Automated seismo-volcanic event detection applied to popocatépetl using machine learning

Abstract

Popocatépetl, one of Mexico's most active volcanoes, exhibits high seismicity since its reactivation in 1994. Identifying and classifying its different seismo-volcanic events is crucial for understanding the volcano's dynamics. Machine Learning (ML) methods have proven effective in this task; however, their accuracy often relies on large-scale labeled datasets and they are typically tested on single stations. This may limit a broader applicability across seismic networks monitoring volcanoes. Here, we present an updated ML-based workflow for the automated detection and classification of long-period (LP) events, tremors (TR), and volcano-tectonic (VT) earthquakes at Popocatépetl, using continuous seismic recordings from 2019 to 2023. The workflow leverages data from a network of up to 19 seismic stations, enhancing event classification with a limited labeled dataset and improving reliability across multiple stations. The workflow is divided into two stages: the first stage generates LP and TR catalogs by training a classification model using Popocatépetl's data. The second stage creates the VT catalog using a pre-trained phase-picking and phase association models, alongside standard seismological methods for event location. The automatic catalogs generated by our workflow accurately captured the temporal and spatial trends of seismicity at Popocatépetl over more than four years, including periods of increased volcanic activity. Our approach also identified additional events not reported in manual analyses, improving the detection of trends related to volcanic processes, such as activity related to dome emplacement, periods of explosive activity, and potential system pressurization.