GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 1.1 GNGTS 2024 sequence is a signifcant case study for testng and refning automated techniques to detect and locate microearthquakes using machine learning. The RSC is composed of 10 statons and has monitored microseismic actvity potentally induced by underground gas storage actvites since 2012. Rigorous manual processing conducted by the RSC, involving daily and monthly ofine procedures, is undertaken to guarantee data accuracy and metadata reliability. Nevertheless, this approach can be tme consuming and demanding, partcularly in densely populated seismic sequences where very fast analysis is preferable. Comparing seismic catalogs derived from associatons approved by experienced analysts and revised manual picks with those generated using the PhaseNet integrated with LOC-FLOW provides a unique opportunity to evaluate the performance of ML methods in detectng and localizing local microearthquakes. Our partcular focus includes: 1. PhaseNet phase picker performance: we examine the efectveness of the PhaseNet phase picker in comparison to manual phase picks. This evaluaton aims to assess the accuracy and reliability of PhaseNet's automated phase detecton. 2. LOC-FLOW-generated earthquake catalogs: we analyse the earthquake catalogs produced by LOC-FLOW, evaluatng both origin tme and absolute locatons. The examinaton involves comparing catalogs formed with PhaseNet picks against those created with the original RSC manual picks. This analysis ofers insights into the coherence and efcacy of PhaseNet in creatng the overall earthquake catalog. 3. Contributon of template matching: we assess the infuence of template matching on the fnal catalog and compare it to the dataset obtained from the original RSC procedures. This evaluaton aims to clarify the degree to which template matching enhances the accuracy and comprehensiveness of the earthquake catalog within the LOC-FLOW workfow. 4. Spato-temporal characteristcs of seismicity: furthermore, we assess the spato-temporal characteristcs of the acquired seismicity. This examinaton is useful to check the efciency of the method in discerning tectonic structures actvated during the sequence. Gaining insights into the spatal and temporal paterns of seismic actvity ofers valuable understanding of the underlying geologic processes. We fnd that PhaseNet achieved a detecton rate of 79% for manual P arrival tmes and 90% for S arrival tmes at the same statons. While P picks exhibited satsfactory accuracy, a notceable delay was observed for S picks. This delay is presumed to be a common feature, even in other datasets, given the high quality of the manual picking used for comparison. Afer integratng events identfed by the template matching procedure, the fnal LOC-FLOW catalog is characterized by an increased number of events compared to the inital manual catalog. However, in our specifc case study, PhaseNet did not contribute signifcantly to the augmentaton of the earthquake count during the most actve days of the sequence (e.g., 2-3 August), where template matching played a crucial role. Despite the observed lower accuracy in S picks,