Bridging 3D and 2D Convolution for Hyperspectral Images with Cross-dimensional Spectral Attention

Hyperspectral imaging (HSI) captures reflectance characteristics across hundreds of narrow spectral bands, enabling the joint acquisition of fine-grained spatial details and continuous spectral information. However, the inherently high dimensionality of hyperspectral data presents substantial challenges for feature extraction, necessitating effective modeling of complex spatial-spectral interactions while alleviating overfitting and computational inefficiency. To address these challenges, we propose the Hybrid-Dimensional Hyperspectral Network (HDHN), a novel framework that achieves a balance between efficiency and performance. HDHN employs a 3D encoder with SSIR blocks to capture spatial-spectral representations and a 2D decoder with SIR blocks to abstract semantic features and reconstruct spatial resolution. This design not only delivers strong performance but also reduces computational overhead and mitigates overfitting, offering a practical solution for hyperspectral image analysis. Moreover, we introduce the Cross-Dimensional Spectral Attention (CDSA) mechanism, which adaptively refines spectral representations at the pixel level. By dynamically allocating attention weights, CDSA emphasizes informative spectral bands, thereby enhancing spatial-spectral feature interactions and facilitating efficient encoder-decoder fusion. This leads to notable improvements in both feature representation and overall model performance. Additionally, we present HDHN-e, a streamlined variant optimized for computational efficiency. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art approaches on hyperspectral image classification and salient object detection tasks, while maintaining superior efficiency.