<p>In recent years, Transformers have shown remarkable performance in image dehazing tasks by effectively capturing long-range dependencies. However, they often overlook local features and frequency domain information. To solve these issues, a wavelet-guided spatial-frequency Transformer combined with physics-based refinement network, abbreviated as WG-SFFormer, is proposed for remote sensing image dehazing. It is structured around three key components: a backbone, a wavelet-guided branch, and a feature refinement part. Specifically, the backbone incorporates a spatial-frequency Transformer module, composed of two parallel self-attention mechanisms: one operating in the frequency domain and the other in the spatial domain, allowing separate learning of frequency and spatial features. The frequency domain self-attention employs a frequency feature block to extract relevant information, leveraging the Fast Fourier Transform to decouple haze-related features, thereby enhancing haze removal efficiency. The wavelet-guided branch introduces a cascaded wavelet-guided module to capture both high-frequency and low-frequency information, effectively supplementing the missing local information in the backbone. In the feature refinement part, a physics-based refinement module is proposed to further eliminate haze and generate clear remote sensing images. Extensive experimental results indicate that WG-SFFormer achieves superior dehazing performance compared to competing methods. Additionally, ablation studies validate the effectiveness of its core components.</p>