Spectrally-Spatially Coupled Dynamic Reduction for Efficient Hyperspectral Image Super-Resolution

Hyperspectral image super-resolution faces a fundamental challenge in practical applications: achieving high-fidelity reconstruction on resource-constrained edge devices while maintaining manageable computational complexity. Although recent deep learning-based single-image methods have shown remarkable success, their substantial computational overhead hinders deployment. Meanwhile, emerging architectures like Mamba offer efficient long-range modeling potential but struggle with the inherently high dimensionality of hyperspectral data, leading to prohibitively long sequences and excessive computational loads. Existing lightweight strategies often process spectral and spatial dimensions independently, failing to exploit their intrinsic coupled correlation and resulting in a suboptimal performance-efficiency tradeoff. To address this, this article proposes a spectrally-spatially coupled dynamic lightweight network (S²CD-Net). Its core innovation is a novel paradigm of coupled reduction and independent learning. We first design a dual-stage collaborative dimensionality reduction mechanism, where spectral selection and spatial compression mutually guide and refine each other in a closed loop, enabling intelligent and precise removal of cross-dimensional redundancy. Subsequently, the resulting low-redundancy spectral and spatial sequences are fed into dedicated, lightweight SS2D scanners for independent and in-depth global modeling, ensuring rich feature representation while reducing complexity. Experiments demonstrate that our method requires only minimal parameters and computational cost, yet achieves reconstruction quality comparable to and even surpassing current nonlightweight state-of-the-art methods, establishing a new optimal balance between performance and efficiency. The code is available at https://github.com/XYLGroup/S2CD-Net