Deep-Learning-Based Denoising for Earthquake Monitoring in an Urban Setting with Dense Array Data

Abstract:

Earthquake monitoring in urban settings is important but challenging, because relatively weak earthquake signals can be overwhelmed by strong cultural noise sources. The deep-learning-based DeepDenoiser algorithm, which was originally trained on Northern California earthquake datasets recorded in relatively quiet places, has shown remarkable performance for the denoising of seismic data recorded in similar regions. In this study, we extend the application of DeepDenoiser to seismic data in an urban environment, by training DeepDenoiser with a large catalog of urban noise sources from the Long Beach dense nodal deployment and an expanded earthquake signal dataset using additional training samples from the San Jacinto Fault Zone nodal deployment. The Long Beach deployment comprised thousands of vertical component sensors deployed in two phases in 2011 and 2012. These data represent a rich source of urban seismological noise for training an enhanced deep-learning-based denoising algorithm. The San Jacinto Fault Zone dense nodal array uses the same type of nodal sensors as those in the Long Beach deployment, but it has more records of earthquakes, and these have higher signal-to-noise ratio (snr).

The new version of DeepDenoiser trained on these data recovers earthquake signals more reliably from noisy urban seismic settings. To explore whether the recently reported widespread earthquakes in the upper mantle under Long Beach are real or artifacts, we use the denoised Long Beach data for earthquake location, but do not find events from that depth. More generally, the new denoising algorithm should help improve the performance of existing earthquake monitoring networks in urban settings by enhancing the snr and clear identification of earthquake signals.