Generalized Volcanic Eruption Forecasters Using Seismic Data and Machine Learning: Examples in Magmatic and Phreatic Eruptions

 https://agu.confex.com/agu/fm23/meetingapp.cgi/Paper/1361493(external link)

Volcanic eruptions pose a significant threat to human life and infrastructure, and accurate forecasting is challenging due to limited data availability for certain volcanoes. In this study, we explore using well instrumented volcanoes to train machine learning forecast models with seismic data, demonstrating their potential to generalize precursor information and expand applicability to other volcanoes. By testing on over 10 volcanoes, we assess the reliability of volcanic forecasters and their discrimination ability in distinguishing imminent eruptions from non-eruptive periods using extended seismic records. Through transfer learning, our models successfully identified eruption precursors for volcanoes with limited eruptive records and issued warnings for previously unobserved eruptions.

We categorized the volcanoes into three groups (Fig): a Magmatic group consisting of volcanoes in Alaska and Kamchatka, sharing a common tectonic origin, region, and eruption style; a Phreatic group comprising eruptions from New Zealand's Whakaari, Ruapehu, and Tongariro volcanoes. Additionally, we created a mixed pool containing all the volcanoes in the study.

We define a classification problem where we are interested in signals occurring 48 hours which focuses pattern recognition algorithms onto signals occurring prior to eruptions. Our focus is on distinguishing eruptive and non-eruptive data using a random forest model composed of decision trees trained on various data subsets. For testing, a cross-validation strategy is employed, where eruptions are split into testing and training sets, and model performance is averaged across out-of-sample eruptions.

Overall, all our models exhibit good performance before out-of-sample eruptions (model response in eruptions not part of the training) and show good discriminability. These results highlight the viability of machine learning in volcanic eruption forecasting and emphasize the need for further research and data collection to improve their effectiveness.