A new technology has been developed that can rapidly predict city-scale structural damage even when portions of satellite imagery are obscured by clouds or smoke immediately after a disaster.
A research team led by Professor In Ho Cho of the Department of Architecture at Seoul National University College of Engineering has presented a scientific artificial intelligence (Scientific AI) and data science framework that uses satellite imagery and various forms of open data to rapidly and accurately predict disaster-related structural damage at the city scale. By incorporating a statistical method for correcting incomplete satellite data, the researchers developed an approach that can estimate structural damage immediately after a disaster without requiring a separate, computationally expensive training stage.
The findings were published in Scientific Reports , an international journal in the Nature Portfolio.
As climate change intensifies and infrastructure systems become increasingly complex, structural damage caused by natural and human-induced disasters—including typhoons, floods, wildfires, earthquakes, and explosions—is occurring on an increasingly broad, city-wide scale. Rapidly determining where structures have been damaged and the extent of that damage immediately after a disaster is a critical first step in rescue operations, setting recovery priorities, and developing disaster-response measures.
In actual disaster settings, however, complete satellite imagery is not always available. Portions of satellite images may be obscured by clouds, smoke, precipitation, shadows, and other factors, which can reduce the performance of subsequent AI or statistical analyses. Existing AI-based damage assessment methods also often require large volumes of training data and substantial computational resources, making them difficult to deploy for immediate damage prediction in the aftermath of a disaster.
To address these challenges, the research team combined a Scientific AI approach, which integrates structural engineering knowledge into artificial intelligence, with a general-purpose statistical imputation method that rapidly corrects incomplete data. Rather than simply learning from data, Scientific AI incorporates engineering knowledge of the physical phenomenon of structural damage into AI-based analysis.
The framework developed in this study uses both pre- and post-disaster satellite imagery together with a range of publicly available data. In particular, it uses image entropy, a measure of disorder in satellite imagery, to analyze the extent of changes in structures and their surrounding environments before and after a disaster and to predict structural damage at the city scale. Entropy indicates how complex and irregular the information within an image has become and can be used to quantify changes in structures and surrounding environments following a disaster.
The researchers also applied a statistical data imputation method to enable damage prediction even when parts of satellite imagery are obscured or missing. Rather than discarding incomplete data or simply filling in missing values with averages, the method uses patterns and similarities within the data to statistically reconstruct missing information. This makes it possible to predict disaster damage even from satellite imagery in which some information has been lost because of clouds, smoke, or other obstructions.
A key feature of the framework is that, unlike state-of-the-art AI approaches, it does not require a costly training stage in which large-scale datasets must be newly assembled and trained. Instead of building extensive training datasets and retraining a model whenever a new disaster occurs, the framework can immediately estimate damage using pre- and post-disaster satellite imagery and publicly available data.
The research team validated the performance of the framework using a case involving damage caused by a major typhoon. The results confirmed that city-scale structural damage and destruction could be rapidly predicted even when using incomplete satellite data together with open data. Notably, even when more than 50% of the imagery was obscured by clouds, the framework successfully predicted the extent of damage to urban buildings beneath the obscured areas with an error of less than 5%.
The study also demonstrated that the approach could potentially be applied to disasters beyond major typhoons, including wildfires. The researchers applied the framework to a major wildfire damage case, demonstrating its potential to be extended to large-scale structural damage prediction across a range of natural and human-induced disasters.
The significance of this work lies in expanding disaster damage prediction beyond a problem of simply interpreting satellite imagery or training artificial intelligence models, instead treating it as a Scientific AI problem that integrates structural engineering knowledge with statistical data correction techniques. In particular, the study demonstrates the potential to rapidly estimate city-scale damage even under the incomplete-data and time-constrained conditions typical of the immediate aftermath of a disaster.
Rapid decision-making is critical in disaster response. If city-scale structural damage and destruction can be quickly identified immediately after a disaster, the information could help determine which areas should receive rescue personnel and equipment first, where recovery efforts should begin, and which areas require additional safety inspections. In the longer term, the approach is expected to have applications in urban disaster-prevention planning, post-disaster recovery strategies, and infrastructure management.
Professor Cho’s primary research focuses on Scientific AI, which combines scientific knowledge with artificial intelligence algorithms to address challenging problems in engineering and science. This study represents an advance that combines Scientific AI and data science to open new possibilities for city-scale disaster damage prediction.
Professor In Ho Cho said, “This study demonstrates that combining structural engineering knowledge with Scientific AI and data science can enable complex engineering information to be predicted rapidly and at low cost.” He added, “It is particularly meaningful that reliable engineering judgments can be derived even from damaged or incomplete image data.”
He continued, “We expect that this approach could eventually be extended to a wide range of fields, including the management of aging structures, image-based performance assessment of machinery, ships, and aircraft, and image-based analysis for defense applications.”
The study involved Professor In Ho Cho of the Department of Architecture at Seoul National University, researchers from Iowa State University, and a researcher from the U.S. Air Force Research Laboratory. The research was supported by the Seoul National University New Faculty Research Settlement Fund.
□ Introduction to the SNU College of Engineering
Seoul National University (SNU) founded in 1946 is the first national university in South Korea. The College of Engineering at SNU has worked tirelessly to achieve its goal of ‘fostering leaders for global industry and society.’ In 12 departments, 323 internationally recognized full-time professors lead the development of cutting-edge technology in South Korea and serving as a driving force for international development.
Scientific Reports
Data/statistical analysis
Not applicable
The authors declare no competing interests.