Assessing the Efficacy of Machine Learning Algorithms to Denoise Seismic Data

Abstract:

With the explosion of digital seismic data since the early 1990s, machine learning has become a powerful tool to explore this ever-growing dataset. The flexibility of machine learning algorithms allow us to explore datasets that are too big to evaluate manually and even provide novel approaches to address known problems with seismic data. Machine learning is particularly useful for removing non-seismic signals from inherently noisy datasets based on pattern recognition. In this study, we train a machine learning algorithm to remove non-seismic signals from records collected by the Japanese National Research Institute of Earth Science and Disaster Resilience’s (NIED) Seafloor Observation Network for Earthquakes and Tsunamis (S-Net) cabled ocean bottom sensors. The training dataset is created using traces from NIED’s High Sensitivity Seismograph Network (Hi-Net), an on-land borehole geophone network. Instrumental differences between the two networks can be dealt with by removing instrument response. However, the difference in site characteristics between Hi-Net and S-Net are more difficult to overcome. Because of this difference in site characteristics, it is important to use a multifaceted approach to determining how effective our model results are. In addition to using the traditional metric of signal-to-noise increase, we propose other metrics that are rooted in geophysical realities to systematically determine the effectiveness of our training dataset for the problem at hand. These metrics include STA/LTA trigger times, spectral content of waveforms, and envelope shape; we aim to use these metrics to look for problems in our model output ranging from cycle slipping to amplitude changes. The goal of this analysis is to determine whether the model output data we generate will be useful to travel-time tomography applications in the Japan Arc.