GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.1 GNGTS 2024 Machine learning classifcaton of seismic tremor associated with the Mefte d’Ansanto CO 2 gas emission. S. Panebianco 1,2 , C. Satriano 3 , G. Vivone 2 , M. Picozzi 4 , A. Strollo 5 , T.A. Stabile 2 1 Dipartmento di Fisica e Scienze della Terra, Università di Ferrara, Italy 2 Consiglio Nazionale delle Ricerche (CNR-IMAA), Italy 3 Université Paris Cité, Insttut de physique du globe de Paris, France. 4 Isttuto Nazionale di Oceanografa e di Geofsica Sperimentale - OGS, Italy. 5 GFZ German Research Centre for Geoscience, Potsdam, Germany. The focus of this study is framed around the development of automatc methods to track the spato-temporal evoluton of deep-origin, non-volcanic gas emissions. It is widely recognized that crustal fuids play a crucial role in the earthquake nucleaton process, and the characterizaton of their emissions on the surface can be essental for beter understanding crustal source mechanisms. We investgated seismic tremors recorded at the gas-emission of Mefte d’Ansanto, situated within one of the highest seismic hazard areas of Southern Apennines, at the northern end of the fault system that generated the Mw 6.9 1980 Irpinia Earthquake. Mefte d’Ansanto is currently considered to be the largest natural emission of deep-source, non-volcanic, CO 2 -dominated gases on Earth, with an estmated total gas fux of about 2000 tons per day. We studied seismic tremor recorded between 30-10-2019 and 02-11-2019 recorded by a dense temporary seismic network (4 broadband and 7 short-period sensors) deployed around the emission area. Tremor signals were identfed by developing an automated detecton algorithm, based on non-parametric statstcs of the recorded signal amplitudes. At the same tme, we extracted signals characteristcs parameters like the RMS amplitude and the statstcal moments of amplitude distributons, both in tme and frequency domains. These were used as features for training and optmizing staton-based KNN (k-Nearest-Neighbors) binary classifers, which were then used to classify and discriminate the target tremor from anthropogenic and background noise. The trained classifers showed good performances, with a median overall accuracy of 92.8%. The comparison of the classifed tremor across all statons revealed common features: a variable duraton (16 s to 30-40 mins), a broad-frequency band (4-20 Hz) with varying amplitudes peaks at diferent statons, and two kinds of signals: (a) long-duraton, high-amplitude tremors, (b) pulsatng tremors. The higher amplitude of classifed tremors recorded at statons located close to bubbling and pressurized vents suggests the presence of multple local sources. Corresponding author: serenapanebianco@cnr.it