S2CIFTNet: Spatial–Spectral Coupling and Interactive Fusion Transformer Network for Hyperspectral Image Classification

Recently, the fusion network model using Convolutional Neural Networks and Transformers has become a research hotspot in Hyperspectral Image Classification. However, these models applied to Hyperspectral Image Classification often adopt a single branch heterogeneous architecture, which only achieve the fusion of the local details of CNNs and the global attention mechanism of Transformers through simple concatenation. Challenges remain to address both the nonlinear relationships between substantial redundant information in local complex features and the generalization ability at a single scale. To address these issues, we propose the Spatial-Spectral Coupling and Interactive Fusion Transformer Network (S2CIFTNet), which enables efficient extraction of multiscale spatial-spectral features, precise modeling of nonlinear feature relationships,and effective fusion of multimodal features. We have designed a module for Spatial-Spectral Fusion Second-Order Distance Covariance (S2FDC), which can perform multiscale spatial-spectral feature extraction and effectively integrate different branch features through second-order distance covariance, accurately captures the complex interdependency structures between variables, thus achieving cross-scale spatial-spectral information fusion. Additionally, we designed an Interactive Attention Fusion (IAF) module that maps spatial-spectral features into feature markers and dynamically integrates these representations to extend the conventional multichannel self-attention mechanism. This extension promotes sufficient interaction and fusion between diverse input information, thereby achieving effective capture of both local details and global context, which significantly enhances classification performance of the model. Experimental results show that on four public datasets, S2CIFTNet achieves excellent classification results with fewer training samples compared to existing methods. The source code link is available at https://github.com/huitailang99/S2CIFTNet.