Automated Detection and Machine Learning‐Based Classification of Seismic Tremors Associated With a Non‐Volcanic Gas Emission (Mefite d’Ansanto, Southern Italy)

Abstract

A major aim in the study of crustal fluids is the development of automatic methodologies for monitoring deep‐source, non‐volcanic gas emissions’ spatio‐temporal evolution. Crustal fluids play a significant role in the generation of large earthquakes and the characterization of their emissions on the surface can be essential for better understanding crustal processes generating earthquakes. We investigate seismic tremors recorded over 4 days in 2019 at the Mefite d’Ansanto (southern Apennines, Italy) that is located at the northern end of the fault system that generated the Mw 6.9 1980 Irpinia Earthquake. The Mefite d’Ansanto is hypothesized to be the largest natural, non‐volcanic, CO2‐rich gas emission on Earth. The seismic tremor is studied by employing a dense temporary seismic network and an automated detection algorithm based on non‐parametric statistics of the recorded signal amplitudes. We extracted signal characteristics (RMS amplitude and statistical moments of amplitudes both in time and frequency domains) for use in the subsequent supervised machine‐learning classification of the target tremor and accidently detected anthropogenic and background noise. The data set is used for the training and optimization of station‐based KNN (k‐Nearest‐Neighbors) binary classifiers obtaining good classification performances with a median overall accuracy across all stations of 92.8%. The classified tremor displayed common features at all stations: variable duration (16 s to 30–40 min), broad peak frequency (4–20 Hz) with varying amplitudes, and two types of signals: (a) long‐duration, high‐amplitude tremor and (b) pulsating tremor. Higher tremor amplitudes recorded at stations closer to local bubbling and pressurized vents suggest multiple local tremor sources.