A national experiment in data completion
Official flood maps cover only one-third of U.S. river channels. We trained a deep-learning model on the places FEMA had mapped and used it to complete a continuous 30-meter map of the lower 48—revealing 11 million people and 4.1 million buildings not represented in the national database.
Based on the Nature Communications study “Deep learning completes US flood hazard maps revealing millions exposed to previously unrecognized risk.”
The United States began its national flood-insurance program in 1967. More than half a century later, detailed federal studies cover only about one-third of the country’s river channels. In 2023, more than 40% of counties were still absent from the National Flood Hazard Layer, and only one-quarter of its models had been updated within five years.
Engineers can model a place in detail, cover a large territory, or keep the work affordable. Doing all three at once has been prohibitively slow and expensive—leaving many communities with coarse, outdated, or missing information.
Before this work
Existing FEMA flood patterns provide the training signal. Terrain provides the physical context. The model learns the relationship between the two, then estimates flood extent where detailed studies are missing.
Instead of commissioning a new engineering study for every river, the method turns the country’s authoritative—but incomplete—record into a reusable source of training data.
National scale. High resolution. Low cost.
Read the method in the paper →Pale blue shows FEMA’s published 100-year flood extent. Deep blue shows the model-generated extent. Pan, zoom, or change the comparison to inspect where the official record ends and the model continues.
FEMA’s national database places 16.06 million people and 5.07 million buildings inside mapped 100-year flood zones. The completed map raises those totals to 27.09 million people and 9.19 million buildings—69% and 81% more.
Some communities had detailed federal studies. Others had boundaries that stopped at administrative lines—or virtually no official flood extent at all. The model makes that uneven geography measurable.
Areas with little or no FEMA coverage had the highest median share of children and older adults. But the pattern is not a simple rich-versus-poor divide: areas completed by FEMA tended to have lower household incomes and higher income inequality. Mapping priorities reflect several kinds of vulnerability.
St. Louis shows underestimated tributaries, Grand Rapids shows abrupt administrative cutoffs, and Charlotte shows conservative boundaries and shoreline omissions. Each map is interactive.
Many public datasets are authoritative yet incomplete because expert surveys are costly. Conditional generative models can learn from the completed portions, combine them with physical context, and estimate the missing geography at scale.
The innovation is not a substitute for domain expertise. It is a way to make existing expertise travel farther—turning scattered observations into new, high-resolution public insight.
The paired bars compare FEMA-mapped exposure with the completed estimate on one scale.
This list ranks metropolitan areas with at least 10,000 exposed people by the share outside FEMA’s extent.