Saliency-Modulated Group Dynamic Attention for Hyperspectral Image Classification

Hyperspectral image classification (HSIC) remains challenging due to severe spectral redundancy, complex spatial–spectral dependencies, and the scarcity of labeled samples. While convolutional neural networks (CNNs)-based frameworks improve local feature extraction, they struggle to capture long-range spatial–spectral interactions. Vision transformers offer a promising alternative; however, their uniform global self-attention mechanism can lead to attention dilution on dense spatial–spectral token grids, particularly under limited labeled-data conditions. To address these challenges, this article proposes SGDAFormer, a transformer-based framework for HSIC that incorporates the saliency-modulated group dynamic attention (SGDA) mechanism as its core contribution. SGDA partitions tokens into semantically coherent groups, applies intragroup self-attention, and adaptively reweights the group outputs via a learnable saliency gate, thereby suppressing redundant responses and emphasizing discriminative spatial–spectral regions. SGDA mitigates the representational inefficiency of uniform global self-attention and improves learning stability under limited annotation availability. Since SGDA operates on tokenized representations, we adopt a CNN-based tokenizer to produce spatially coherent spatial–spectral tokens and incorporate learnable positional embeddings to preserve relative spatial ordering within each patch. Extensive experiments on five public hyperspectral benchmarks under a strict low-label regime demonstrate that SGDAFormer consistently outperforms state-of-the-art methods, achieving an overall accuracy of 97% on the Pingan dataset.