CONVOLUTIONAL AND FOCUSED LINEAR TRANSFORMER FOR HYPERSPECTRAL IMAGE CLASSIFICATION

Hyperspectral image (HSI) classification remains a challenging task due to the high-dimensional nature of the data and complex spatial-spectral relationships. Traditional convolutional neural network (CNN)-based methods typically focus on local spatial features, often neglecting long-range dependencies. In contrast, Transformer-based approaches excel at capturing long-range dependencies but struggle to extract local features. Additionally, the computational burden of self-attention, especially when modeling fine-grained relationships within large feature maps, has been a persistent challenge in HSI classification. To address these limitations, we propose a novel approach that integrates CNN and Transformer blocks with a focused linear attention (FLA) mechanism, named Convolutional and Focused Linear Transformer (CFLT) model. The proposed model includes two core components: the FLA-based Transformer and a center-region-enhanced feature fusion module. The FLA-based Transformer performs efficient linear attention across all pixels, capturing long-range dependencies without excessive computational cost. The latter enhances cooperation between CNN and Transformer features by modulating them with center-region information, thus improving both local and global feature extraction. Extensive experiments on real HSI datasets demonstrate that the proposed model achieves a strong balance of high classification accuracy and computational efficiency.