[Paper Review] A Machine Learning-enhanced Robust P-Phase Picker for Real-time Seismic Monitoring

Dazhong Shen, Qi Zhang+8 more|arXiv (Cornell University)|Nov 21, 2019

Seismology and Earthquake Studies44 references4 citations

TL;DR

This paper proposes EL-Picker, a machine learning-enhanced, ensemble learning-based framework for real-time automatic detection of seismic P-phase arrivals in noisy, continuous waveforms. By combining a low-complexity trigger, a diverse ensemble of classifiers, and a refinement module, EL-Picker achieves superior performance—identifying 120% more P-phase arrivals than expert-labeled data—while revealing subtle, previously missed seismic patterns due to human oversight.

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.

Motivation & Objective

• To address the challenge of detecting weak, low signal-to-noise ratio (SNR) P-phase arrivals in continuous, noisy seismic waveforms from aftershocks.
• To reduce reliance on labor-intensive manual labeling by automating P-phase detection with machine learning.
• To improve detection accuracy and completeness by integrating multiple machine learning models through ensemble learning.
• To uncover seismic patterns overlooked during manual inspection, particularly in high-magnitude or complex aftershock sequences.
• To develop a scalable, efficient, and stable framework suitable for real-time deployment in operational seismic monitoring systems.

Proposed method

• The EL-Picker framework consists of three modules: Trigger (using STA/LTA for initial event candidate selection), Classifier (an ensemble of multiple machine learning models), and Refiner (for post-processing and improving arrival time precision).
• An ensemble learning strategy combines diverse classifiers—such as SVM, HMM, and deep neural networks—to enhance robustness and generalization across different seismic station data.
• The Trigger module uses a short-term average (STA) and long-term average (LTA) energy ratio to identify potential seismic phases, reducing computational load before machine learning inference.
• The Classifier module applies multiple models to time windows flagged by the Trigger, leveraging their complementary strengths to improve detection accuracy.
• The Refiner module fine-tunes arrival time predictions using signal characteristics and attention mechanisms, especially in low-SNR regions.
• A recurrent attention (RA)-CNN model is used to analyze spectral waterfalls and identify discriminative features in waveforms, particularly around S-phase peaks, to explain missed detections.

Experimental results

Research questions

• RQ1 Can an ensemble of machine learning models outperform individual models in detecting P-phase arrivals in real-time, continuous seismic waveforms?
• RQ2 To what extent can automated detection reduce the burden of manual labeling while improving detection completeness?
• RQ3 What structural or signal-based patterns in seismic waveforms are consistently missed during manual inspection, especially in low-SNR or high-magnitude aftershock scenarios?
• RQ4 How does the integration of a lightweight trigger with a sophisticated classifier and refiner improve detection efficiency and accuracy?
• RQ5 Can attention mechanisms in deep learning models reveal hidden regularities in seismic waveforms that correlate with missed P-phase detections?

Key findings

• EL-Picker achieved the highest detection performance among evaluated methods on the aftershock dataset of the MS 8.0 Wenchuan earthquake.
• The framework identified 120% more P-phase arrivals than the expert-labeled dataset, indicating substantial improvement in detection completeness.
• A significant number of missed arrivals were found to have low signal-to-noise ratios (SNR), particularly in the 15–25 second window after the P-phase onset, where waveforms were less distinct.
• The RA-CNN model revealed that attention mechanisms focus on S-phase peaks, highlighting that subtle P-phase signals near these regions are often overlooked in manual inspection.
• The ensemble classifier demonstrated adaptability across different seismic stations, confirming the framework’s flexibility in handling heterogeneous waveform data.
• Manual labeling tends to prioritize high-amplitude, high-SNR events, leading to systematic underdetection of low-SNR, regional aftershocks—especially after major quakes.