Cross-Stage Attention Edge Enhancement and Fourier-Wavelet Transformer Integrated Network for Hyperspectral Image Classification

Hyperspectral image classification (HSIC) is a crucial task in remote sensing. In existing HSIC architectures, convolutional neural networks (CNNs) excel at capturing local information through regional feature representations, while transformers are adept at establishing long-range dependencies with the self-attention mechanism. However, these methods still encounter challenges of imbalanced global-local feature explorations and boundary feature extractions. To address these issues, this study proposes the cross-stage attention edge enhancement and Fourier-wavelet transformer integrated network (CAEEFT Net), which effectively balances global context modeling with local detail preservation and boundary feature extraction for HSIC tasks. Specifically, for spatial feature refinement, three key modules are designed: the cross-stage attention (CSA) module to enable the interaction of features across different stages, thereby strengthening the model's feature representation ability, the global-local attention (GLA) module to jointly enhance global and local features, and the pyramid-stripe attention (PSA) module to capture discriminative edge features. For spectral feature extraction, this paper proposes a spectral Fourier-wavelettrans former (FWT) to integrate the strengths of both global frequency domain patterns and local token-level features. Experimental results on three benchmark datasets demonstrate that CAEEFT Net achieved superior performance compared to state-of-the-art methods, validating the effectiveness of the proposed CAEEFT Net model for HSIC.