Automatic precursor recognition and real-time forecasting of sudden explosive volcanic eruptions at Whakaari, New Zealand

Abstract:

Sudden steam-driven eruptions strike without warning and are a leading cause of death for tourists visiting volcanoes. Recent fatalities following the 2019 Whakaari eruption in New Zealand suggest there is a need for forecasts that can update on short time scales: hours to days. We used time-series feature engineering to identify eruption precursors in real-time seismic data streamed from Whakaari, New Zealand. We identified four-hour energy bursts that occur hours to days before most eruptions as particularly significant signals. We suggest these indicate charging of the vent hydrothermal system by hot magmatic fluids. Other important signals include 40-min RSAM oscillations that could indicate boiling instabilities in the hydrothermal system. We trained a model to recognise these precursors in real-time data and developed a system for short-term alerts during periods of elevated eruption likelihood. Under cross-validation (pseudoprospective) testing and an optimized alert threshold, this forecasting system could provide advanced warning of an unseen eruption in four out of five instances, including sixteen hours warning for the 2019 eruption. On the other hand, an eruption is 2016 was not caught, likely because a lack of magmatic contribution meant its precursors were unrecognized. The average alert duration for this system was five days and there was a 1 in 12 likelihood for an eruption to occur during an alert. Total alert duration was 8.5% of the study period: about one month each year on average. We suggest this level of accuracy and specificity represents an improvement over the current Volcano Alert Level system. However, future eruptions exhibiting different eruption precursors are unlikely to be detected.