9:30 AM - 9:45 AM
[SSS04-03] Noise Reduction for Ocean Bottom Seismometer Waveforms by Generative Models
Keywords:ocean bottom seismometer, denoising, machine learning, generative model
Ocean bottom seismometers (OBS) yield valuable data from both natural and artificial sources, shedding light on the subsurface structure beneath the ocean. Artificial sources, with their dense, precise, and controllable waves, are ideal for resource exploration and for obtaining detailed insights into fine-scale features such as fault zones and volcanoes. However, these records are often plagued by varying levels of noise due to factors such as weather conditions, ocean currents, and background seismic activity, complicating their interpretation. In this study, we leverage the capabilities of generative models to reduce noise in seismograms recorded by OBS.
We trained a denoising diffusion probabilistic model (DDPM), which has gained acclaim for its performance in image generation, to enhance signals in seismic profiles. The training data was 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. The repetition of the survey allows for diverse stacking to eliminate low-quality traces, and synthetic reference profiles serve as ground truth for machine learning.
Our study represents the first application of DDPM to actual OBS data, marking a departure from previous research that has only examined the model using simulated waveforms (e.g., Durall et al., 2023). We tested the trained model on a separate dataset from the Noto survey, as well as data from other surveys. Our model effectively removes random background noise and extreme noise present in certain traces, performing comparably to the diversity stacking method, regardless of variations in locations and instruments. Furthermore, undesired direct waves due to short airgun shot intervals can be suppressed as well. This not only highlights the potential of generative models to enhance seismic signals but also suggests significant time and cost savings in conducting marine seismic surveys to obtain high-quality data.
We trained a denoising diffusion probabilistic model (DDPM), which has gained acclaim for its performance in image generation, to enhance signals in seismic profiles. The training data was 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. The repetition of the survey allows for diverse stacking to eliminate low-quality traces, and synthetic reference profiles serve as ground truth for machine learning.
Our study represents the first application of DDPM to actual OBS data, marking a departure from previous research that has only examined the model using simulated waveforms (e.g., Durall et al., 2023). We tested the trained model on a separate dataset from the Noto survey, as well as data from other surveys. Our model effectively removes random background noise and extreme noise present in certain traces, performing comparably to the diversity stacking method, regardless of variations in locations and instruments. Furthermore, undesired direct waves due to short airgun shot intervals can be suppressed as well. This not only highlights the potential of generative models to enhance seismic signals but also suggests significant time and cost savings in conducting marine seismic surveys to obtain high-quality data.