GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 1.1 GNGTS 2024 ML-based workfow for earthquake detecton and locaton: preliminary results from the northern Apennines with a model trained on local waveforms G. Poggiali 1 , S. Bagh 2 , L. Chiaraluce 2 , C. J. Marone 1 , Z. E. Ross 3 , E. Tint 1 , W. Zhu 4 1 La Sapienza Università di Roma, Rome, Italy 2 Isttuto Nazionale di Geofsica e Vulcanologia, Rome, Italy 3 Seismological Laboratory, Division of Geological and Planetary Sciences, California Insttute of Technology, Pasadena, CA, USA 4 University of California, Department of Earth & Planetary Science, Berkeley, CA, USA The analysis of microseismicity has a fundamental role in understanding earthquakes driving processes such as seismic sequences evoluton and preparatory phase. Recent advances in machine learning (ML)-based detecton and locaton techniques, coupled with dense seismic networks, have dramatcally increased the quantty of low-magnitude earthquakes that can be recorded and properly located. This has led to the development of a new generaton of earthquake catalogs illuminatng fault systems in unprecedented detail. We study data from the Alto Tiberina Near Fault Observatory (TABOO-NFO) located in the Apennines of central Italy where earthquake actvity is intense. This is an ideal site to apply modern detecton techniques and study in detail complex fault systems evolving in a shallow crust surrounded by deep fuid circulaton. We build an earthquake catalog for the TABOO area using data from 2010 to present using a ML-based workfow tailored to this area. We started from an updated version of the deep learning phase picker PhaseNet (Zhu and Beroza, 2019). This version of PhaseNet includes several improvements, such as polarity estmaton and beter detecton of close-in-tme events, which were ofen undetected in the original version. We trained PhaseNet on waveforms collected by the local TABOO seismic network that had been manually labelled by analysts (Cataneo et al., 2017) with P, S phases and frst arrivals polarity. We show that training the model with local data results in a signifcant improvement mainly in the picking accuracy of S-phases and also in polarity estmaton. The result of the detecton step is a massive phase dataset composed of tens of millions of P and S phases making the associaton (e.g., binding) of events a challenge. For this task we used the Gaussian Mixture Model Associaton (GaMMA, Zhu et al., 2021) algorithm, which treats earthquake phase associaton as an unsupervised clustering problem in a probabilistc framework to estmate earthquake (preliminary) locatons, origin tmes and magnitudes. We show the