Tsunamis trigger atmospheric disturbances that are picked up by GPS satellites – and an AI-powered monitoring system that detects the signals could alert us before the tsunami reaches coastal areas.
A better early warning system for tsunamis could become a reality with the help of artificial intelligence. Researchers have shown how widely available AI technology can detect the subtle disturbances in the atmosphere caused when a tsunami’s destructive waves begin to form far from the shore – a demonstration that could help provide earlier warnings for coastal communities long before the tsunami reaches them.
 |
| Flooding caused by a tsunami that hit the coast of Sri Lanka in 2004 DigitalGlobe via Getty Images |
“There is no global network for detecting tsunami waves, and installing physical hardware, like buoy-based systems, is expensive,” says Valentino Constantinou at Terran Orbital Corporation, a satellite manufacturing company based in Florida. “But we know that small satellite constellations are just proliferating everywhere.”
Those GPS satellites and other global navigation satellite systems are important because they constantly exchange radio signals with ground stations. Crucially, the speed of the radio signals is affected by the density of charged particles in an area of Earth’s ionosphere some 300 to 350 kilometres above the planet’s surface. Tsunami-triggered shock waves travelling up into the atmosphere affect the density of these particles, generating small but measurable changes in the satellite radio signals.
In 2017, research groups at NASA’s Jet Propulsion Laboratory in California and the Sapienza University of Rome in Italy developed a computer algorithm for measuring changes in the density of charged particles in the ionosphere caused by the formation of a tsunami. Constantinou and his colleagues transformed the data produced through that technique – a one-dimensional line showing changes in charged particle density over time – into two-dimensional images that can be analysed by off-the-shelf AI models. They then tasked the AI with identifying tsunami-related features within the images.
Constantinou and his colleagues trained and tested the AI on data from three earthquake-triggered tsunamis – one that struck Chile in 2010, a 2011 event in Japan, and an event that occurred off Canada’s west coast in 2012. The team then validated the AI’s performance on data from a fourth tsunami, this one triggered by the 2015 Illapel earthquake off the coast of Chile. This process allowed the researchers to see how well the AI could distinguish tsunami-related disturbances from normal background noise in the ionosphere.
However, the researchers were concerned that the AI might falsely detect periods of tsunami-related disturbances that did not exist. To reduce the possibility of this happening, they filtered out any ionospheric disturbance patterns that were not detected by at least 70 per cent of ground stations in contact with the individual satellites passing by overhead.
This approach produced “pretty good results” with a reported detection performance of more than 90 per cent, says Quentin Brissaud at NORSAR, a seismic research foundation in Norway, who was not involved in the study. But he says it remains to be seen if this AI performance based on data from just four tsunami events can lead to accurate detection of a more diverse set of tsunamis.
Initial research into atmospheric disturbances generated by tsunamis was spurred by the 2004 Indian Ocean earthquake and tsunami that caused billions of dollars in damages and killed about a quarter of a million people worldwide, along with the 2011 Tohoku earthquake and tsunami that killed more than 18,000 people and triggered a nuclear disaster in Japan. But the rarity of huge tsunamis makes it challenging to analyse and predict such events, says Brissaud.
A truly effective tsunami detection system would also require international cooperation to share data from satellite constellations. “Data is often held by governments or commercial partners operating the satellites,” says Constantinou. “There is no one place to grab the data for a global system.”
Reference: