Deep learning-based earthquake detection for tectonic interpretation and hazard implications

This dissertation explores the detection, location, and seismotectonic interpretation of microseismicity within the Val d’Agri region of Southern Italy. This region is characterized by both active tectonics and anthropogenic activities, such as hydrocarbon extraction and reservoir impoundment. Two deep-learning-based approaches for earthquake detection and phase identification are evaluated and compared, substantially enhancing the sensitivity and comprehensiveness of seismic catalogs compared to conventional STA/LTA methodologies. Deep-learning-based approaches enable the identification of nearly twice as many events while successfully recovering the majority of previously identified earthquakes. Additionally, they provide fully automated workflows that are well-suited for near-real-time monitoring. Building on these enhanced catalogs, I then apply waveform-based clustering and high-precision relocation to reveal the geometry of active fault structures. The clustering approach groups similar events, likely originating from the same fault segments, while the relocation highlights planar features consistent with fault planes. Moment tensor inversion of the largest events (M > 2) further constrains the faulting mechanisms, indicating predominantly SW-dipping normal faults aligned with the regional extensional regime. A thorough examination of a representative seismic sequence is undertaken, analyzing its spatial-temporal progression and identifying potential hazards. This comprehensive work presents a systematic integration of modern machine learning and advanced seismological techniques. By employing these methodologies, I have achieved significant improvements in earthquake detection capabilities. Furthermore, this research contributes to a more nuanced comprehension of active deformation and seismic hazards within an intricate tectonic environment.