Researchers have used artificial intelligence to develop a new tool for assessing earthquake damage, a leap that could ultimately help first responders in making critical rescue decisions, suggests a new study.
The team's AI, called the LoRA-Enhanced Ground-view Generation (LEGG) diffusion model, is trained on real aerial drone images that it uses to create highly photorealistic 3D reconstructions of the ground. Creating imagery detailed enough to fully capture a region's physical characteristics distinguishes this synthetic model, enabling it to recognize complex visual patterns and predict where structures may be damaged, even in densely populated urban areas.
"What our algorithm does is generate thousands of pairs of semi-realistic photos of what a building looks like on the top and from the ground," said Rongjun Qin, co-author of the study and a professor of civil, environmental and geodetic engineering at The Ohio State University. "Having such data is vital, as drones gather important information from above, but people actually make emergency decisions from ground-level views."
Similar studies on the aftermath of devastating earthquakes relied on UAV or lidar-based detection methods to survey collapsed buildings and structures from above, but none had addressed how damage might have looked on the ground prior to prolonged rescue efforts. Moreover, depending on the severity of the earthquake, manual damage assessments can take days or weeks to fully complete, which isn't ideal for rapid recovery missions.
In this paper, Qin and his colleagues introduce a framework for bridging these gaps using AI-generated images, with the aim of laying the foundation for more accurate disaster assessment and better earthquake preparedness.
"This simulation is essentially a map, but an experienced and well-trained AI could offer an additional supply of information that would be really helpful for emergency crews in making quick decisions about where to go when the clock is ticking," said Qin.
The study was published in the International Journal of Remote Sensing.
To test the applicability of their proposed algorithm, researchers conducted a case study on a real-world disaster, the 2023 Kahramanmaras, Turkey, earthquake, a powerful 7.8 magnitude quake that destroyed 280,000 buildings and damaged at least 700,000 more. Comparing drone imagery from 2015 to photos taken in the days after the shake revealed dramatic changes in the local built environment, such as collapsed buildings and temporary shelters in open areas.
After showing their AI a dataset of only 3,000 of these city structures, the model was able to create images that enhanced the recognition of a number of building issues, including façade cracks, building tilts and partial collapses, demonstrating that it could extract subtle cues from multiple sources to generate high-resolution, photorealistic street-level views.
This advanced capability stems from the combination of drone and ground imagery that researchers injected it with to ensure the model had a strong starting point for understanding potential structural damage and its community effects, said Qin.
"As long as you have good data, AI can serve as a very generous predictor of past and future outcomes," he said. "It's a tool that can be incredibly helpful."
In the future, applying the team's framework to novel scenarios or areas could inspire governments and engineers to design more resilient infrastructures as well as reshape post-disaster assessment and emergency management policies.
"This work presents a great opportunity for engineers and other decision makers to remotely assess the damage in structures soon after a disaster," said Halil Sezen, co-author of the paper and a professor of structural engineering in civil, environmental and geodetic engineering at Ohio State.
That said, their algorithm will likely be utilized in tandem with other emergency or resource planning tools, said Qin, noting that with more in-depth experiments, the model could help anticipate destruction levels in other earthquake-prone environments, like Japan or California.
"There is still a lot of work to be done to bring in the kind of perspective AI offers," said Qin. "But the more good quality data that we have, the faster we're going to achieve our goals."
Co-authors include Ohio State's Ningli Xu, Abdullah Türer, Abdulmajeed Batarfi, and Hessah Albanwan from Kuwait University. This work was supported by the Scientific and Technological Research Council of Türkiye, the Ministry of Environment, Urbanization, and Climate Change of the Republic of Türkiye as well as the Intelligence Advanced Research Projects Activity (IARPA) and the Office of Naval Research.
