Deep-learning-based denoising of prestack seismic angle gathers with application to time-lapse data

Deep-learning algorithms have become powerful tools for denoising, leveraging their automatic learning and feature extraction abilities to detect and remove noise from images while preserving fine details and textures. These methods reduce reliance on manual adjustments and parameter tuning, improving efficiency and consistency. In geophysics, many researchers have applied various deep-learning algorithms for seismic denoising; however, most applications have focused on prestack shot/receiver gathers or poststack seismic data. The use of deep learning to denoise offset- or angle-domain prestack seismic gathers remains rare. Prestack seismic gathers are essential in exploration geophysics, playing key roles in velocity model building, amplitude-variation-with-angle analysis, and imaging complex geologic structures. Given the strong performance of deep-learning algorithms in denoising and the importance of seismic gathers, this study applies deep learning to denoise angle-domain prestack seismic gathers (angle gathers). Specifically, we use a conditional generative adversarial network known as Pix2Pix to denoise PP-wave angle gathers. Tests on synthetic and field data demonstrate its feasibility, with results showing that the trained model effectively removes random and correlated noise in angle gathers while filling in missing signals. Its application to synthetic time-lapse data further demonstrates strong performance, which is particularly significant for carbon capture and storage (CCS) monitoring, where high-fidelity time-lapse seismic data are crucial for tracking subsurface [Formula: see text] migration and storage integrity. The study examines the network’s ability to capture physical relationships inherent in angle gathers. Specifically, the Zoeppritz equation is incorporated into the training data synthesis, ensuring that the model learns the angular dependence of reflection coefficients. This integration of domain knowledge enhances the denoising process, making it a valuable tool for advancing CCS monitoring and evaluation efforts.