Of the 1,500 active volcanoes worldwide, about 6 percent of them erupt each year, or 50 to 85. Less than half of all volcanoes have sensors, and even fewer are considered well-monitored, the result of high costs and difficulty in maintaining equipment in such unforgiving environments. Volcanoes that are considered dormant rarely have any monitoring, despite surprises like the 2008 eruption of the Chaitén volcano in Chile after 8,000 years of inactivity.
Now, volcanologists are turning to satellite imagery and artificial intelligence to keep a closer eye on more volcanoes and, eventually, forecast eruptions. MOUNTS (Monitoring Unrest from Space), currently tracks 18 volcanoes, including Mount Fuego in Guatemala and Mount Etna in Italy.
With 800 million people living within 62 miles of an active volcano, there are plenty of reasons to increase monitoring.
Humans have settled around volcanoes throughout history, even if they are dangerous, largely due to volcanic soil being so fertile, according to Sébastien Valade, the project lead of MOUNTS and a researcher at the Technical University of Berlin (TU Berlin).
A recent paper in Remote Sensing describes how MOUNTS uses multiple sets of satellite data to provide continuous monitoring of volcanoes, looking for changes in signals that usually mean a change in volcanic activity. These signals—ground deformations, gas emissions, and temperature increases—are detectable at different wavelengths across the electromagnetic spectrum, and they are all captured by different satellite images.
But such a large dataset, and one that is updated regularly, is difficult to go through manually. So part of the initial phase of the project was to see if machine learning algorithms could be integrated into the data analysis process. Researchers developed an artificial neural network, a type of AI, to automatically detect large deformation events, a sign that magma is moving underground. The neural network could compare two photos from different days and spot any changes.
"We don't want to monitor them all the time," says Andreas Ley, a TU Berlin researcher who worked on MOUNTS. "We want the system to tell us when something interesting is going on."
MOUNTS has already been able to spot signs from several recent eruptions, including the Erta Ale eruption in Ethiopia, as well as eruptions in Hawaii and Italy. The system also detected a deformation related to the eruption last month at Reunion Island's Piton de la Fournaise, and sent automated email alerts to users who had signed up for updates. As of now, all MOUNTS data is freely accessible online for any researchers to tap into.
But the project and future use cases have problems to overcome. The AI can get confused by how water vapor in the atmosphere distorts satellite images, as well as when it comes to comparing photos of the same area but over different seasons and with varying tree coverage. It also doesn't remove the need for the expensive ground sensors.
While satellites provide monitoring when ground-based sensors aren't available, they don't cover every situation. Some eruptions can be detected two years in advance; others may only show signs 10 minutes beforehand, and satellite images can't spot that in time to evacuate or warn residents.
Still, it's early days regarding what AI can do for volcano monitoring and our knowledge of eruptions. The next step for MOUNTS is to use AI to combine the data from gas emissions, temperature increases, and ground deformations to see if such integration can create better predictions.
"We are trying to go to AI to detect more complex signals that are trickier to recognize without AI," Valade says.
Eyes where there were none
Much of that progress relies on free images provided by government satellites like the European Sentinels and the American Landsats. Previously, satellite images were expensive, forcing researchers to concentrate on just a couple of volcanoes. Now, it's possible to look at a lot more, according to Fabien Albino, a researcher from the University of Bristol in the U.K. who worked on a project that ran 30,000 images of 900 volcanoes through a machine learning algorithm and was able to winnow the shots down to a hundred that possibly showed ground deformations.
"It's much more than we were able to do in the past," Albino says.
These satellite-based projects could even help developing countries, which typically have fewer ground-monitoring stations. An alert based on data from satellite images could let countries deploy sensors and more ground monitoring when necessary, allowing better planning of evacuations. Albino said his team is working with colleagues in Ethiopia and Ecuador to figure out what sort of information would allow the countries to better react to potential threats.
"We don't have the right to say to the local authorities to evacuate or not," he says. "But we can give a better understanding."