Multi-scale feature enhancement and cross-dimensional attention mechanism fusion network for hyperspectral image classification

Abstract

Hyperspectral image (HSI) classification is widely used in various fields but faces significant challenges such as high dimensionality, spatial complexity, and spectral redundancy, which limit classification accuracy. Traditional machine learning methods rely on handcrafted features and struggle with high-dimensional data, while deep learning approaches still encounter difficulties in multi-level feature fusion, global-local collaborative modeling, and efficient cross-dimensional interaction. To address these challenges, this paper proposes the multi-scale feature enhancement and cross-dimensional attention mechanism fusion network (MSFE-CAMF), which integrates three key modules: a multi-scale feature extraction module (MS-FEM), a large-kernel attention and local feature fusion module (LKALFFM), and a cross-dimensional attention module (CDAM). First, the MS-FEM employs a multi-branch 3D convolutional architecture to capture multi-scale spatial dependencies and spectral correlations while maintaining computational efficiency. Additionally, a residual connection mechanism is incorporated to enhance model stability and convergence. Second, the LKALFFM combines large-kernel attention with local feature enhancement, facilitating cross-scale information capture through multi-scale learning. At the same time, it strengthens fine-grained feature sensitivity via local feature fusion, enabling a more refined representation of hyperspectral data. Finally, the CDAM module integrates a cross-dimensional attention mechanism with multi-pooling channel gating (MCG) to enhance multi-dimensional feature modeling of HSI through spatial-channel information interaction, improving classification accuracy and adaptability in complex scenarios. Extensive evaluations on four popular HSI datasets show that, with 10% of the training samples, the proposed method achieves 99.61% overall accuracy on the Houston 2013 dataset, 99.96% on the Salinas dataset, 99.92% on the WHU-Hi-LongKou dataset, and 99.90% on the WHU-Hi-HanChuan dataset. These results demonstrate the competitive performance of our architecture compared to state-of-the-art methods, underscoring its effectiveness and robustness in HSI classification.