<p>Surface waves are a kind of coherent noise in seismic data, characterized by low frequency, high amplitude, low velocity, and strong spatial continuity. These characteristics significantly complicate the identification and extraction of effective signals. Because conventional methods rely solely on single-domain feature extraction and limited representation capacity, they often fail to suppress overlapping surface waves and preserve weak signals. To alleviate these issues, this paper proposes Spatial-Frequency Domain Learning with Refinement Module for Surface Wave Suppression (SFNet) to capture more comprehensive features in two stages. In the first stage, a Dual-domain Hybrid Feature Decoupler extracts spatial and frequency features through two branches. The spatial branch incorporates a Refinement Module that combines edge enhancement with adaptive feature recalibration to preserve weak signals and maintain boundary continuity. The frequency branch captures global contextual information to facilitate the discrimination of surface waves. Then, both branches are enhanced by a Convolutional Block Attention Module to adaptively emphasize critical features. In the second stage, a Multi-Kernel Convolution Module is employed to improve the separation of overlapping regions. Experiments are undertaken on both synthetic and field datasets. Results show that SFNet outperforms five mainstream and advanced methods, not only effectively suppressing surface waves but also successfully preserving the integrity and continuity of weak effective signals.</p>