Spectral–Spatial Transformer With Multiscale Convolutional Attention for Hyperspectral Image Classification

Hyperspectral image (HSI) classification plays a vital role in remote sensing by leveraging rich spectral and spatial information for accurate material recognition. However, existing methods, particularly Transformer-based approaches, still face challenges in effectively modeling multiscale spatial–spectral features, preserving local details, and maintaining robustness to noise. To mitigate these limitations, we propose TMCANet, a spectral–spatial Transformer with multiscale convolutional attention, designed to effectively leverage both local and global contextual dependencies for HSI classification. Our design is guided by three core strategies: first, a convolutional feature extraction module, consisting of four convolutional layers, to learn hierarchical spectral multiscale representations and enhance local feature capture; second, a hybrid spatial–spectral cross-fusion attention block, combining local spatial attention, and spectral tokenization attention, to dynamically integrate spatial textures and spectral dependencies, and third a spectral–spatial tokenization transformer equipped with an adaptive cross-layer fusion mechanism to aggregate multilevel features, reduce redundancy, and strengthen global contextual modeling. Furthermore, an enhanced focal loss is adopted to alleviate class imbalance and improve multiclass classification robustness. We validate TMCANet on three benchmark datasets—Indian Pines, Pavia University, and Houston—achieving overall accuracies of 94.91%, 95.49%, and 89.89%, respectively, with performance comparable to or exceeding state-of-the-art baselines. These results demonstrate that the proposed multiscale convolutional attention and hybrid spatial–spectral fusion strategies effectively preserve both local detail and global context, thereby enhancing classification performance and supporting practical applicability in remote sensing.