Improved Transformer Network Based on Multiscale Grouping Feature Fusion for Hyperspectral Image Classification

In recent years, hyperspectral image classification methods based on convolutional neural networks and Transformer architectures have achieved remarkable success. However, existing Transformer-based models exhibit several limitations, including insufficient capability in modeling local spectral structures, difficulty in capturing complex correlations between nonadjacent spectral bands, and excessive model parameter counts, which hinder their deployment in resource-constrained environments. To address these issues, this article proposes a novel model, multiscale grouped feature fusion vision transformer (MGFF-ViT), designed to enhance spectral feature representation while maintaining parameter efficiency. First, we introduce a new spectral aggregation module, spectral odd–even fusion, which reshapes spectral vectors into highly correlated spectral matrices. This allows the model to extract both local spectral features and long-range dependencies across frequency bands. Second, we propose a multiscale grouped weight attention mechanism that captures spatial contextual information at multiple scales. By applying a sorting-based strategy—“strong to strong, weak to weak”—the module strengthens relevant feature interactions while suppressing noise and redundancy. In addition to improving classification accuracy, MGFF-ViT significantly reduces the number of model parameters, making it lightweight and efficient for practical applications. Extensive experiments on four publicly available hyperspectral datasets demonstrate that MGFF-ViT achieves competitive or superior performance compared to state-of-the-art methods, even with limited training data.