Back to Bigger Picture

BIGGER PICTURE Disaster preparedness

Listening for a volcano’s change of rhythm

Researchers used continuous seismic records to search for eruption precursors at Whakaari. The most useful follow-up tested what could fool the system in real time.

Whakaari / White Island seen from Ohope Beach in November 2017. This photograph shows the volcano studied, not an AI warning or the 2019 eruption.Original photograph
Whakaari / White Island seen from Ohope Beach in November 2017. This photograph shows the volcano studied, not an AI warning or the 2019 eruption.Krzysztof Golik / Wikimedia CommonsCC BY-SA 4.0
Historical test
Four of five eruptions identified in the 2020 study
Prospective evaluation
18 months during 2020–2021
Important finding
Three false alerts traced to regional earthquakes

How the idea works

  1. Read continuous tremor

    Seismic sensors record changing ground motion.

  2. Find precursor patterns

    Learned time-series patterns produce an eruption score.

  3. Test and calibrate warnings

    Real-time results expose noise and inform probability estimates.

The ground carries more than one kind of signal

Whakaari’s sudden steam-driven explosions make short-term forecasting difficult. Researchers at the University of Auckland examined years of seismic records for patterns preceding eruptions. The question was whether continuous tremor measurements contained useful warning information that an algorithm could recognize repeatedly, rather than relying only on occasional assessments of the volcano’s general state.

Sources: [1][3]

A promising result came from looking backward

The 2020 work identified precursors in four of five historical eruptions. The university reported that the December 2019 event could have received 16 hours’ warning under the tested system. That is a retrospective result: the warning was not issued then. It suggests useful information existed in the record, not that future eruptions can be predicted with a guaranteed lead time.

Sources: [1][3]

The real-time test found a different problem

A follow-up study ran the forecaster from February 2020 to August 2021, updating it every ten minutes from recent seismic data. No eruption occurred, but three false alerts did. Regional earthquakes had contaminated the tremor signal. Researchers developed filtering and normalization changes that would have prevented those particular false positives in the reanalysis.

Sources: [3]

Learning what fools a warning is progress

The 2022 paper also distinguishes a model’s numerical score from a calibrated eruption probability. That matters when people must act on a warning. This remains evaluated research, not proof of lives saved or a substitute for official volcanic guidance. Its inspiring contribution is a measurable path from hidden signals to testing, error diagnosis and a more interpretable decision aid.

Sources: [3]

The original sources

  1. University of AucklandEruption alert system would have given 16 hours’ warning at Whakaari
  2. Nature CommunicationsAutomatic precursor recognition and real-time forecasting of sudden explosive volcanic eruptions at Whakaari, New Zealand
  3. Dempsey et al., Bulletin of Volcanology; author-hosted published PDFEvaluation of short-term probabilistic eruption forecasting at Whakaari, New Zealand

What to explore next

Mapping the flood even when clouds hide the groundTwenty-five minutes back from the science planning pile
Explore all 99 applications