Presentation Information
[S22-10]Denoising Ocean Bottom Seismometer Data with Diffusion Models: A Real-World Application of DDPM to Geophysical Signal Enhancement
*Jun Su1, Ryoichiro Agata1, Gou Fujie1, Yasuyuki Nakamura1 (1. JAMSTEC)
In this study, we explore the application of denoising diffusion probabilistic models (DDPM) for enhancing seismic data captured by ocean bottom seismometers (OBS). OBS data, while rich in information about subsurface structures from both natural and artificial seismic sources, often suffer from noise introduced by instrumental and environmental factors such as wind, operating current, natural seismicity, previously fired shots, and even marine mammals. We present a novel approach leveraging generative models to effectively reduce this noise, focusing on a dataset from a seismic survey near the Noto Peninsula, Japan, conducted from August 30 to September 6, 2024. In this survey, airgun shots were strategically fired at 200-meter intervals, and the densely recorded data allowed for diverse stacking techniques to synthesize high-quality traces.
Our implementation of DDPM marks the first instance of training this model with its input derived from field data, differing from prior research that utilized only simulated waveforms (e.g., Durall et al, 2023). We trained the model using the Noto survey dataset and evaluated its performance on both the Noto dataset and additional datasets from other surveys. The results illustrate that our DDPM successfully mitigates random background noise and impulsive noise without compromising data quality, distorting waveforms, or altering the actual amplitude of specific frequencies. The denoised results match the performance of traditional diversity stacking methods across varying locations and instruments. Additionally, the model effectively suppresses unwanted direct waves resulting from short airgun shot intervals.
To address the practical deployment challenges in marine seismic survey environments, we developed a web-based application that enables remote processing of SEGY format files on a GPU-enabled server. This cloud-based inference system allows survey teams to upload raw seismic data from shipboard locations with limited computational resources and receive denoised results with a fast response, facilitating immediate data quality assessment and survey optimization decisions at sea. This work demonstrates the practical viability of diffusion models for seismic data enhancement and opens new directions for applying deep learning to reduce data acquisition costs in geophysical surveys.
Our implementation of DDPM marks the first instance of training this model with its input derived from field data, differing from prior research that utilized only simulated waveforms (e.g., Durall et al, 2023). We trained the model using the Noto survey dataset and evaluated its performance on both the Noto dataset and additional datasets from other surveys. The results illustrate that our DDPM successfully mitigates random background noise and impulsive noise without compromising data quality, distorting waveforms, or altering the actual amplitude of specific frequencies. The denoised results match the performance of traditional diversity stacking methods across varying locations and instruments. Additionally, the model effectively suppresses unwanted direct waves resulting from short airgun shot intervals.
To address the practical deployment challenges in marine seismic survey environments, we developed a web-based application that enables remote processing of SEGY format files on a GPU-enabled server. This cloud-based inference system allows survey teams to upload raw seismic data from shipboard locations with limited computational resources and receive denoised results with a fast response, facilitating immediate data quality assessment and survey optimization decisions at sea. This work demonstrates the practical viability of diffusion models for seismic data enhancement and opens new directions for applying deep learning to reduce data acquisition costs in geophysical surveys.
