Improving microearthquake detection in the Val d’Agriregion (Southern Italy) with deep learning

The detection and monitoring of low-magnitude earthquakes are crucial for situational awareness and risk assessment. We employ two advanced methodologies for seismic arrival time picking, detection, and localization of microseismicity in the Basilicata region (southern Italy). Both approaches rely on deep neural networks for detecting and picking P- and S-wave arrivals. This region exhibits complex seismicity due to tectonic setting, reservoir impoundment, and hydrocarbon extraction, as it hosts Europe’s largest onshore oil field and a dammed water reservoir. We compare our results with a reference catalog based on the classical short-time average over long-time average (STA/LTA) method and analyst reviews. The machine-learning-based catalogs identify approximately twice as many earthquakes as the reference bulletin, with recall rates (indicating the proportion of retrieved events also present in the reference catalog) of 93% and 77%, respectively. Our findings demonstrate that deep learning significantly improves the magnitude detection threshold while ensuring high reliability. A significant advantage is the fully automated and rapid workflow, which produces a homogeneous catalog and can be integrated into near-real-time seismic monitoring. These tools thus provide valuable advancements in earthquake detection and sequence analysis.