Learning Spectral–Spatial-Former Deep Prior for Hyperspectral Image Superresolution

The superresolution (SR) technique is a leading solution for achieving high spatial–spectral resolution in hyperspectral (HS) images, which current sensors struggle to provide due to cost and physical constraints. This study presents a multistage optimization framework that leverages high- and low-frequency components, along with a quadratic splitting method, to address the SR problem. Traditional model-based approaches often use shallow architectures with limited generalization. To overcome this, we integrated our model into a deep convolutional neural network enhanced by a Transformer module for regularization. Although the Transformer’s capabilities are noteworthy, it can improve in capturing local self-similarity and spectral correlations. Furthermore, these models frequently overlook the importance of multiscale and short-range information. Therefore, we introduce a multiscale architecture that allows Transformers to better capture short- and long-range dependencies. By implementing multiscale spatial-aware and multidepth channel-aware modules, we generate comprehensive deep spatial–spectral prior feature maps. The spatial branch focuses on using local–global prior features for HS image reconstruction, while the spectral branch emphasizes the most informative channels and their correlations. Experiments demonstrate that our method significantly outperforms state-of-the-art fusion-based SR techniques in terms of efficiency.