Official flood maps shape disaster preparedness, insurance decisions and urban planning, but large parts of the United States remain unmapped or under-mapped. This means some communities may be left unaware of the risks they face, limiting their ability to prepare for future floods.
Researchers from the National University of Singapore (NUS) College of Design and Engineering (CDE) and Tsinghua University School of Architecture, led by Associate Professor Rudi Stouffs from the Department of Architecture at NUS, have co-developed a deep learning framework that completes missing and under-mapped flood hazard zones across the contiguous United States.
By learning from existing official flood records and terrain data, the framework generated a spatially complete 30-metre flood hazard map that reveals a much larger scale of flood exposure than current official maps indicate. The paper, published in Nature Communications on 15 June 2026, highlights the potential of AI to strengthen public access to flood risk information and to guide more targeted resilience planning.
Revealing the scale of overlooked flood risks
The researchers found that official databases may have omitted around 11 million people and 4.1 million buildings from mapped flood zones. Taken together with the official baseline, the findings suggest that flood exposure across the contiguous United States may be substantially greater than currently recognised.
The study also found that these gaps are not evenly distributed. Many under-mapped and unmapped areas include socially vulnerable populations, particularly the elderly and children, pointing to wider implications for risk communication, resilience planning and the allocation of public resources to communities most in need.
That said, the research also showed that a concerted effort had been made to ensure more complete mapping in densely populated areas and in economically weaker regions.
How the model completes patchy and inconsistent data
The framework was trained to learn the relationship between terrain features and known flood hazard zones. Using elevation data and existing official flood records, the model identified patterns associated with flood-prone areas and applied that understanding to places where flood mapping was incomplete or absent.
Researchers generated a continuous 30-metre flood hazard layer across the contiguous United States. The approach offers a scalable complement to conventional flood mapping, which can be costly and time-intensive.
A key strength of the AI model is its capacity to generalise the underlying relationship between modern terrain data and flood inundation patterns. Although trained on a noisy dataset containing a mix of high-quality and outdated maps, the model learned to prioritise hydrologically consistent patterns over historical errors.
This capability is evident in its output. For instance, the model disregards the erroneous, abrupt boundaries found in some maps, which are often artefacts of administrative limits. Instead, it generates flood zones that logically follow the terrain. In other cases, it extends the boundaries of underestimated flood regions to better reflect the topographical context. In areas lacking official maps entirely, the model can still produce credible flood risk assessments.
"The model functions as a robust validation and correction system. It does not simply replicate outdated information; it actively leverages current topographical data to produce flood maps that are more consistent and, in many cases, more accurate than the source material,” said Assoc Prof Stouffs.
“Our study shows that AI can do more than reproduce existing maps. It can help detect where flood risk has been missed, examine who are potentially affected, and uncover what kind of social impacts it may have,” said Adjunct Associate Professor Ye Zhang from the Department of Architecture (and Associate Professor at Tsinghua University), a co-author of the paper. “That creates an opportunity to improve public awareness of risk and its social implications and to support more strategic flood resilience efforts.”
While the researchers note that the AI-generated maps are not intended to replace official regulatory flood maps, they can serve as a public guide to highlight overlooked risk and support better targeting of future mapping and adaptation efforts.
“By making hidden risk more visible, this approach can help communities and decision-makers act earlier and plan more effectively,” said Dr Abraham Noah Wu, first author of the paper and formerly PhD student at the Department of Architecture and now Research Fellow at Tsinghua University.
