GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 1.1 GNGTS 2024 Comparing Machine Learning to Manual Earthquake Locaton procedures: evaluatng the performance of LOC-FLOW on a microseismic sequence occurred in Collalto area (NE Italy) M. Sugan 1 , L. Peruzza 1 , M.A. Romano 1 , M. Guidarelli 1 , L. Morato 1 , D. Sandron 1 , M.P. Plasencia Linares 1 , M. Romanelli 1 1 Natonal Insttute of Oceanography and Applied Geophysics - OGS, Italy Detectng earthquakes and picking seismic phases are fundamental elements in numerous seismological processes, being crucial in both seismic monitoring and in-depth seismological investgatons. The utlizaton of machine learning (ML) techniques has experienced a notable improvement lately, ofering a promising way to face the complexites associated with earthquake detecton and localizaton. The reliability of ML methods remains an open queston in setngs with dense and localized seismic networks. In such contexts, fast and accurate detecton and localizaton of earthquakes are essental for decision-making, playing a pivotal role in seismic risk mitgaton strategies, even for events of very low magnitude. In the feld of microseismic monitoring, ML applicatons are similar to those of earthquake monitoring, but have the task of processing weak seismic signals characterized by low signal-to- noise ratos at individual receivers or very short target tme signals (Anikiev et al., 2023). Therefore, evaluatng the performance of ML models trained on regional datasets in a microseismic sequence is challenging but crucial, especially for applicatons in the feld of induced seismicity (e.g., Mousavi et al., 2016) or for actvites of observatories near faults. This study focuses on evaluatng the performance of the PhaseNet algorithm (Zhu and Beroza, 2018), a prominent deep learning model for earthquake phase identfcaton. The evaluaton is conducted within the extensive LOC-FLOW workfow for earthquake locaton proposed by Zhang et al. (2022). The study is performed on the seismic events of the Refrontolo sequence that occurred in August 2021 on an antthetc fault segment of the Montello thrust system in the Pedemontana district of Southeastern Alps (Peruzza et al., 2022). The seismic sequence displayed remarkable actvity despite its low energy release (M L 2.5 for the main event). The sequence consisted of 374 events occurring at approximately 9 km depth within a confned volume, and was monitored by the permanent Collalto Seismic Network (RSC). This