GAST: A graph-augmented spectral–spatial transformer with adaptive gated fusion for small-sample hyperspectral image classification

Accurate hyperspectral image (HSI) classification under scarce labels and class imbalance requires models that couple long-range spectral reasoning with irregular local spatial context. We present GAST, a Graph-Augmented Spectral-Spatial Transformer with Adaptive Gated Fusion for Small-Sample Hyperspectral Image Classification. GAST pairs a lightweight spectral Transformer with a GATv2-based spatial branch on an 8-neighbour pixel graph, and fuses them via a center-conditioned, channel-wise gating mechanism that uses the centre-pixel representation to modulate all tokens in the patch. Unlike conventional static fusion strategies (e.g., concatenation or summation) that assign fixed importance to modalities regardless of image content, this adaptive fusion dynamically modulates the spectral and spatial streams at the pixel level, allowing the model to prioritise spatial texture for complex urban structures while shifting focus to spectral signatures for subtle vegetation classes. Training is further stabilized by an imbalance-aware objective that switches between weighted cross-entropy and focal loss according to a measured class ratio, and by a two-stage Bayesian hyperparameter search that aligns capacity with scene statistics. Across eight public benchmarks under a 5%-label protocol, GAST consistently matches or surpasses recent hybrid graph-Transformer architectures while remaining compact and fast at inference. Ablation studies confirm the complementary roles of both branches and the benefit of gated fusion. The resulting architecture offers a strong accuracy-efficiency trade-off and reliable performance across seeds, making it a practical solution for low-data HSI applications. The code is publicly available at https://github.com/fesihkeskin/GAST . • Introduces GAST, a dual-branch spectral–spatial architecture that couples a lightweight spectral Transformer with a GATv2-based spatial module and centre-conditioned channel-wise gating. • Achieves state-of-the-art accuracy on eight benchmarks using only 5% labelled pixels per class, with standard deviations below 0.6 pp across five seeds. • Offers a strong accuracy-efficiency trade-off, with up to 5 × fewer parameters and 2 × faster inference than other hybrid Transformer-GNN models. • Mitigates class imbalance via an adaptive loss that switches between focal and weighted cross-entropy depending on the dataset imbalance ratio. • Provides a fully reproducible PyTorch/Optuna pipeline, with code and trained models released publicly.