The Nava team has developed a machine learning algorithm that can quickly, accurately and autonomously identify landslidesEvastraw for Varsity
Landslides are an important global danger, killing thousands of people each year and affecting millions around the world. After the disaster, aid workers rely on satellite images to see where the roads are blocked and the village is blocked, but it is time to spot slides with their eyes. Now, researchers at the Ministry of Earth Science in Cambridge, led by Lorenzo Nava, are exploring how new image processing techniques using machine learning can help identify landslides quickly and accurately.
When landslides attack in mountainous and remote areas, satellite images are often the first resource for emergency workers to check. However, manual analysis of these satellite images can be laborious and time-consuming. This is just as boring and slow as modifying your Tripos exam reading. To aid in this scanning process, Nava's team has developed a machine learning algorithm that can quickly, accurately, and autonomously identify landslides. “In high-stakes scenarios like disaster response, trust in AI-generated results is important,” explains Nava. “Through this challenge, we aim to bring transparency to the model's decision-making process and empower decision-makers on the ground to act with confidence and speed.”
“Manual analysis of these satellite images is as boring as modifying Tripos test readings.”
This AI-powered detection system uses neural networks as a machine learning model to mimic complex functions of the human brain and recognize patterns. By supplying thousands of data points to the system, from stable slopes to fresh slide scars, AI learns to find potentially fatal slides.
Landslide identification often relies on two types of satellite information: optical imagery and radar data. In itself, there is a gap between each type of data. When thick clouds block the sky, traditional optical satellites are blind. In contrast, radar satellites can be seen through the darkness, but the information they provide is often more difficult to interpret. AI takes sharp boundaries of all-weather sustainability of radar and optical images, sewing them into rapid maps of thousands of landslides, creating more reliable pictures.
“AI takes sharp boundaries of all-weather sustainability of radar and optical images, sewing them into rapid maps of thousands of landslides.”
In addition to improving system accuracy, NAVA explains that the next step involves shedding light on the algorithmic process. “AI can feel like a black box. Its internal logic is not always transparent and people can hesitate to act on that output,” he said. Viewing tool inferences means that responders can better understand the data and increase confidence in the results.
The Cambridge team is not the only pit crew in this AI race with landslides. Researchers at NASA and the USGS have built an open-source landslide detection system that categorizes satellite photos as “objects” and then machine learning. Transfer learning pipelines have also been adopted to adapt the model to new landslide detection tasks.
In the event of a disaster, automation has already been deployed. During the Haiti earthquake and the bounty of the associated tropical storm, several groups carried out parallel rapid remote mapping efforts, providing responders with the first map of blocked roads and blocked valleys. Cambridge's works are within the global lineup that pushes landslide mapping over days or minutes.
The impressive feature of this project is, in this case, a different discipline of AI and geoscience, coming together to help people. Such interdisciplinary collaboration addresses complex challenges by leveraging the various technologies and technologies available, while fostering advancements in both fields.
So we know that the next time the Earth shaking, a Cambridge AI algorithm might be ringing the alarm.
