Denoising Offshore Distributed Acoustic Sensing Using Masked Auto-Encoders to Enhance Earthquake Detection

Abstract

Offshore distributed acoustic sensing (DAS) has emerged as a powerful technology for seismic monitoring, expanding the capacities of cable networks and coastal seismic networks to monitor offshore seismicity. However, offshore DAS data often combine signals unfamiliar to seismologists, including new types of instrumental noise and ocean signals that overprint those from tectonic sources, which may hinder seismological research. We develop a self-supervised deep learning algorithm, a masked auto-encoder (MAE), to denoise DAS data for seismological purposes. The model is trained on DAS recordings of local earthquakes with randomly masked channels acquired on fiber-optic cables in the Cook Inlet offshore Alaska. To demonstrate the benefits of denoising for seismological research, we conduct the most fundamental steps to build any earthquake catalog: seismic phase picking, signal-to-noise ratio (SNR) estimation, and event association. We leverage the generalizability of ensemble deep learning models with cross-correlation to predict phase picks with sufficient precision for post-processing (e.g., earthquake location). The SNR of the denoised S waves of testing DAS data increased by 2.5 dB on average. The MAE denoised, on average, DAS data allows 2.7 times more S picks than the original noisy data for smaller regional earthquakes. The results demonstrate that our self-supervised MAE can elevate the accuracy and efficiency of seismic monitoring with higher earthquake detectability.