Seismic Noise Attenuation in Imaging Gathers Using a Residual U-Net: Application to Seismic Data from the Xihu Sag

Seismic imaging gathers are often contaminated with residual noise introduced by pre-stack denoising and migration imaging algorithms, which compromises imaging accuracy and the reliability of subsequent interpretation. Traditional denoising methods often lack precise control over the filtering process: excessive filtering can remove seismic details and blur fault structures, while insufficient filtering yields only limited enhancement of seismic data quality. Moreover, many noise removal processes require manual intervention, reducing overall processing efficiency. To address this, the present study proposes a noise removal method for seismic imaging gathers based on a U-Net neural network with residual learning. It innovatively introduces deep-learning-based denoising into the domain of seismic imaging gathers, a field that has received limited attention in prior studies. To better accommodate the noise characteristics of imaging gathers, we customized the U-Net architecture by adjusting network depth, residual connections, and feature extraction layers. The incorporation of residual connections enhances the network’s generalization ability, enabling it to effectively distinguish noise from valid signals and achieve high-precision denoising. When applied to seismic data from the Xihu Sag, the method significantly improves the signal-to-noise ratio of the imaging gathers, produces high-quality seismic profiles for oil and gas exploration, and supports exploration efforts in the region.