FG-MoE: Frequency-Guided Mixture of Experts for Unified Remote Sensing Image Restoration

Abstract

In remote sensing (RS), image restoration is an essential but difficult problem because degradation patterns are diverse and scene content is highly heterogeneous. Existing RS image restoration methods still exhibit two major weaknesses: 1) they are built for a single degradation category, resulting in limited generalization across diverse restoration tasks; and 2) they process all frequency components of the image uniformly, making it difficult to reconstruct both global semantics and fine textures. In this article, we propose a frequency-guided mixture-of-experts (FG-MoE) framework, which enables unified RS image restoration across diverse degradation types by leveraging frequency-aware modeling to achieve accurate recovery of both semantic information and fine details. Specifically, FG-MoE consists of three key modules. First, we introduce the frequency mask prediction module (FMPM), which utilizes the fast Fourier transform (FFT) to generate frequency-aware masks for separating low- and high-frequency components. Guided by these masks, we design a frequency-guided mixture-of-experts (FG-MoE) module that dynamically assigns frequency components to specialized expert branches for targeted degradation processing. Furthermore, an expert fusion module (EFM) is proposed to adaptively integrate the outputs of both branches, enabling comprehensive and balanced reconstruction. Comprehensive experiments show that FG-MoE delivers strong restoration results under diverse RS image degradation conditions, including dehazing, cloud removal, and low-light enhancement.