Detecting slow slip events in the Cascadia subduction zone from GNSS time series using deep learning

Abstract

Slow Slip Events (SSEs) are like long-duration slow earthquakes during which stress is gradually released over several days to months, and a comprehensive catalog of SSEs is essential for a better understanding of the earthquake cycle. However, SSEs usually only produce mm to cm surface deformations, making them a challenge to identify from raw Global Navigation Satellite System (GNSS) time series, which are often obscured by low-frequency background noise. We devise an approach that first employs variational Bayesian Independent Component Analysis to improve the signal-to-noise ratio of GNSS time series and then utilizes deep learning combining bidirectional Long Short-Term Memory and two different attention mechanisms to identify SSEs. We apply this new method to the GNSS three-component time series at 240 stations along the Cascadia subduction zone from 2012 to 2022. A total of 56 SSEs are detected, 18 more than the number in the existing SSEs catalogs during the same period. The starting time, duration, spatial and propagation pattern of the 56 SSEs are consistent with the tremor catalog, which helps to gain new insights into the slip behavior in the Cascadia subduction zone. In general, our work provides an effective framework for extracting subtle signals hidden in GNSS time series.