[Paper Review] Slow slip detection with deep learning in multi-station raw geodetic time series validated against tremors in Cascadia

Giuseppe Costantino, Sophie Giffard‐Roisin+4 more|arXiv (Cornell University)|May 31, 2023

earthquake and tectonic studies4 citations

TL;DR

This study presents the first end-to-end deep learning detector for slow slip events (SSEs) in raw, multi-station GNSS time series, using a physics-informed synthetic training dataset and a hybrid CNN-Transformer architecture. It detects 78 SSEs in Cascadia (2007–2022), recovering 87.5% of known events and revealing a day-long time lag between slow deformation and tremor activity, suggesting SSEs may trigger nearby aseismic ruptures.

ABSTRACT

Slow slip events (SSEs) originate from a slow slippage on faults that lasts from a few days to years. A systematic and complete mapping of SSEs is key to characterizing the slip spectrum and understanding its link with coeval seismological signals. Yet, SSE catalogues are sparse and usually remain limited to the largest events, because the deformation transients are often concealed in the noise of the geodetic data. Here we present the first multi-station deep learning SSE detector applied blindly to multiple raw geodetic time series. Its power lies in an ultra-realistic synthetic training set, and in the combination of convolutional and attention-based neural networks. Applied to real data in Cascadia over the period 2007-2022, it detects 78 SSEs, that compare well to existing independent benchmarks: 87.5% of previously catalogued SSEs are retrieved, each detection falling within a peak of tremor activity. Our method also provides useful proxies on the SSE duration and may help illuminate relationships between tremor chatter and the nucleation of the slow rupture. We find an average day-long time lag between the slow deformation and the tremor chatter both at a global- and local-temporal scale, suggesting that slow slip may drive the rupture of nearby small asperities.

Motivation & Objective

• To develop a systematic, automated method for detecting slow slip events (SSEs) in raw geodetic time series, overcoming limitations of manual and pre-processing-dependent approaches.
• To address the scarcity of labeled training data for SSE detection by generating a realistic, physics-based synthetic GNSS dataset with spatial correlations and data gaps.
• To improve detection of low-magnitude SSEs (Mw < 6) that are often masked by noise in standard geodetic analyses.
• To validate the model against independent tremor catalogs and assess temporal relationships between SSEs and tremor activity.

Proposed method

• The SSEgenerator synthesizes realistic multi-station GNSS time series by modeling slow slip deformation on a fault interface, incorporating spatial correlations, seasonal signals, and realistic data gaps.
• The SSEdetector employs a hybrid deep learning architecture combining 1D convolutional neural networks (CNNs) for local feature extraction and a self-attention mechanism (Transformer) for long-range temporal dependencies across stations.
• The model is trained end-to-end on raw, unprocessed GNSS position time series without pre-filtering, enabling it to learn noise signatures and distinguish true deformation transients.
• Spatial stacking of signals is guided by physical knowledge of SSEs, aligning station data according to their proximity to the subduction zone and expected deformation patterns.
• The model outputs a probability time series indicating the likelihood of an SSE at each time step, which is then thresholded to identify events.
• Validation uses cross-correlation between SSE detections and daily tremor counts, with statistical filtering (correlation > 0.4) to ensure meaningful associations.

Figure 1: Schematic architecture of SSEgenerator and SSEdetector. (a) Overview of the synthetic data generation (SSEgenerator). In the matrix, each row represents the GNSS position time series for a given station, color-coded by the value of the position. The 135 GNSS stations considered in this stu

Experimental results

Research questions

• RQ1 Can a deep learning model trained on synthetic, physics-based GNSS data detect real slow slip events in raw, unprocessed time series without pre-processing?
• RQ2 How does the temporal relationship between slow slip deformation and tremor activity vary across global and local timescales in Cascadia?
• RQ3 To what extent can the model retrieve known SSEs from existing catalogs, and how does its performance compare to manual or filtering-based detection methods?
• RQ4 Can the model provide reliable estimates of SSE duration and temporal overlap with other events, and how do these compare to independent observations?
• RQ5 Does the observed time lag between slow slip and tremor activity suggest a causal relationship, such as SSEs nucleating nearby aseismic ruptures?

Key findings

• The model successfully detects 78 slow slip events in Cascadia between 2007 and 2022, recovering 87.5% of previously catalogued events.
• All detected SSEs occur within a peak of tremor activity, confirming strong spatiotemporal correlation between slow slip and tremor signals.
• A consistent day-long time lag is observed between the onset of slow deformation and the peak of tremor activity at both global and local temporal scales.
• The model provides reliable proxies for SSE duration, with inferred durations matching observed tremor pulse widths when filtered for statistical significance.
• The overlap percentage between detected events indicates that 26% of events have significant temporal overlap, suggesting complex, possibly interacting deformation processes.
• The study demonstrates that deep learning models trained on realistic synthetic data can detect low-magnitude SSEs in raw geodetic data, overcoming the signal-to-noise limitations of traditional methods.

Figure 2: Performance of SSEdetector on synthetic data . (a) The blue curve represents the true positive rate (probability that an actual positive will test positive), computed on synthetic data, as a function of the magnitude. (b) Map showing the spatial distribution of the magnitude threshold for