Machine Learning-enhanced Realistic Framework for Real-time Seismic Monitoring - The Winning Solution of the 2017 International Aftershock Detection Contest

Dazhong Shen, Qi Zhang, Tong Xu, Hengshu Zhu, Wenjia Zhao, Zikai Yin, Peilun Zhou, Lihua Fang, Enhong Chen, Hui Xiong

Abstract

Identifying the arrival times of seismic P-phases plays a significant role in real-time seismic monitoring, which provides critical guidance for emergency response activities. While considerable research has been conducted on this topic, efficiently capturing the arrival times of seismic P-phases hidden within intensively distributed and noisy seismic waves, such as those generated by the aftershocks of destructive earthquakes, remains a real challenge since most common existing methods in seismology rely on laborious expert supervision. To this end, in this paper, we present a machine learning-enhanced framework based on ensemble learning strategy, EL-Picker, for the automatic identification of seismic P-phase arrivals on continuous and massive waveforms. More specifically, EL-Picker consists of three modules, namely, Trigger, Classifier, and Refiner, and an ensemble learning strategy is exploited to integrate several machine learning classifiers. An evaluation of the aftershocks following the MS 8.0 Wenchuan earthquake demonstrates that EL-Picker can not only achieve the best identification performance but also identify 120% more seismic P-phase arrivals as complementary data. Meanwhile, experimental results also reveal both the applicability of different machine learning models for waveforms collected from different seismic stations and the regularities of seismic P-phase arrivals that might be neglected during manual inspection. These findings clearly validate the effectiveness, efficiency, flexibility and stability of EL-Picker.

Introduction

Seismologists rely on the precise identification of "P-phase arrivals"—the initial, non-destructive waves of an earthquake—to activate early warning systems and coordinate emergency responses. However, this task becomes exceptionally difficult during aftershock sequences. In these periods, hundreds of small tremors can be buried in noise or overlapping waveforms, overwhelming traditional detection methods.

Existing solutions often fall into two extremes: traditional detectors like the Short-Term Average/Long-Term Average (STA/LTA) are computationally efficient but prone to false alarms, while manual expert inspection is accurate but too slow for real-time needs. While deep learning models (e.g., CNNs, LSTMs) offer high accuracy, their computational demands on continuous, massive data streams often preclude real-time deployment. This paper presents EL-Picker, a modular framework that combines the speed of traditional triggers with the accuracy of ensemble machine learning.

Methodology: The EL-Picker Framework

EL-Picker is designed as a three-stage pipeline to balance efficiency and accuracy:

  1. Trigger Module: This module serves as a high-speed initial filter. It utilizes a traditional broadband phase picker (FilterPicker) to scan continuous waveforms. By flagging only potential candidates, it prevents the system from wasting computational resources on background noise.
  2. Classifier Module: Candidates passed by the trigger are analyzed by an ensemble of machine learning models. The system extracts between 679 and 715 features, including amplitude fluctuations and "Spectral Waterfalls," across multiple frequency bands and channels. These features are fed into nine different classifiers (including Random Forest, SVM, and AdaBoost), whose outputs are aggregated by a Logistic Regression "Meta-Model" to produce a final confidence score.
  3. Refiner Module: For accepted candidates, this module pinpoints the exact arrival time using the Akaike Information Criterion (AIC). It also performs multi-station cross-validation to ensure the detected signal is physically consistent across the geographic network.
The EL-Picker framework architecture, showing the interaction between the Trigger, Classifier, and Refiner modules.

The robustness of the Classifier module stems from its extensive feature engineering, which captures the complex nuances of seismic signals across different time windows and bandwidths.

Summary of the 679–715 features extracted for the Classifier module, categorized by type, time window, and bandwidth.

Experimental Evaluation

The framework was rigorously tested on two major datasets: the MS 8.0 Wenchuan earthquake (2008) in China and the MW 6.0 South Napa earthquake (2014) in California. The Wenchuan dataset provided over 12,000 labeled P-phase arrivals across 12 stations, while the Napa dataset served to test the model's robustness across different geological regions.

Key Results

  • Superior Accuracy: EL-Picker consistently outperformed baseline models, including deep learning architectures like LSTM and CNN, in both precision and recall.
  • Discovery of Hidden Events: In a blind test of 1,000 random samples, EL-Picker identified 120.6% more P-phases than were originally labeled by experts. Upon review by the China Earthquake Administration, 90.7% of these "extra" picks were confirmed as real, previously missed aftershocks.
  • Geographic Personalization: The study found that different ML models performed optimally at different stations based on local geology. For instance, Random Forest excelled in mountainous regions, suggesting that "one-size-fits-all" models may be suboptimal for large-scale networks.
  • Operational Efficiency: The Trigger module processes signals in just 0.0025 ms, making the system highly suitable for real-time monitoring on standard hardware.

Conclusion

EL-Picker demonstrates that a hybrid approach—combining traditional seismic triggers with ensemble machine learning—can significantly improve the speed and accuracy of earthquake monitoring. By identifying events missed by human experts and maintaining low computational overhead, it provides a realistic framework for real-time seismic response. Future work will address the identification of S-phases and the direct integration of spatial data into the machine learning features.

Summary updated

Abstract

Identifying the arrival times of seismic P-phases plays a significant role in real-time seismic monitoring, which provides critical guidance for emergency response activities. While considerable research has been conducted on this topic, efficiently capturing the arrival times of seismic P-phases hidden within intensively distributed and noisy seismic waves, such as those generated by the aftershocks of destructive earthquakes, remains a real challenge since most common existing methods in seismology rely on laborious expert supervision. To this end, in this paper, we present a machine learning-enhanced framework based on ensemble learning strategy, EL-Picker, for the automatic identification of seismic P-phase arrivals on continuous and massive waveforms. More specifically, EL-Picker consists of three modules, namely, Trigger, Classifier, and Refiner, and an ensemble learning strategy is exploited to integrate several machine learning classifiers. An evaluation of the aftershocks following the MS 8.0 Wenchuan earthquake demonstrates that EL-Picker can not only achieve the best identification performance but also identify 120% more seismic P-phase arrivals as complementary data. Meanwhile, experimental results also reveal both the applicability of different machine learning models for waveforms collected from different seismic stations and the regularities of seismic P-phase arrivals that might be neglected during manual inspection. These findings clearly validate the effectiveness, efficiency, flexibility and stability of EL-Picker.
BibTeX
@misc{shen2019machine,
  author = {Dazhong Shen and Qi Zhang and Tong Xu and Hengshu Zhu and Wenjia Zhao and Zikai Yin and Peilun Zhou and Lihua Fang and Enhong Chen and Hui Xiong},
  title = {{Machine Learning-enhanced Realistic Framework for Real-time Seismic Monitoring - The Winning Solution of the 2017 International Aftershock Detection Contest}},
  year = {2019},
  eprint = {1911.09275},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1911.09275},
}
Figure