Machine learning event detection workflows in practice: A case study from the 2019 Durr\"es aftershock sequence

TL;DR

This study compares machine learning workflows for seismic event detection during the 2019 Durrës aftershock sequence, demonstrating their effectiveness and consistency with manual methods in practical earthquake monitoring.

Contribution

It applies and benchmarks two advanced ML-based seismic detection workflows on real aftershock data, highlighting their advantages over traditional manual processing.

Findings

ML methods detect more events than manual pickers.

Results are consistent with manual picks, with minimal bias.

ML approaches improve detection during quiet seismic periods.

Abstract

Machine Learning (ML) methods have demonstrated exceptional performance in recent years when applied to the task of seismic event detection. With numerous ML techniques now available for detecting seismicity, applying these methods in practice can help further highlight their advantages over more traditional approaches. Constructing such workflows also enables benchmarking comparisons of the latest algorithms on practical data. We combine the latest methods in seismic event detection to analyse an 18-day period of aftershock seismicity for the $M_{w}$ 6.4 2019 Durr\"es earthquake in Albania. We test two phase association-based event detection methods, the EarthQuake Transformer (EQT; Mousavi et al., 2020) end-to-end seismic detection workflow, and the PhaseNet (Zhu & Beroza, 2019) picker with the Hyperbolic Event eXtractor (Woollam et al., 2020) associator. Both ML approaches are benchmarked against a data set compiled by two independently operating seismic experts who processed a subset of events of this 18-day period. In total, PhaseNet & HEX identifies 3,551 events, and EQT detects 1,110 events with the larger catalog (PhaseNet & HEX) achieving a magnitude of completeness of ~1. By relocating the derived catalogs with the same minimum 1D velocity model, we calculate statistics on the resulting hypocentral locations and phase picks. We find that the ML-methods yield results consistent with manual pickers, with bias that is no larger than that between different pickers. The achieved fit after relocation is comparable to that of the manual picks but the increased number of picks per event for the ML pickers, especially PhaseNet, yields smaller hypocentral errors. The number of associated events per hour increases for seismically quiet times of the day, and the smallest magnitude events are detected throughout these periods, which we interpret to be indicative of true event associations.