Skip to content

Random Forest classificators for Stromboli volcano (Italy) to detect periods of higher probability of major or paroxysmal eruptions

Laura Sandri 1, Alexander Garcia1, Alberto Ardid2

  • Affiliations: 1Istituto Nazionale di Geofisica e Vulcanologia, Sezione di Bologna, Bologna, Italy 2University of Canterbury, Christchurch, New Zealand

  • Presentation type: Poster

  • Presentation time: Monday 16:30 - 18:30, Room Poster Hall

  • Poster Board Number: 153

  • Programme No: 3.1.45

  • Theme 3 > Session 1

Abstract

Machine Learning (ML) is nowadays one of the most rapidly advancing and widely used techniques to analyse massive sets of data from complex processes, in search of regularities or common and repeated patterns that cannot be recognized by human eye or modelled by analytical formulation. One of the most common application of ML in volcanology, to date, has been in terms of classification of pre-eruptive versus non-eruptive time periods. To this end, time series of Real-Time Seismic Amplitude (RSAM) in different frequency bands have been engineered to define a large set of features characterizing time-periods, and ML methods have been trained and tested for their skill in recognizing in advance the imminence of relevant events, such as eruptions or phreatic explosions. In this contribution we show the preliminary application and results of a ML algorithm called Random Forest (RF) to data from Stromboli volcano, in order to recognize periods of higher probability of occurrence of large explosive events. First, we trained and tested the RF on continuous amplitude RMS measurement data that are available over a period of approximately 8 years, in which about 30 major eruptions and/or paroxysms have occurred. This number of target events (30 events) enabled a better-than-usual testing of the RF skills. Secondly, focussing on a shorter period of time in 2020-2021 during which approximately 8 target events have occurred, we exploited an available dataset of multiparametric time series of seismic, deformation and gas observables, to train and test the performance of a multiparametric RF algorithm.