Skip to content

Evaluation of a random forest model to forecast paroxysms at Volcán de Fuego, Guatemala

Amelia Bain 1,3, Oliwia Pedrycz2, Andrew Bell3, Sophie Butcher4, Silvio De Angelis5, Eliza Calder3, Amilcar Roca Palma6, Fabrizio Ponce6, Rüdiger Escobar Wolf7, Elaine Spiller8

  • Affiliations: 1Ludwig-Maximilians-Universität (LMU), München, Germany; 2University of Strathclyde, Glasgow, United Kingdom; 3University of Edinburgh, Edinburgh, United Kingdom; 4British Geological Survey (BGS), Edinburgh, United Kingdom; 5University of Liverpool, Liverpool, United Kingdom; 6Instituto Nacional de Sismología, Vulcanología, Meteorología e Hidrología (INSIVUMEH), Guatemala City, Guatemala; 7Michigan Technological University, Houghton (Michigan), USA; 8Marquette University, Milwaukee (Wisconsin), USA

  • Presentation type: Poster

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

  • Poster Board Number: 163

  • Programme No: 3.1.55

  • Theme 3 > Session 1

Abstract

The high level of activity at persistently active volcanoes, characterised by large numbers of events recorded in geophysical monitoring data, often with emergent onsets and extended durations, calls for statistical and data-driven approaches. Here, we evaluate the performance of a forecasting model applied to seismic data collected by INSIVUMEH's monitoring network at Volcán de Fuego, Guatemala. Fuego displays a background of low-level explosive and effusive activity, punctuated by more intense explosive episodes (called paroxysms) on average two to three times per year, which are frequently accompanied by pyroclastic flows. Due to the difficulty and challenges related to evacuations at Fuego, it is crucial to forecast the onset of a paroxysm as early and with as much confidence as possible. This context motivated our exploration of forecasting models based on machine learning. The selected forecasting model was developed by Dempsey et al. (2020, Nature Comms 11) to forecast sudden phreatic explosions at Whakaari volcano (Aotearoa New Zealand). We evaluate the model's performance in the very different context of a persistently active volcano with a magma-filled conduit to shallow levels. The model was first trained using 48-h data windows labelled according to whether or not a paroxysm occurred in the subsequent window, omitting 1 month of data before and after a chosen 'target' eruption. The model's forecasting performance was tested using 5, 7 and 10 labelled eruptions. We describe important differences with the Whakaari context that required special consideration, the conditions under which the model performed best, limitations, and future perspectives.