Enhancing Ocean Bottom Seismometer Data: Diffusion Model Applications for Noise Reduction in Marine Surveys Near Noto Peninsula, Japan

Seismic profiles obtained from ocean bottom seismometers (OBS) during seismic surveys are crucial for understanding subsurface structures. However, these profiles often contain records that are affected by varying levels of noise due to factors such as weather conditions, ocean currents, and background seismicity, which complicates interpretation. One common approach to mitigate this uncertainty is diversity stacking, which involves retrieving seismic records with the same source-receiver pairs multiple times to filter out noise. Unfortunately, this technique can increase the costs of marine seismic surveys and limit data availability.

In this study, we utilize OBS data collected from a seismic survey near the Noto Peninsula, Japan, conducted between August 30 and September 6, 2024. Airgun shots were fired at 200-meter intervals, repeated five times along a 100-kilometer survey line, and recorded by 40 OBS with 2-kilometer spacing. We trained a machine learning model to reduce the noise in seismic profiles from each shot, using profiles processed through diversity stacking as a reference. Specifically, we applied a denoising diffusion probabilistic model (DDPM) based on the methodology outlined by Durall et al. (2023). This model, which has recently demonstrated efficacy as an image generator, takes a list of words, sentences, or images as input and iteratively refines the result towards the desired output using a neural network. While Durall et al. (2023) trained their model solely on simulated seismograms as target images, our approach leveraged diversity stacking and incorporated real-world waveforms as training data for the first time.

An example profile from the test set indicates that the trained model effectively addresses both random background noise and extreme noise present in certain traces, successfully reducing noise levels in profiles from a singular shot to be comparable to those achieved through diversity stacking. These results suggest that by enhancing OBS data with the DDPM, it is possible to obtain a clearer seismic structure of the deeper subsurface and a broader range of data with fewer airgun shots.