How Much Warning Is Worth a False Alarm? Machine Learning and New Zealand's Sudden Eruptions

When Whakaari/White Island erupted suddenly in December 2019, it killed 22 people and severely injured 25 more, New Zealand’s deadliest volcanic disaster in recent history. Writing in The Conversation, Alberto Ardid, a lecturer at the University of Canterbury, and colleagues make a plain point about that day: it was not a freak event. Over the past 20 years, they note, New Zealand has experienced “another half-dozen sudden volcanic explosions with the potential to kill or injure people nearby.” Many were near-misses only because they happened at night or when few people were present. None was forecast early enough to warn or evacuate anyone.

The danger is not unique to New Zealand. In 2014, Japan’s Mt Ontake erupted with little warning and killed 63 people, many of them hikers near the summit. The researchers’ new study asks whether machine learning could buy people more time, by detecting subtle changes in the continuous vibrations around a volcano that can signal an impending eruption hours beforehand.

The problem such systems are meant to address is a timing one. Under conventional warning systems, experts interpret complex unrest signals and assess the risk before sounding an alarm. That careful judgement is essential, but it “faces a fundamental challenge when escalation occurs over minutes or hours rather than days or weeks.” Automated warnings, drawing directly on real-time monitoring, could respond quickly when unrest suddenly changes.

To test how well that might work, the team re-analysed a machine-learning eruption forecaster at five volcanoes in New Zealand, Japan, Chile and Russia, using years of seismic data to estimate the probability of an eruption within a rolling 48-hour window. The Whakaari result gives the trade-off a concrete shape. Tested on data it had not seen during training, the forecaster anticipated four of five eruptions. The cost was around 15 days each year when a warning would have been in place without any eruption occurring.

Fifteen days of false alarms sounds like a lot. Whether it is too many, the authors argue, depends on what is at stake. They used a cost-loss model to weigh the economic disruption of precautionary action against the losses a successful forecast could avoid, and found precautionary action “could reduce preventable losses by 30 to 90 percent.” The balance differs by volcano. Closing Mt Ruapehu during the busy ski season, they acknowledge, “could come at a steep price in lost revenue and disruption,” which has to be set against the deaths and lifelong injuries a timely evacuation could prevent.

That framing has precedent in how other hazards are handled. Tsunami warnings, the researchers point out, produce far more alarms than damaging waves, yet communities keep evacuating when sirens sound, because good risk communication and public trust hold the system together. Volcano monitoring, they suggest, may need a similar tolerance for uncertainty where eruptions develop too fast for conventional processes.

The authors are careful about what they are and are not proposing. Automated warnings “would not mean replacing volcanologists.” The systems depend on monitoring data collected by organisations such as New Zealand’s GeoNet, and human expertise remains essential for interpreting volcanic behaviour and managing a crisis as it unfolds. Automation is most useful during sudden escalation, they write, “providing an initial warning while experts assess what is happening.” Earlier warnings could also help protect skifields, walking tracks, roads, power lines and nearby communities.

The researchers are equally clear about the study’s limits. Their modelling is “not a comprehensive cost-benefit analysis of operational volcano monitoring,” but a demonstration of how the cost of false alarms can be weighed against the benefit of earlier action. That places the work alongside the United Nations’ Early Warnings for All initiative, which aims to give everyone access to multi-hazard warning systems by 2027.

Their conclusion is measured rather than triumphant. For volcanoes capable of erupting with little warning, they write, “waiting for certainty carries its own risk,” and accepting more false alarms may sometimes be a reasonable price for giving people more time to get out of harm’s way. The full article, with links to the underlying research, is worth reading in The Conversation.