Detection and characterization of slow deformation from GNSS data by deep learning in the Cascadia subduction zone

The stress that accumulates on faults due to tectonic plate motion can be released seismically and aseismically. The seismic release of stress takes place during earthquakes at short time scales (seconds to minutes), and can be identified on seismic records. The aseismic part of this stress release occurs during Slow Slip Events (SSEs), that last from days to years and do not radiate energetic seismic waves. SSEs are monitored with dense Global Navigation Satellite System (GNSS) networks that record the deformation induced at the surface. A precise identification of slow slip events is key to better understand the mechanics of active faults and to better describe the role of aseismic slip in the seismic cycle. Yet, the characterization of SSEs of various sizes from existing GNSS networks is challenging, and extensive SSEs catalogs remains sparse and incomplete: for example, 64 events for SSEs in Cascadia (Michel et al., 2018), 24 long-term SSEs (Takagi et al., 2019) and 284 short-term SSEs (Okada et al., 2022) in Nankai, Japan. Traditional SSE characterization either focus on specific events, identified visually with high signal to noise ratio (e.g. Radiguet et al., 2016), use time series decomposition approaches such as ICAIM (independent Component analysis inversion method) (Michel et al., 2018; Radiguet et al., 2020), or, for small events, geodetic template matching (Okada et al., 2022; Rousset et al., 2017).

We focus on the Cascadia subduction zone, where a link between slow slip and bursts of tremors has been established (Rogers & Dragert, 2003). In this direction, tremor catalogues can be used to validate potential SSEs detections against the spatiotemporal distribution of tremors. Moreover, a catalogue of SSEs has been recently assessed by (Michel et al., 2019), providing additional benchmark to our analyses. We generate synthetic SSEs from synthetic dislocations (Okada, 1985) using the slab2 model (Hayes et al., 2018). Each SSE template, assumed as a sigmoidal-shaped transient, is further added to a window of noise obtained from real GNSS data (Costantino et al., in prep.).

We develop a deep learning-based method for the systematic detection and characterization of SSEs using a Convolutional Neural Network (CNN) in combination with a Graph Neural Network (GNN). We test our method both on synthetic and real position time series. Results on synthetic data are consistent and show a detection trade-off between the SSEs location, magnitude and the density of the GNSS network. Results on real GNSS positional time series show a good agreement with existing catalogues (cf. Michel et al., 2019). Moreover, new detections have been carried out, which correlate well with the temporal distribution of tremors, suggesting that those events could be new SSE detections, which will be further validated by assessing their spatial-temporal consistency through scaling laws output by the deep learning model.