Spatio-Temporal Graph Structure Learning for Earthquake Detection

TL;DR

This work addresses earthquake detection in EEW contexts by integrating multi-station seismic data through a Spatio-Temporal Graph Convolutional Network that learns both static and dynamic spatial relationships via Spectral Structure Learning Convolution. It represents the network as a graph ${\mathcal G}$ with learnable adjacency implemented through a static component ${W^s}$ and a dynamic component ${W^d= X^T W_\phi X}$, applying Chebyshev polynomial filters ${T_k}$ to produce ${F_s}$ and ${F_d}$, then merging them as ${H=\text{ReLU}(F_s)+\text{ReLU}(F_d)}$, with temporal dependencies captured by a GRU and a sigmoid head to output per-time-step probabilities ${\tilde D}$. Evaluated on MeSO-net data around Tokyo, the approach outperforms a baseline conventional GCN, achieving higher true positive rates at lower false positive rates and providing per-station probability time series suitable for real-time EEW. The method advances earthquake detection by leveraging learnable spatio-temporal graph structures in a lightweight architecture, with potential extensions to sliding-window real-time processing and adaptive graph handling for evolving noise conditions.

Abstract

Earthquake detection is essential for earthquake early warning (EEW) systems. Traditional methods struggle with low signal-to-noise ratios and single-station reliance, limiting their effectiveness. We propose a Spatio-Temporal Graph Convolutional Network (GCN) using Spectral Structure Learning Convolution (Spectral SLC) to model static and dynamic relationships across seismic stations. Our approach processes multi-station waveform data and generates station-specific detection probabilities. Experiments show superior performance over a conventional GCN baseline in terms of true positive rate (TPR) and false positive rate (FPR), highlighting its potential for robust multi-station earthquake detection. The code repository for this study is available at https://github.com/SuchanunP/eq_detector.

Figures (4)

• Figure 1: (a) Map of the Kanto district relative to Japan. (b) Locations of the MeSO-net stations (gray and blue dots) in the Kanto district. (c) The 13 seismic stations (blue dots) used in this study.
• Figure 2: (a) The architecture of our GCN. The input comprises three-component waveform data from each of the 13 seismic stations, while the output is a detection probability time series for each station. The GCN consists of a spatial-temporal feature extraction module and an output module. The spatial-temporal feature extraction module contains 5 Spectral SLC layers. The output module includes a fully connected layer followed by a sigmoid layer. (b) The Spectral SLC layer integrates a spatial component to capture static and dynamic graph structures, and a temporal component to model temporal dependencies in the signals.
• Figure 3: ROC curves of our proposed method and the baseline (Conventional GCN) model, evaluated on the test dataset. The ROC curves illustrate the trade-off between the true positive rate (TPR) and false positive rate (FPR) at various minimum detection probabilities (MDP). The color bar represents the corresponding MDP values. Our method demonstrates superior performance, achieving a higher TPR and lower FPR. The optimal MDP, representing the FPR-TPR point nearest to the upper-left corner, is indicated by an arrow for each model.
• Figure 4: Detection results for 13 seismic stations using unseen waveform data from the North-South component (September 4, 2011, 00:20:14 to 00:20:34, Japan time). The top panel of each subplot shows the waveform (blue), while the bottom panel compares detection probabilities from our method (dashed brown), the baseline model (dotted green), and the ground-truth labels (solid orange).