Submitted by Dr C.M. Martin-Jones on Tue, 01/07/2025 - 14:54
Researchers from the University of Cambridge are using AI to speed up landslide detection following major earthquakes and extreme rainfall events—buying valuable time to coordinate relief efforts and reduce humanitarian impacts.
On April 3, 2024, a magnitude 7.4 earthquake hit Taiwan’s eastern coast— the country’s strongest in 25 years. Owing to the country's earthquake preparedness, many buildings remained intact. The main threat instead came from landslides, which swept through isolated and mountainous villages, causing significant destruction.
Following the earthquake, Lorenzo Nava, who is jointly based in Cambridge’s Departments of Earth Sciences and Geography, used AI to detect about 7,000 landslides from satellite images.
In the aftermath of disasters such as earthquakes, floods and landslides, responders often turn to satellite imagery to quickly assess the extent of damage and co-ordinate relief efforts across vast and often inaccessible areas.
Mapping landslides from satellite imagery by eye can be time intensive, said Nava. “In the aftermath of a disaster, time really matters,” he said. Using AI, he identified the Taiwan landslides within three hours of the imagery being acquired.
AI is already used in landslide mapping, but Nava hopes to improve its detection capabilities by employing a suite of satellite technologies—including those that can see landslides through clouds and at night.
Rescue teams at one of the landslides following the Taiwan earthquake. Credit: Wikipedia Commons/Taitung County Government.
Multiplying Hazards
Landslides often follow major earthquakes or intense rainfall, and their occurrence can be further amplified by human activities such as deforestation and construction on unstable slopes. In certain environments, landslides may trigger secondary hazards, transforming into fast-moving debris flows or contributing to severe flooding.
Nava’s work fits into a larger effort at Cambridge to understand how landslides and other hazards can set off cascading ‘multihazard’ chains. The CoMHaz group, led by Maximillian Van Wyk de Vries, Professor of Natural Hazards in the Departments of Geography and Earth Sciences, recently collaborated with international researchers to identify the events leading up to the 2023 Sikkim flood in northeastern India. The flood happened when a landslide slipped into a glacial lake, triggering a massive outburst of water that destroyed a hydroelectric power plant and wiped-out downstream communities.
Van Wyk de Vries and team draw on information from satellite imagery, computer modelling and fieldwork to locate landslides, understand why they happen and ultimately predict their occurrence. Beyond data analysis, they are working with local communities to raise awareness of landslide hazards.
During a recent trip to Nepal, Nava and Van Wyk de Vries partnered with local scientists and the Climate and Disaster Resilience in Nepal (CDRIN) consortium in an initial effort to build a landslide early warning system for Butwal, a city lying in the shadow of a massive, unstable slope.
Nava and Van Wyk de Vries (5th and 8th from left in top image) meeting with local scientists and town planners in Butwal.
Building Trust
Effective management of disasters depends on open dialogue between scientists, local planners, and affected communities. Likewise, transparency is essential for the responsible use of AI in disaster response, said Nava—helping build trust and ensure model outputs are interpretable and actionable by decision-makers.
“Very often, the decision-makers are not the ones who developed the algorithm,” said Nava. “AI can feel like a black box. Its internal logic is not always transparent, and that can make people hesitant to act on its outputs.”
Nava is training AI to identify landslides in two types of satellite images—optical images of the ground surface and radar data, the latter of which can penetrate cloud cover and even acquire images at night.
Optical satellite images give a view of landslides over vast areas, but they can't see through cloud or at night like radar. Image: Landslide scars in Taiwan in 2010. Credit: NASA Earth Observatory.
Radar images can however be difficult for to interpret, as they use greyscale to depict contrasting surface properties and landscape features can also appear distorted. These challenges make radar data well-suited for AI-assisted analysis, helping extract features that may otherwise go unnoticed.
By combining the cloud-penetrating capabilities of radar with the fidelity of optical images, Nava hopes to build an AI-powered model that can accurately spot landslides even in poor weather conditions.
His trial following the 2024 Taiwan earthquake showed promise, detecting thousands of landslides that would otherwise go unnoticed beneath cloud cover. However, he acknowledges that there is still more work needed to improve the model’s accuracy and transparency.
“It’s important to make it easier for end users to evaluate the quality of AI-generated information before incorporating it into important decisions,” Nava added.
This is something he is now addressing as part of a broader partnership with the European Space Agency (ESA), the World Meteorological Organization (WMO), the International Telecommunication Union’s AI for Good Foundation and Global Initiative on Resilience to Natural Hazards through AI Solutions.
At a recent working group meeting at the ESA Centre for Earth Observation in Italy, the researchers launched a data-science challenge to crowdsource efforts to improve the model. “We’re opening this up and looking for help from the wider coding community,” said Nava.
Beyond improving the model’s functionality, Nava says the goal is to incorporate features that explain its reasoning—potentially using visualizations such as maps that show the likelihood of an image containing landslides to help end users understand the outputs.
“In high-stakes scenarios like disaster response, trust in AI-generated results is crucial. Through this challenge, we aim to bring transparency to the model’s decision-making process, empowering decision-makers on the ground to act with confidence and speed.”
Take part in the Challenge:
The Challenge is open to anyone with beginners-level knowledge of coding, and is ideal for anyone looking to apply their experience to. More details here.
Further reading: