A time–space and frequency–wavenumber dual-domain fusion network for coupling noise separation

SUMMARY Distributed acoustic sensing (DAS) technology has gained widespread attention in seismic exploration due to its high spatial resolution and low deployment cost. However, the presence of coupling noise in DAS data significantly affects the accurate extraction and interpretation of seismic signals. Coupling noise typically appears as narrow-band stripe-like or zigzag-like interference and shares similar characteristics with seismic signals in the time–space (T–S) domain, making it challenging for traditional denoising methods to achieve effective signal-noise separation without residual noise or signal leakage. To address these challenges, this paper proposes a deep learning-based dual-domain fusion approach that integrates both T–S and frequency–wavenumber (F–K) domain information to enhance the accuracy of coupling noise separation. The method leverages the narrow-band characteristics of coupling noise in the F–K domain while incorporating spatiotemporal information from the T–S domain to achieve cross-domain feature fusion, thereby improving the separability between signals and coupling noise. Experimental results demonstrate that the proposed method significantly improves coupling noise suppression performance on both synthetic and field DAS vertical seismic profile data while minimizing signal leakage. Furthermore, in corridor stacking experiments, the method effectively reduces the impact of coupling noise on seismic interpretation, improving the reliability of subsurface formation analysis. Compared to conventional F–K filtering, single-domain network and denoising diffusion model, the proposed approach achieves superior performance in terms of coupling noise suppression and signal amplitude preservation.