Wavelet-Enhanced Spatial–Spectral Prior Injection Mamba Network for Hyperspectral Image Super-Resolution

Deep learning-based methods, particularly Mamba-based models, hold huge potential for hyperspectral image super-resolution (HSI-SR) reconstruction. However, the intrinsic spatial-spectral details in hyperspectral images are hard to be completely represented on insufficient data solely utilizing the hierarchical learning manner in most Mamba structures, which results in the degradation of restoration quality on inherent fine-grain textures. To address this challenge, this paper proposes a Wavelet-Enhanced Spatial-Spectral Prior Guided Mamba Network (WESSP-Mamba) for HSI-SR, which leverages wavelet-enhanced spatial-spectral priors to refine detail recovery with cross-layer embedment. Specifically, this model strengthens multiscale spatial correlations with a Wavelet transform and produces spectral responses with a spectral attention mechanism, which adaptively highlights specific bands to acquire an instructive spatial-spectral joint prior. The learnable prior is then injected into the hierarchical deep features from the backbone network—built upon a Shuffle-ReShuffle Mamba (SRM) module—through a spatial selection mechanism, thereby contributing to the impressive representation capability. Extensive experiments on three representative datasets show that the WESSP-Mamba model achieves superior HSI-SR performance to other deep learning approaches in terms of quantitative and qualitative metrics. Especially, the peak signal-to-noise ratio of WESSP-Mamba exceeds that of the second-best method by 0.0608dB, 0.0520dB, 0.0443dB on the Chikusei, Houston and PaviaC datasets at the ×4 scale factor, respectively.