By Paul Voosen, covering Earth and Planetary Science, Science
Satellites are providing torrents of data about the world’s active volcanoes, but researchers have struggled to turn them into a global prediction of volcanic risks. That may soon change with newly developed algorithms that can automatically tease from that data signals of volcanic risk, raising the prospect that within a couple years scientists could develop a global volcano warning system.
Without such tools, geoscientists simply can’t keep up with information pouring out the satellites, says Michael Poland, the scientist-in-charge of the U.S. Geological Survey’s Yellowstone Volcano Observatory in Vancouver, Washington, who was not involved in either study. “The volume of data is overwhelming,” he says.
Andrew Hooper, a volcanologist at the University of Leeds in the United Kingdom who led the development of one method, says the new algorithms should benefit the roughly 800 million people who live near volcanoes. “About 1400 volcanoes have potential to erupt above the sea,” he says. “About 100 are monitored. The vast majority aren’t.” Both methods were presented this week (Dec. 10, 2018) in Washington, D.C., at the semiannual meeting of the American Geophysical Union (AGU).
Over the past few years, with the launch of the European Space Agency’s satellites Sentinel 1A and Sentinel 1B, the field of volcanology has received frequent, repeated views of how the ground shifts around the world’s volcanoes. The Sentinel 1 satellites use a technique called radar interferometry, which compares radar signals sent to and reflected from Earth to track changes in the planet’s surface. The method isn’t new, but, uniquely, the Sentinel 1 satellites revisit each spot on the planet once every 6 days, and the Sentinel team releases those high-resolution observations rapidly. A research group in the United Kingdom called the Centre for Observation and Modelling of Earthquakes, Volcanoes and Tectonics (COMET) had already begun to create a database of these ground-movement snapshots, called interferograms, for the world’s volcanoes. Overlaying this database with automated detection seemed natural given the success machine learning has had in other forms of pattern detection, says Hooper, who works with COMET.
Changes in ground motion typically reflect magma shifting beneath the volcano and do not perfectly predict eruptions. But unlike thermal hot spots or ash plumes, which can be automatically detected with weather satellites, land shifts can help predict eruptions, not simply indicate their occurrence. “Deformation doesn’t always mean eruption,” Hooper says. “But there are few cases where we don’t have an eruption without deformation.”
First, the teams had to teach their algorithms not to confuse atmospheric shifts for ground motion, something interferograms are prone to do. To do that, Hooper’s team settled on a technique called independent component analysis, which learns to break apart a signal into different pieces: such as stratified atmosphere or short-term turbulence, along with ground shifts in a volcano’s caldera or flank. The technique allows them to catch both brand-new ground motions, or changes in rate, both of which can be signs of pending eruption.
Meanwhile, another COMET team led by Juliet Biggs, a volcanologist at the University of Bristol in the United Kingdom, has built a second algorithm using a increasingly popular form of artificial intelligence called convolutional neural networks, which use layers of biologically inspired “neurons” to break apart features of images into ever-more-abstract pools, learning how to tell, for example, cats from dogs. The researchers first trained their neural network using raw interferograms from Envisat, Sentinel’s precursor, for which they had existing examples of eruptions. Although the algorithm had some success on an analysis of 30,000 Sentinel interferograms, it still produced too many false positives. There were simply too few examples to learn from, says Fabien Albino, a volcanologist who works with Biggs at Bristol. “For machine learning, 100 is nothing. They want thousands and thousands.”
To overcome that problem, Biggs and her colleagues create a synthetic data set of computer-simulated eruptions, generated for a few known physical patterns. These synthetic data dropped the fraction of false positives from some 60% to 20%, as they reported today at the AGU meeting. That trend will only continue to get better as more Sentinel examples are poured into the algorithm, Albino says. “The system is just going to tune like Google, [inputting] millions of cats and dogs, and afterward the system knows. It doesn’t have to learn anymore. It’s stable.”
Read the source article in Science.