Improvements from incorporating machine learning algorithms into near real-time operational post-processing

Area of Science:

• Earth Science
• Geophysics
• Seismology

Background:

• Real-time seismic monitoring relies on automatic data analysis for event identification.
• Post-processing of small seismic events (M < 3) refines data before analyst review.
• Machine learning (ML) algorithms are increasingly viable for seismic phase picking.

Purpose of the Study:

• To implement and evaluate ML algorithms for seismic event post-processing.
• To improve the accuracy and efficiency of seismic event detection and location.
• To reduce the workload for seismic data analysts.

Main Methods:

• The Southern California Seismic Network integrated the deep-learning PhaseNet for improved phase picking.
• An automatic post-processing pipeline (ST-Proc) was developed using PhaseNet and the GaMMA ML algorithm.
• PhaseNet was compared to the traditional STA/LTA picker for pick quantity and accuracy.

Main Results:

• PhaseNet yielded 2-3 times more picks, especially S phases, with improved accuracy compared to STA/LTA.
• ML-driven post-processing led to enhanced epicenter accuracy.
• The ST-Proc pipeline achieved 65-70% event detection from sub-network triggers with a 5% false event rate.
• GaMMA-determined epicenters showed good accuracy, within a few kilometers of final locations.

Conclusions:

• ML algorithms significantly improve seismic event detection and characterization.
• Automated pipelines using PhaseNet and GaMMA streamline seismic data processing.
• These advancements reduce analyst workload and increase overall monitoring efficiency.