Low-Rank Representation Learning With Spectral-to-Spatial Aggregation Transformer for Hyperspectral Image Super-Resolution

Hyperspectral image (HSI) faces significant challenges in achieving high-resolution (HR) due to the inherent limitations of imaging devices. To tackle this issue, HSI super-resolution (SR) techniques have been introduced to reconstruct HR images from low-resolution (LR) inputs, demonstrating remarkable progress in enhancing spatial fidelity. Notably, the spectral domain of HSIs inherently exhibits low-rank characteristics resulting from strong interband correlations. Inspired by this observation, many approaches have been developed to leverage the low-rank property of HSIs for performance improvement. However, most existing methods rely on linear low-rank formulations, which are often insufficient to capture the complex spectral structures of HSIs. Moreover, approaches relying on singular value decomposition (SVD) or tensor factorization usually require explicit decomposition operations, which may increase optimization complexity. In this article, we propose a nonlinear low-rank representation (LRR) learning strategy combined with a spectral-to-spatial aggregation Transformer for HSI SR tasks. Specifically, the learned nonlinear LRR is embedded into the spatial attention module to exploit both spatial self-similarity and spectral low-rank characteristics, enabling compact and informative representations that preserve fine spatial and spectral details. In parallel, the spectral-to-spatial aggregation Transformer bridges the spectral and spatial domains, enabling more effective integration of spectral priors into spatial feature learning. Furthermore, we incorporate a spectral convolution attention module into our Transformer framework to mitigate the limitation of Transformers in extracting high-frequency information. Experimental results demonstrate that our method effectively reconstructs SR images with sharper edges, finer details, and enhanced spatial-spectral consistency, outperforming existing approaches in both qualitative and quantitative evaluations.