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

Dempsey, D. E.; Cronin, S. J.; Mei, S.; Kempa-Liehr, A. W.

doi:10.1038/s41467-020-17375-2

We trained a machine learning model on nine years of seismic tremor from one station at Whakaari. It recognised four-hour tremor bursts before most eruptions. In tests on eruptions withheld from its training, it warned of four of five eruptive episodes, including 17 hours before the fatal 2019 eruption.

The problem

On 9 December 2019, Whakaari (White Island) erupted while 47 tourists and guides were on the island. Twenty-one were killed or fatally injured. Twenty-four days earlier, the volcano had been assessed at alert level 2, on a scale up to 5. The alert level classifies the current state of a volcano and gives no eruption forecast.

Alert levels are set by expert consensus, an approach that has proven useful in slow-building crises (Barberi et al., 2009; Jousset et al., 2012). For small eruptions at a hydrothermally active volcano, consensus can be too slow to issue a warning. Our goal was to test whether an automatic system could recognise eruption precursors in real time and give a fast, transparent forecast.

What we did

We used continuous data from one seismometer (station WIZ) from 2011 to 2019. Every 10 minutes we computed four measures of volcanic tremor: the average strength of shaking in three frequency bands, and a ratio between two of them. The nine years contain five major eruptive episodes and seven eruptions. One episode, in October 2013, contained three eruptions over eight days.

We divided the record into 48-hour windows that overlap by 36 hours. For each window we computed several hundred features, such as its mean, its trend, and the strength of oscillations at particular periods. Each window was labelled by whether an eruption occurred in the following 48 hours. A statistical test then identified the features whose values differed before eruptions. We treat these as candidate precursors.

Only about 1 in 400 windows was followed by an eruption. We balanced the data by randomly discarding most of the no-eruption windows. Because this step is arbitrary, we repeated it 100 times and trained a decision tree on each version. Every 10 minutes, each tree uses the last 48 hours of tremor to predict whether an eruption will occur in the next 48 hours, and the fraction of trees predicting one is the ensemble mean. An alert is issued when the ensemble mean exceeds a chosen threshold, and is withdrawn once it has stayed below the threshold for 48 hours.

We tested the model by cross-validation, withholding one month of data either side of each eruptive episode and forecasting that episode with a model trained on the rest.

What we found

At a threshold of 0.8 (80 of the 100 trees predicting an eruption), the model warned in advance of four of the five episodes, or five of the seven eruptions. It issued an alert 17 hours before the fatal 2019 eruption. Warning times for the other eruptions ranged from 7 to 75 hours. It missed the 8 October 2013 eruption, the second of three in that episode, and the April 2016 episode.

Raising the threshold reduces false alarms and increases the number of missed eruptions. This is because more trees must agree before an alert is issued. At 0.8, the model issued 58 alerts covering 8.5% of the nine years, or about one month a year, and 8.6% of them contained an eruption. At 0.93, it issued 26 alerts covering 2.5% of the period, or about nine days a year, and 15% of them contained an eruption. However, it also missed the 3 October 2013 eruption, while still warning of the 2019 eruption at least four hours ahead. To warn of five of the seven eruptions by chance, randomly placed alerts would need to cover about 70% of the period, or about eight months a year.

We also tuned a rule based on two manually selected features on all the data, with no cross-validation. Its 12 alerts, totalling 55 days, warned of six of the seven eruptions, and half of the alerts contained an eruption. However, the rule was scored on the eruptions it was fitted to. We expect it to perform poorly on future data.

The most informative precursors were bursts of tremor lasting about four hours, overlaid by a 40-minute oscillation. They preceded the 2019 eruption and both 2013 eruptive episodes. We suggest these bursts record magma rising beneath the vent and releasing hot gas into the shallow hydrothermal system. In this interpretation, the gas raises the pressure until the rock capping the system fails. Before the 2016 eruption these signals were weaker, and rock fragments thrown out by that eruption contained no magmatic material (Kennedy et al., 2020). We interpret it as a hydrothermal explosion of trapped groundwater.

Limitations

The forecaster runs in under a minute on a desktop computer and needs only one seismometer. However, it is a prototype. A stronger test is to fix the model and evaluate it on data recorded afterwards (Schorlemmer et al., 2018).

Trained only on data recorded before each eruption, the model warned of the October 2013 and December 2019 eruptions and missed the August 2013 eruption. A data-driven model forecasts only eruptions whose precursors resemble those in its training data. This probably explains the missed 2016 eruption. An ensemble mean of 0.8 also cannot be read as an 80% chance of eruption. This is because the trees were not calibrated to give probabilities. Our 48-hour window and 48-hour forecast horizon were tuned on the same nine years of data.

Who sets the alert threshold, and who has access to the forecasts, are open questions with implications for public safety, economic activity, and public trust. They need to be resolved before automatic forecasts can meet public expectations of volcano warnings.

Glossary

A steam-driven (phreatic) eruption is an explosion caused by heated water flashing to steam.

The volcanic alert level is New Zealand's numbered scale, with a highest level of 5, describing the current state of a volcano.

An eruptive episode is a single eruption or a group of eruptions close in time, which we treat as one event.

Volcanic tremor is continuous low-level shaking of the ground at an active volcano.

A hydrothermal system is a body of hot water and gas circulating through rock beneath a volcano.

A feature is a single number that summarises a segment of the tremor record, such as its mean or trend.

A decision tree is a classifier that assigns cases to classes by a sequence of threshold tests on their feature values.

The ensemble mean is the fraction of the 100 decision trees that predict an eruption in the next 48 hours.

The alert threshold is the ensemble mean above which the model issues an alert.

Cross-validation tests a model on data withheld from its training.

Abstract

Sudden steam-driven eruptions strike without warning and are a leading cause of fatalities at touristic volcanoes. Recent deaths following the 2019 Whakaari eruption in New Zealand expose a need for accurate, short-term forecasting. However, current volcano alert systems are heuristic and too slowly updated with human input. Here, we show that a structured machine learning approach can detect eruption precursors in real-time seismic data streamed from Whakaari. We identify four-hour energy bursts that occur hours to days before most eruptions and suggest these indicate charging of the vent hydrothermal system by hot magmatic fluids. We developed a model to issue short-term alerts of elevated eruption likelihood and show that, under cross-validation testing, it could provide advanced warning of an unseen eruption in four out of five instances, including at least four hours warning for the 2019 eruption. This makes a strong case to adopt real-time forecasting models at active volcanoes.