Spectral-Aware Lightweight Mamba Network with Multi-Kernel Fusion for Hyperspectral Image Super-Resolution

Hyperspectral image super-resolution (HSI-SR) has been widely applied in ground object classification, target detection, and environmental monitoring. Its core is to reconstruct high-quality spatial–spectral information from low-resolution observations. However, existing models generally suffer from severe redundancy and high computational costs, which limit their practical deployment on resource-constrained devices. Therefore, we propose a Spectral-Aware Lightweight Mamba Network (SALMNet). In general, SALMNet consists of two stages, namely Group-wise Spectral Reconstruction (GSR) and Multi-Kernel Bidirectional Fusion (MKBF). Firstly, in the GSR stage, a joint representation module based on SA-MobileMamba is constructed to significantly reduce the computational complexity of visual state space modeling (VSSM) while maintaining the ability to capture long-range spectral–spatial relationships. Subsequently, In order to significantly reduce the amount of model parameters and maintain the fusion ability of cross-scale features, we designed a lightweight multi-level encoder–decoder and a lightweight bidirectional feature fusion module based on multi-kernel deep convolution (MK-DeConv) in the MKBF stage. Finally, a spectral–spatial knowledge distillation framework based on the teacher–student model is designed to achieve effective transfer of the rich spectral–spatial representation in the teacher model to the lightweight student model, namely SALMNet.Experimental results on the CAVE and Chikusei datasets demonstrate that SALMNet achieves significant model compression and computational acceleration while maintaining high-quality reconstruction performance.