PLGMamba: A New Progressive Local–Global State-Space Model for Hyperspectral Image Super-Resolution

Convolutional neural networks (CNNs) and Transformer architectures are among the most popular techniques for hyperspectral image super-resolution (HSI-SR). However, both architectures suffer from inherent deficiencies: CNNs are constrained to a limited receptive field, which hinders their ability to capture a wider range of spatial contexts, while Transformers are computationally intensive, making them expensive to train and deploy in large-scale scenes. Furthermore, existing methods tend to employ end-to-end spectral and spatial modeling paradigms for the HSI-SR task while neglecting the natural spectral patterns of high-dimensional HSIs. To overcome the above problems, this paper proposes a new progressive local–global state-space model for HSI-SR, named PLGMamba, which fully takes into account feature extraction and model design from a local to a global perspective. Initially, a progressive learning paradigm to extract spectral sequences (from local to global) is applied to the low-resolution HSI to ensure spatial and spectral fidelity. Then, a residual attention Mamba (RatMamba) is developed to capture the local–global spectral–spatial features of the reconstructed HSI, which integrates spectral–spatial local perception and global modeling into a unified residual learning framework. Finally, a residual Mamba (ResMamba) is designed to fuse progressive local spectral–spatial features and obtain the final high-resolution HSI. Experiments on 3 benchmark HSIs and 1 Gaofen-5 HSI demonstrated that the proposed PLGMamba outperforms classical and state-of-the-art HSI-SR approaches including the mainstream CNN, Transformer, and Mamba. The proposed method is available online: https://github.com/chengle-zhou/PLGMamba .