Ground truth: AI can detect Earth’s movements

All around the village of Kimtang, in central Nepal, are tell-tale signs that something isn't right. There are cracks in the concrete steps of houses, trees growing at strange angles – evidence that the ground is shifting beneath the villagers' feet. The question is, just how much is the ground moving?

"This is not good," says Antoinette Tordesillas, a mathematician at the University of Melbourne, as she shows me an overhead view of Kimtang via Zoom. There's a large red spot on the image, which is not any old satellite image but a coloured map created by an artificial intelligence (AI) system.

The AI has identified a large unstable area, right beneath the village, colouring it bright red amid the dark blue of the rest of the hillside. It means the village sits right on top of a spot at high risk of a potentially devastating landslide. "Their village, where they live and farm, are actually on the slope," says Tordesillas. She and her colleagues have visited Nepal, and in some cases interviewed villagers about the slowly developing situation.

Landslides, she adds, might seem like sudden catastrophes that are impossible to predict. But satellite images captured using radar can reveal otherwise invisible signs of the ground starting to move days, weeks or even years in advance of a collapse. Granules of earth begin subtly separating from one another. "Dancers, if you like, following some kind of unwritten choreography," says Tordesillas.

In this precarious village in the mountains of Nepal, lives may be at stake. The AI-generated image is a warning. But it's also a chance.

Landslides are becoming more common, partly due to climate change, but also more direct human activities such as construction works and mining. In the US, landslides kill 25-50 people each year and cause billions of dollars in damage. Worldwide, they claim thousands of lives annually. Predicting when and where they will happen is incredibly difficult. But developments in AI are making it possible, helping geologists identify thousands of slopes around the world that are at high risk of slipping.

Nepal is home to some of the highest reaches of the Himalaya and is particularly prone to landslides and avalanches. In October 2025, a spate of landslides killed around 60 people in this mountainous country.

While it's possible to monitor such places using satellites, or ground-based sensors, observing large areas in this way generates vast amounts of data. Analysing it manually would be "beyond human capability", says Tordesillas. Fortunately, well-established forms of AI such as machine learning can do this work.

"We use our knowledge of the physics of [slope] failure to guide AI," says Tordesillas. That is to say, this is a much more specialised system. That doesn't mean it could never make a mistake – but it does mean that scientists can try to ensure that it reflects their own deep knowledge of how landslides unfold.

Kimtang's villagers may be grateful for that. In 2019, they were relocated after a landslide struck a nearby area where they had been farming. "The unfortunate thing is, this area where they've been relocated to is the most unstable part in this whole region," says Tordesillas.

Getting the data to reveal this was, in part, a stroke of luck. It came from a European satellite called Sentinel-1, which bounces radar off the ground – at roughly 2,000 flashes per second – to map terrain in fine detail. Sentinel-1 happened to fly over this part of Nepal in just the right way. "The angle of the satellite on this particular village is such that we were able to get good measurements," says Tordesillas.

The latest imagery dates from January 2025 and, thankfully, no landslide has yet occurred there in the time since. But the information and analysis it allowed the AI to perform has given Tordesillas, and her partners at other institutions, an opportunity to inform villagers of the risk, work with them to develop ways of monitoring the situation at ground-level, and plan evacuation routes or muster points.

Again, the AI map could help with that.

"The high school […] happens to be one of the most stable spots – that's something worth telling the mayor," says Tordesillas, who has also shared the AI analysis with government officials.

But this isn't the only place where AI is helping to identify communities that could be at imminent risk of a landslide.

Researchers in the UK, for example, are using AI algorithms to crunch Sentinel-1 radar imagery to analyse 300,000 slopes around the island of Great Britain. They have found 3,000 that are actively moving.

"We are talking about millimetres per year – it's something you don't notice with the naked eye," says Alessandro Novellino, a geoscientist at the British Geological Survey (BGS) who uses remote sensing to monitor for landslides. These tiny movements, however, could be a sign of a major slide likely in the future. And even if they don't lead to a major landslide, the movements over time will affect 8,700 miles (14,000km) of road and 220 miles (360km) of railway track.

It would take months or even years for humans do the same analysis that machine learning systems can do in a matter of minutes or hours, Novellino adds. "This is essentially enabling new science that we couldn't do in the past."

He stresses that the BGS does not have access to real-time data for this project since that is much more expensive. But, as in Nepal, data that is a year or two old can give you an indication of which slopes could be at risk of failing based on their historic movement.