An adaptive graph fusion network for small-sample hyperspectral image classification

Hyperspectral image (HSI) classification under small-sample conditions remains a fundamental challenge in remote sensing. Reliable pixel-level annotations are prohibitively costly to obtain, while land-cover categories often exhibit substantial spectral overlap, spatially intermingled distributions, and high intra-class variability. Under such limited supervision, classifiers relying on a single feature representation or a fixed relational topology are prone to unstable decision boundaries and error-amplifying feature propagation. To address these issues, this paper proposes an Adaptive Graph Fusion Network (AGFNet) for small-sample HSI classification. The proposed framework first extracts complementary spatial, spectral, and contextual representations from each HSI patch and constructs branch-specific sparse graphs to refine sample relationships in the corresponding feature spaces. A reliability-gated cross-branch interaction module is then introduced to facilitate the exchange of complementary information across branches. Subsequently, a sample-aware fusion module adaptively weights branch contributions for each input sample. In this manner, feature extraction, graph refinement, cross-branch interaction, and adaptive fusion are jointly optimized within a unified end-to-end framework. Experiments are conducted on three representative HSI datasets using a consistent fixed-sample protocol, encompassing class-imbalanced agricultural scenes, fine-grained crop categories, and complex urban land-cover patterns. The results demonstrate that AGFNet achieves competitive or superior performance compared with recent CNN-, Transformer-, and graph-based methods while producing more spatially coherent classification maps. Ablation experiments, branch-weight analyses, and hyperparameter sensitivity analyses further validate the effectiveness, interpretability, and robustness of the proposed framework under limited-label conditions.