Communications of the ACM
Predicting When Volcanoes Will Blow
Researchers at the University of Canterbury in Christchurch, New Zealand, were motivated to develop a real-time warning system that uses machine learning after the White Island volcanic eruption in New Zealand in 2019 killed 22 people.
Credit: Unsplash
There are about 500 volcanoes worldwide that could potentially erupt while over 500 million people live close to a volcano, according to Active Volcanoes of the World, a book series of the International Association of Volcanology and Chemistry of the Earth's Interior (IAVCEI). Predicting eruptions, however, can be difficult, since the behavior of many volcanoes still is not well understood.
Furthermore, developing predictions can be a slow process, since it typically involves arriving at a consensus of experts who interpret real-time data from cameras and sensors. "It takes a lot of time to get people in a room and get them to agree," says David Dempsey, a senior lecturer in the Department of Civil and Natural Resources Engineering at the University of Canterbury in Christchurch, New Zealand.
Machine learning is now being explored to develop alternative ways of forecasting eruptions. An automated system could constantly analyze incoming data, speeding up the process. And while experts typically look for particular signs of an eruption based on prior experience, machine learning could uncover new indicators. "The nice thing about a machine learning approach is it's pretty ruthless," says Dempsey. "It just looks at all of the signals and it says, 'these are convincing on a statistical basis, they occur more frequently prior to eruptions than during non-eruptive periods'."
Dempsey and his team were motivated to develop a real-time warning system that uses machine learning after the White Island volcanic eruption in New Zealand in 2019 that killed 22 people. A human-based system currently used for that area provides an alert level, varying from 0 to 5 depending on the degree of unrest, which is updated every few weeks or months. Before the 2019 eruption, the alert level did indicate the volcano was at its most hazardous state, but many tour operators still took the risk and brought tourists to the site. Dempsey thinks an automated system could help support existing methods. "It would essentially provide you with real-time commentary of what's happening at the volcano, is this similar or not to previous eruptions, and should you be worried on that basis," he says.
To create such a model, the team used seismic data captured from the White Island volcano, just one kilometer away from the main vents, from January 2011 to January 2020. The data was processed to extract signals coming from certain depths which are relevant for predicting an eruption. During that period of time, the volcano had erupted five times, signals preceding each eruption were used to train a machine learning model called Random Forest, to see if it could detect any patterns of activity that were similar across eruptions.
The model was trained on data related to four eruptions, then tested on the fifth; it was able to 'predict' four out of the five eruptions, which were of a similar nature, by recognizing a four-hour-long burst of seismic energy that occurred either days or hours before each eruption. Dempsey thinks that represents pressure starting to build up inside the volcano prior to an eruption. "It's very much something that the machine learning model recognized by itself," he says.
When the model makes a prediction, it also provides a degree of suspicion ranging from 0 to 100, depending on how sure it is that there will be an eruption. One of the challenges for Dempsey and his team was to decide on the threshold for the degree of suspicion that would trigger an alert, if it was deployed to monitor a volcano. The team decided on a value of 80, which means there is a 1 in 12 chance an eruption will occur during an alert. Using this threshold, the model would have been able to predict the fatal 2019 eruption about 17 hours beforehand.
The team has been testing its warning system since February 2020, constantly pulling in the latest data from seismic stations on White Island. If predictions reach a high value, it sends emails to people monitoring the volcano.
Dempsey and his colleagues think the model's performance can be improved by incorporating seismic data from other stations on the island, which could help pinpoint where exactly beneath the volcano there is a lot of activity. The team is also developing a filter that can separate out earthquake signals, since they often interfere with a volcano's seismic data, even when an earthquake is up to 100 kilometers away.
Another team used machine learning to try to uncover seismic patterns related to eruptions at Piton de la Fournaise on Réunion Island in the Indian Ocean, one of the most active volcanoes in the world. Christopher Ren, a researcher at Los Alamos National Laboratory in New Mexico, and his colleagues focused on volcanic tremor, a continuous seismic signal that typically occurs when a volcano is in an eruptive state. As magma, or molten rock, starts filling up the hollow chamber inside a volcano, it emits certain resonant frequencies linked to magma movement. "The tremor will start slow and build up and then above a certain level, that's officially when an eruption is happening," says Ren.
Machine learning algorithms were trained using data captured continuously over six years' time from a network of seismometers around the volcano. A supervised learning approach, where data was labelled to highlight when eruptions occurred, was first used to try to identify differences in volcanic tremor between eruptive and non-eruptive periods. Then, an unsupervised learning method called spectral clustering was used to try and determine similarities and differences between eruptions, as well as different phases of an eruption.
Ren and his team found the first model could pick out eruptive periods very well, where the tremors it learned to associate with eruptions were in a consistent frequency range across different events. The model was tested by using part of the dataset that was withheld during training. "It accurately picked out most of the eruptive days in the blind test that it didn't have access to," says Ren. "I was pretty happy."
The unsupervised approach was able to pick out different phases of an eruption, which was confirmed by looking at historical records of the types of signals that characterize each stage. It was also able to recognize characteristics particular to different types of eruptions, such as an unusual event in 2015 in which a great deal of lava flowed out.
These models could be developed further to help predict if an eruption might occur. Similar to Dempsey's model, they would output a degree of suspicion that an eruption was imminent, and they would have to decide what threshold would warrant alerting people nearby. "Where people's lives are at risk, there's a lot of policy decisions that need to come into play," says Ren. "It's absolutely something that needs to be thought about when deploying the algorithm to a real-world application."
Ren and his colleagues would like to incorporate other types of data into their models, which could better capture a volcano's behavior during an eruption. A technique called Interferometric synthetic-aperture radar, or InSAR, which involves using radar images of the Earth's surface captured by satellites, could help spot tiny deformations (often a few centimeters long, or less) on a volcano's surface. Small amounts of swelling or cracking, for example, relate to changes inside a volcano, such as the movement of magma, which can be linked to eruptions.
Ren is particularly interested in using InSAR to see if it can spot early signs of an eruption. A volcano typically expands before an eruption, but the onset of such inflation is hard to identify from seismic signals alone, because of noise. "It would be really interesting to see whether there's any signal associated with the very beginnings of inflation," says Ren. "We're going to look at whether we can fuse information from seismic data and satellite imagery to characterize the volcano even better."
Sandrine Ceurstemont is a freelance science writer based in London, U.K.
No entries found