Machine learning versus manual earthquake location workflow: testing LOC-FLOW on an unusually productive microseismic sequence in northeastern Italy

Abstract

AbstractIt is an open question whether machine-learning (ML) methods can be trusted in areas where dense and localized seismic networks are in operation, and prompt and accurate detection and location of earthquakes are essential to guide decision-making processes that contribute to seismic-risk-mitigation-strategies, even for very low-magnitude events. To address these concerns, we compare the performance of a widely-used ML phase picker, PhaseNet, integrated with several popular earthquake location methods (included in LOC-FLOW), with the results obtained by the workflow adopted since 2012 by the Collalto Seismic Network, installed to monitor natural and potentially induced microearthquakes nearby an underground gas storage. The tested dataset concerns the most populated microseismic sequence observed so far (374 events, ML⩽2.5, August 2021, Refrontolo, NE-Italy), as its unusual productivity raised some criticalities in the combination of automatic routines, and time-consuming manual revision of phase picks adopted by the standard workflow. LOC-FLOW is able to detect the majority of the events listed in the manually revised catalog, demonstrating its ability to efficiently and accurately build earthquake catalogs from continuous seismic data. We highlight both the advantages and limitations of the ML-picker and recommend the use of template-matching-techniques in the final stage of processing to increase the number of events.

Keywords

• Machine learning
• microseismicity
• template matching
• seismic monitoring
• Collalto Seismic Network