NEWS

How AI and satellites could help predict volcanic eruptions

Emerging monitoring methods will allow scientists to keep an eye on many more volcanoes.

Search for this author in:

View of Mount Agung erupting in the afternoon.

Agung volcano on the Indonesian island of Bali erupts in November 2017.Credit: Donal Husni/Zuma

Volcanologists are combining satellite measurements of ground movements with artificial intelligence to more accurately monitor — and eventually predict — volcanic eruptions.

Although about 800 million people live within 100 kilometres of a volcano, very few of these potential natural hazards are monitored consistently. But emerging methods are now enabling researchers to keep a constant eye on volcanoes, says Juliet Biggs, a volcanologist at the University of Bristol, UK.

Her team will present its work, which uses machine learning to spot the formation of ground distortions around volcanoes, on 20 March at a conference in Santa Fe, New Mexico.

Biggs and her colleagues use radar observations from two satellites that make up the European Sentinel-1 mission. Depending on their location as they orbit Earth, the craft collect data on the world’s volcanoes every 6, 12 or 24 days. As they repeatedly pass over the same spot, the satellites measure the distance between themselves and the ground. This can indicate whether that distance has changed over time — such as when the ground lifts or drops as magma shifts beneath a volcano.

But there are issues with such data. Water vapour in the atmosphere can mimic the signal of shifting ground, and researchers must account for this when looking at radar observations. These atmospheric distortions are particularly problematic when scientists are trying to work in near-real time.

Seeing clearly

Biggs's team got an early glimpse of these challenges when they started studying Sentinel-1 images of the Agung volcano on the Indonesian island of Bali, in the lead up to a November 2017 eruption. Hundreds of small earthquakes had started shaking the region two months before, forcing the evacuation of 140,000 people.

Atmospheric distortions around Agung complicated the team's efforts to study ground distortions around the volcano. But once Biggs and her colleagues devised a way to correct for the atmospheric signals, they found that the ground had lifted by up to 10 centimetres on Agung’s northern flank, towards a neighbouring volcano. That ground movement was a sign that magma was probably shifting in a natural plumbing system that connects the two volcanoes, the team reported1 last month in Nature Communications. The team hadn’t tried to predict the Agung eruption, but they “learned a lot by looking at this one example”, Biggs says.

Team member Fabien Albino, a geophysicist at the University of Bristol, is now developing ways to correct for atmospheric distortions quickly with the help of a weather model that runs in near-real time.

If it predicts atmospheric disturbances in a given area, then he can identify unusual signals in the satellite radar data that might be caused by water vapour rather than volcanic unrest. The work is still in its early stages, Albino says — but it could eventually provide a way to more quickly assess what's happening in situations such as those at Agung.

Biggs and her colleagues are now pushing to monitor volcanoes more rapidly all around the globe. They have created a neural network that has churned through more than 30,000 Sentinel-1 images of more than 900 volcanoes and flagged about 100 images for closer examination. Of those, at least 39 were accurate detections of actual ground distortions, the team reported2 last year. By getting an algorithm to do the initial work of sorting through the data, researchers save time that they can better spend following up on volcanoes of interest, Biggs says.

The team is also training its neural network on synthetic data generated from simulated eruptions. That work roughly doubled the precision of the algorithm, says Pui Anantrasirichai, an electrical engineer at the University of Bristol who will present the work at the Santa Fe meeting.

Alternative methods

At the University of Leeds, a group led by geophysicist Andrew Hooper is developing another way to automatically detect potential signs of unrest. Rather than sorting the radar images from Sentinel-1 as Biggs's team does, Hooper and his colleagues use a technique that searches for changes in the satellite data3. If the ground is already deforming at a volcano, Hooper’s method can flag if that distortion starts to speed up, slow down or change in some other way. That would allow researchers to detect even small ground alterations over long periods of time.

It is a different type of analysis than Biggs's work, but both groups' ultimate goal is “to process data for all of the volcanoes, all of the time”, says Hooper.

Biggs and Hooper plan to test their approaches on a global database of volcano ground distortions, which is hosted by the Centre for Observation and Modelling of Earthquakes, Volcanoes and Tectonics, a UK collaboration. But because the database has had some technical problems, the researchers have not yet run a side-by-side comparison of their techniques.

Other scientists, such as Matt Pritchard, a volcanologist at Cornell University in Ithaca, New York, are trying to develop algorithms that can spot changes in volcanoes using other types of satellite data, such as surface temperature or ash and gas emissions4. Working with Biggs and others, Pritchard hopes to use machine-learning techniques to sift through 17 years of data from NASA’s Terra and Aqua satellites, which measure heat coming off erupting volcanoes on Earth.

But he and his colleagues are just getting started with their algorithms, which have a long way to go. For now, at least, undergraduate students are much better than the machines at picking out eruptions.

Nature 567, 156-157 (2019)

doi: 10.1038/d41586-019-00752-3

Updates & Corrections

  • Correction 07 March 2019: Pui Anantrasirichai's work training a neural network on simulated eruption data roughly doubled the algorithm's precision, not its accuracy.

References

  1. 1.

    Albino, F., Biggs, J. & Syahbana, D. K. Nature Commun. 10, 748 (2019).

  2. 2.

    Anantrasirichai, N., Biggs, J., Albino, F., Hill, P. & Bull, D. J. Geophys Res. Solid Earth 123, 6592–6606 (2018).

  3. 3.

    Gaddes, M. E., Hooper, A., Bagnardi, M., Inman, H. & Albino, F. J. Geophys. Res. Solid Earth 123, 10226-10251 (2018).

  4. 4.

    Furtney, M. A. et al. J. Volcanol. Geotherm. Res. 365, 38-56 (2018).

Download references

Nature Briefing

Sign up for the daily Nature Briefing email newsletter

Stay up to date with what matters in science and why, handpicked from Nature and other publications worldwide.

Sign Up