectral similarity-based graph is first constructed to model meaningful afnities between pixels. Latent embeddings are then generated via GCN propagation, which indirectly incorporates local spatial context. These embeddings are refined using a joint contrastive objective combining Centroid Metric Loss and Triplet Loss with Semi-Hard Negative Mining to enhance intra-class compactness and inter-class separation. A separate classification module is trained with Cross-Entropy Loss, allowing the model to decouple feature discrimination from label supervision. This dual objective design promotes the emergence of minority-class samples, reduces overfitting, and yields a more structured and balanced embedding space. Extensive experiments on the Indian Pines and Pavia University datasets confirm the effectiveness of the proposed method, which consistently improves classification accuracy and robustness over existing models.