A self-supervised random noise attenuation method for seismic data based on self-attention

Abstract Seismic data is a cornerstone for oil and gas exploration, geophysical research, and high-precision 3D geological modeling. However, during data acquisition, inevitable noise interference severely degrades the clarity of geological boundaries (e.g., faults and stratigraphic interfaces), directly impacting the accuracy of subsequent 3D geological model construction and mineral targeting. Furthermore, the high cost and practical impossibility of obtaining clean ground-truth seismic records (i.e., noise-free training labels) pose significant bottlenecks to conventional supervised deep learning methods. To address these issues, this paper proposes a self-supervised learning-based method for 3D seismic image denoising. Unlike traditional supervised approaches, the proposed model eliminates the need for labeled training data. It incorporates a self-attention mechanism to extract spatial structures and geological features, thereby enhancing noise suppression and data reconstruction performance. Experimental results on both synthetic and field datasets demonstrate that the proposed SA-S2S model effectively attenuates noise while preserving key geological structures. By improving the signal-to-noise ratio and sharpening fault boundaries and stratigraphic interfaces, the method provides high-quality seismic inputs that significantly enhance the identification accuracy of geological structural boundaries in subsequent 3D seismic data interpretation. Under three different Gaussian random noise intensities, the model outperforms other denoising methods in both qualitative visual comparisons and quantitative metrics such as PSNR and SSIM in almost all cases. In summary, the proposed method offers an efficient and practical solution for seismic data denoising, with strong potential for broad application in geoscience and exploration workflows.