GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.3 GNGTS 2024 Dynamic Strain Transients Clustering Using an Unsupervised Machine Learning Strategy B. Di Lieto 1 , P. Romano 1 , S. Scarpeta 2 , A. Sangianantoni 1 , G. Messut 2 1 Isttuto Nazionale di Geofsica e Vulcanologia, Osservatorio Vesuviano, Napoli, Italy; 2 Università degli Studi di Salerno, Dipartmento di Fisica "E. R. Caianiello", Salerno, Italy. In recent years, machine learning techniques have been exploited in volcanology in order to assess natural hazards, volcano dynamics changes and early warning informaton. Among all the other approaches followed so far, unsupervised algorithms have shown to be partcularly reliable in dealing with huge datasets, thanks to their ability to exploit the underlying informaton carried by the dataset and classify data characteristcs without the need to label the training dataset. Since assigning target labels to the training dataset may be hard and tme-consuming in many cases, unsupervised strategies that exploit unlabelled data, have been successfully employed as a clustering and visualizaton tool in exploratory data analysis in a wide range of applicatons. Self- organized neural systems (SOM), specifcally, have the intrinsic capability to analyze large sets of high-dimensional data and can be implemented in an online learning manner. A SOM algorithm was successfully applied to classify VLP events recorded from a borehole strainmeter at Stromboli volcano during the explosive sequence that occurred during the summer of 2019, when two distnct paroxysms, happened about a month and a half apart, violently shook the volcano. Stromboli is an actve, open-conduit strato-volcano, characterized by moderately persistent volcanic actvity with a paucity of deformaton episodes, always a candidate as a natural laboratory for researchers investgatng eruptve precursors on open-conduit volcanoes. Following recent research, data recorded from borehole strainmeters carry several pieces of informaton inherent the statc and dynamic deformatons, due to the intrinsic capability of the instrument of recording high precision data within a wide frequency range. The extension of the tme period previously examined, from 2018 to 2020 (fg. 1), has led us to fnd other correlatons between observed phenomenologies and VLP shape variatons. Valuable informaton is embedded in the data used in the current work, which could be used not only for scientfc purposes but also from civil protecton agencies. Such a variety of possible usage needs the setng of principles and legal arrangements to be implemented in order to ensure that data will be properly and ethically managed, used and accessed from the scientfc community.