MFEM: Multiscale Frequency-Enhanced Mamba for Lightweight Remote Sensing Image Super-Resolution
Wangyou Chen, Laigan Luo, Shenming Qu, Chaoxu Dang · IEEE Geoscience and Remote Sensing Letters · 2025
In recent years, remarkable progress has been achieved in lightweight Remote Sensing Image Super-Resolution (RSISR) techniques. However, these methods, in order to maintain lightweight property and efficient computation, are often affected by inadequate feature extraction and limited receptive fields, which restrict the precise reconstruction of structurally complex remote sensing images (RSI). To address this issue, we propose a Multi-Scale Frequency-Enhanced Mamba (MFEM), which employs a unique multi-scale convolutional structure to enhance feature extraction and utilizes a frequency-enhanced Mamba to capture long-range dependencies with linear complexity. Specifically, we design a Multi-Scale Blueprint Residual Block (MBRB) with three specifications, which allows the network to efficiently extract multi-scale features of RSIs while maintaining lightweight property by using different specifications of MBRB at varying network depths. Furthermore, we develop a Frequency Mamba Enhancement Block (FMEB), which enhances Mamba from within through frequency analysis to globally model multi-level features while improving local information loss. Extensive experiments demonstrate that our method achieves better super-resolution performance while maintaining lightweight property.