Skip to content

Optimising timescales for machine learning-based eruption forecasting: Insights from Miyakejima Volcano, Japan

Julie De Groote , Benoit Taisne, Susanna Jenkins

  • Affiliations: Earth Observatory of Singapore, Asian School of the Environment, Nanyang Technological University, Singapore

  • Presentation type: Talk

  • Presentation time: Monday 11:15 - 11:30, Room S160

  • Programme No: 2.4.10

  • Theme 2 > Session 4

Abstract

Volcanic eruption forecasting has the potential to save lives and contribute to community preparedness in volcanic areas. One of the most important aspects to forecast is the timing of eruption onset on short-term (days to months) timescales. This work uses timescales to optimise machine-learning models that forecast eruption onsets from monitoring data. For this, we evaluate forecasting model performance while adjusting the length of two key windows: the "data window" (period of data collection, used by forecasting models to detect signs of impending eruptions) and "forecasting window" (period over which models estimate eruption likelihood) -- each between 3 and 90 days -- to understand the influence of timescale selection on short-term eruption onset forecasting. We applied this method on processed seismic data for Miyakejima volcano (Japan). Preliminary tests on individual eruptions revealed interesting patterns, including an inverse relationship between optimal data and forecasting window lengths (i.e., longer forecasting windows must pair with smaller data windows, and vice versa, to improve performance) for models tuned to the January 2008 eruption; and distinct optimal window combinations for other data-/eruption-cases. These results (1) suggest that model performances may vary with considered timescales, underscoring the importance of conscious data- and forecasting-window choices in future work, and (2) open up possibilities of earlier notice for imminent eruptions and enhancing forecasts at Miyakejima and beyond. We are currently retuning the models to forecast multiple eruptions. Further research will examine factors that contribute to the success of particular data-/forecasting-window combinations, and their potential to generalise across volcanoes.