Spectral Dynamic Attention Network for Hyperspectral Image Super-Resolution
Dynamic sparse attention and frequency-enhanced feed-forward network improve spatial detail.
Hyperspectral imaging captures rich spectral information but often lacks spatial resolution. Existing deep learning methods struggle with high spectral redundancy and limited non-linear modeling. To address this, a team of researchers from Ocean University of China and Mississippi State University propose SDANet (Spectral Dynamic Attention Network). The framework introduces two key innovations: a Dynamic Channel Sparse Attention (DCSA) module that computes channel-wise correlations and dynamically prunes less informative attention responses, and a Frequency-Enhanced Feed-Forward Network (FE-FFN) that combines spatial and frequency-domain representations to boost non-linear expressiveness.
Experiments on two standard hyperspectral datasets show SDANet outperforms previous state-of-the-art methods in both quantitative metrics and visual quality, while remaining computationally efficient. The adaptive sparsification in DCSA effectively reduces spectral redundancy without sacrificing accuracy. This work, accepted at IEEE GRSL 2026, offers a practical solution for applications like remote sensing, agriculture, and medical imaging where high-resolution hyperspectral data is critical. The authors have confirmed the code will be open-sourced.
- DCSA module uses dynamic data-dependent sparsification to keep only the most informative spectral attention responses.
- FE-FFN jointly models spatial and frequency-domain features, improving non-linear modeling over standard FFNs.
- SDANet achieves state-of-the-art results on two benchmark HSI datasets with competitive computational efficiency.
Why It Matters
Enables sharper hyperspectral images for remote sensing, environmental monitoring, and medical diagnostics.