Deep learning completes US flood hazard maps revealing millions exposed to previously unrecognized risk

Abstract

Flood maps are essential for disaster preparedness, insurance and climate adaptation, yet US official maps cover only one-third of river channels. Here we present a deep-learning framework that learns from existing records to produce a spatially complete 30-m flood hazard map for the contiguous US, filling unmapped regions and updating legacy extents. We estimate that the 2023 national database omits 11 million people and 4.1 million buildings in flood zones, adding 69% and 81% beyond the official baseline respectively. Crucially, our socioeconomic analysis of 917 cities reveals that entirely unmapped areas disproportionately harbor highly vulnerable demographics, particularly the elderly and children, exposing them to unrecognized risks and compounding existing inequalities. By solving the trilemma between accuracy, scale, and cost in national scale flood mapping, our open-source method broadens public access to risk information, paving the way for targeted and equitable climate adaptation.

Millions in the US lack official flood risk data. This study uses an AI model to learn from existing maps and complete the US flood hazard map, revealing over 11 million more people exposed to flood risks, a 69% increase from current official records.

Introduction

Flooding ranks among the most destructive natural disasters in the United States, leading to billions of dollars of economic losses every year^1,2. With more frequent floods expected in the future^3–6, accurate estimation of flood risks and the dissemination of such information are crucial to protect lives and investments in the coming decades.

To mitigate these escalating losses, open-access flood maps that communicate detailed risk information are essential to help stakeholders adopt tailored strategies^7,8. In the U.S., the golden standard for this open, unified information is the National Flood Hazard Layer (NFHL). Maintained by the Federal Emergency Management Agency (FEMA) under the National Flood Insurance Program (NFIP), the NFHL aims to provide a single source of flood hazard information across the country. Though it was not intended to be a risk communication platform, it has become synonymous with that in the public view, and the NFHL has served as a vital resource for understanding flood risks, promoting insurance uptake, and has influenced the property market to reprice highly risky regions^9–12. Based on a report by the Association of State Floodplain Managers¹³, the investments to the NFIP program has yielded a 2-to-1 benefit ratio, with $10.6 billion in investment equating to nearly $22 billion in savings from avoided flood damages.

However, the maps provided by the NFHL are far from complete despite decades of efforts. Since the inception of the NFIP in 1967, FEMA has modeled merely one-third of US rivers, with only a quarter of these models receiving updates in the last five years¹³. As of 2023, more than 40% of US counties remain excluded from the NFHL database, and visible omissions can be observed even in counties with existing data, as shown in Fig. 1.

Fig. 1. Systematic spatial gaps and jurisdictional incompleteness in the National Flood Hazard Layer.

Representative examples from Michigan, New Jersey, and Texas illustrate the two primary typologies of data gaps in the current federal regulatory flood map. In Michigan (left), pervasive under-mapping is evident where large riverine corridors visible on the base topography are omitted from official hazard designation. In New Jersey (middle) and Texas (right), entire counties are missing from the NFHL database, resulting in abrupt, unnatural terminations of flood hazard zones at administrative boundaries. These gaps create significant information asymmetry, leaving communities unaware of actual risk levels. Sources of the map samples: OpenStreetMap contributors. Basemap: OpenStreetMap (https://www.openstreetmap.org/copyright). OpenStreetMap content is published under a CC-BY-SA 2.0 Attribution-ShareAlike 2.0 License (https://creativecommons.org/licenses/by-sa/2.0/). Flood layers: FEMA NFHL. Figure created using QGIS. The geographic center coordinates are indicated within the subfigures where applicable.

The incompleteness of the FEMA flood map means that while assets and population in the mapped areas are well-informed and protected, the unmapped, yet vulnerable communities remain exposed and oblivious to increasing flood damages. This is not just a random gap, but a systematic blind spot. Recent research has defined this as federally-overlooked flood risk¹⁴, which is estimated to affect 26 million people nationally¹⁵. Without the mandatory public insurance in the unmapped areas, poorer households in unmapped areas are more economically challenged to bear the negative consequences of a future flood³. This observation is confirmed at a national scale, where studies consistently link lower income to higher risk in these unmapped zones¹⁵. Event-based analyses of disasters like Hurricane Harvey found that racial inequalities in damage were primarily driven by impacts outside the official floodplains, often in Black and Hispanic neighborhoods¹⁶. In addition, properties located outside FEMA flood zones are often mistakenly believed to be devoid of any flood risk, while they could still be vulnerable to flooding to various degrees in reality. This misconception has left millions of property owners unsuspecting of flood damage risks, discouraging both flood mitigation strategies and insurance adoption, which in turn amplifies damage¹⁷. Furthermore, communities located in highly risky areas that are not officially designated as flood zones can only relegate to private insurance, which can be opaque in terms of risk assessment to homeowners compared to the open-access NFHL data and public insurance¹⁸. This creates a double jeopardy¹⁴, as these overlooked, at-risk structures are often of substandard quality, poorly maintained, and uninsured¹⁹, representing billions in unprotected financial liability with minimal insurance coverage²⁰.

In essence, while the benefits of the NFHL are evident, the inaccuracies and inconsistencies in the database create information asymmetry between mapped and unmapped areas, and are causing unintended consequences for the excluded communities, which will bear disproportionately larger economic and social costs as flood losses increase in the future, amounting to tens of billions of dollars of economic and social costs^21,22.

The fundamental limitation to the completeness of the NFHL maps is an inherent trilemma between accuracy, scale, and cost. Achieving high local accuracy typically demands prohibitive labor and administrative costs, as each flood zone in the NFHL is created through conversations between multiple stakeholders and complex administrative processes^23,24. Critically, this process also systematically under-represents certain hazard types, such as pluvial (rainfall-driven) flooding²⁵. While the flood zone delineation that made it into the NFHL is high in resolution, the cost associated with replicating the process towards a continuous, national-scale map is prohibitive.

However, similar to the high cost of acquiring FEMA flood maps, high-resolution computational models that assess flood risks are only made possible with substantial financial investments. While several of these global-scale models are available to academic researchers via data-sharing agreements, they often exist as commercial products that are not as readily accessible to the general public or integrated into local-level planning documents in the same way as the public-facing, regulatory NFHL. As such, while these data have enabled new insights on flood-related socio-economic impacts, limited access to the underlying data restricts the reproducibility of the findings and communities at risk still have no access to such data for resilience planning^26,27.

The crux of the problem lies in the trilemma between accuracy, scale, and cost in the methods used for flood mapping today. With FEMA’s multi-phasic, multi-stakeholder mapping method, the maps created are high in resolution and local accuracy, but the cost of labor and human expertise limits the scale of the mapping effort. With computational methods, large-scale simulations can be done to create a continuous data layer for the entire country. However, the computational complexity and resource requirements increase exponentially with accuracy, limiting access to privileged entities at the moment. If there is a way to generate a continuous, national scale flood map with the local accuracy of FEMA flood maps without astronomical costs, then the vulnerable communities can start resilience planning today.

This paper makes several contributions in addressing the problem. First, we present a flood mapping method based on a Geospatial Data Translation (GDT) model, which learns from existing FEMA flood maps in an unsupervised manner and can accurately demarcate 100-year flood zones in the unmapped regions. This allows us to create a complete flood hazard map of the lower 48 states at 30 m that is consistent with the resolution and extent of the FEMA NFHL 100-year flood zones (1% annual probability). Second, leveraging the data generated by the proposed method, the study reveals the information asymmetry within the FEMA flood maps and quantifies the additional at-risk population and buildings at both the national level and Census Tract level. Third, our analysis of over 900 Core-based Statistical Areas (CBSAs) revealed that, despite FEMA’s efforts in flood mapping in major urban areas, 40% of all metropolitan and micropolitan areas are still mapped inadequately by official flood maps. Fourth, by matching every building to the highly detailed flood map, we show evidence of how the open-access flood map helps with ameliorating inequity from information asymmetry, guiding undermapped communities towards flood resiliency today.

Results

Using a generative AI architecture based on GDT models^28,29, we have created a spatially contiguous, 100-year flood map for the entire lower 48 states. The AI model learns the relationship between terrain data and the FEMA NFHL flood maps in an unsupervised manner and is able to generate flood zones in unseen areas with pixel precision up to 81% when compared to the ground truth. The performance details are substantiated further in the “Methods” section. After training the model on all FEMA data nationally, we deploy the model for all areas in the lower 48 states to create a continuous flood hazard layer that retains the resolution of the native FEMA flood data, achieving the high-accuracy of FEMA data for all river and flood channels in contiguous US.

Substantial population and buildings at risk ignored by FEMA

We begin our analysis by tabulating the total number of exposed population, building area, and building count by the ground truth of NFHL and the additional flooded areas revealed by the AI model. Table 1 shows the comparison between the two datasets. In total, there are 11.04 million people exposed to flood risks as mapped by the AI model in addition to the 16.06 million people mapped by the NFHL. This leads to a total population of 27.09 million potentially exposed to flooding, which is a 69% increase compared to that mapped by the NFHL. Building counts and building areas face even larger increases from the NFHL ground truth, at 81 and 83%, respectively. These statistics echo existing criticisms of the deficiencies in the FEMA dataset in providing flood risk information to the public^3,13,23,30. Major improvement in coverage is required on FEMA’s part to enhance the public flood map registry further.

Table 1.

Buildings and population at risk

	Buildings	Bldg. Area (sqm)	Population
Mapped in NFHL	5.07	1448.98	16.06
Additionally mapped by AI	4.12	1203.38	11.04
Total	9.19	2652.36	27.09

Amounts in millions

That being said, mapping a large country using non-computational methods requires tremendous effort. Understandably, regions with the highest population and assets at stake need to be prioritized to maximize the benefit to the public at the minimum cost. Thus, we also need to analyze how efficient FEMA is in prioritizing areas with high population density for a more comprehensive judgment on the program’s effectiveness.

Overlaying the flood maps with a 30-m gridded population map, we can compare the geographical heterogeneity between the NFHL dataset and the additional exposed population highlighted by the generated flood map (hereafter referred to as mapped by AI interchangeably) as shown in Fig. 2. The map aggregates the 30 m pixels in hexagon cells with approximately 200 m in edge length using the H3 index, and visualizes the number of population in a cell with a quantile color gradient. From the figure, we see that the NFHL has mapped most of the population centers in the east, while omitting the sparser regions in the west (Region D). Similar patterns can be found in region A, where the NFHL has mapped the nearby metropolitan area, while neglecting the sparser, rural communities. However, this rural-urban divide is not always true, as revealed in Regions B and C. In region B, we see that half of the urban population cluster is mapped by the NFHL, while the other half is completely neglected. In region C, a large number of population located in highly populated areas are left out of the NFHL.

Fig. 2. National-scale distribution of known and unrecognized population exposure to flood hazard.

Comparison of flood-exposed population distributions between NFHL regulatory data (red-orange) and generative model predictions (cyan-blue) across the contiguous United States. Hex-grid aggregation (H3 level 9, edge ≈ 200 m) reveals significant information gaps characterized by: A, B jurisdictional discontinuities at metropolitan fringes, C dense urban population centers with absent federal hazard data, and D extensive regional omissions in the western US. The generative layer provides a spatially consistent baseline that rectifies administrative mapping fragmentation.

In general, there is a clear preference for prioritizing large population centers across the country, ensuring that flood mapping efforts on the ground benefit as many residents as possible. However, there are also obvious regions where exposed populations are abruptly cut off and stranded from well-mapped regions, leading to a large number of residents still lacking any official flood map from the NFHL. This analysis demonstrates the power of the AI-generated flood map to unlock unprecedented information for the exposed population and assets that were previously excluded from the NFHL.

Given the high resolution of the generated map, we can aggregate the H3 hexagons to the Census Tract level and calculate the percentage of exposed population for every tract to understand how the AI flood map adds to completing the existing FEMA flood database. Fig. 3 presents a series of maps that visualize the population in flood zones as a percentage of the total population of each census tract. This allows us to understand how much the AI-generated maps can contribute to completing the missing flood information.

Fig. 3. Completion of national flood hazard registry at federal tract level.

Percentage of population within 100-year flood hazard zones for every census tract in the contiguous United States. a Combined extent merging official NFHL data with additional AI-revealed basins; b Current regulatory baseline as documented in the FEMA NFHL; c Unperceived flood risk identified by the generative model. The maps highlight that pervasive under-mapping exists not only in entirely overlooked tracts but also as internal omissions within officially mapped jurisdictional boundaries.

Figure 3b presents the current status quo, showing the percentage of at-risk population as mapped by the FEMA ground truth flood maps. Census Tracts with no FEMA flood maps accessible from the NFHL database are grayed out. In this map, we see the commonly flooded regions, such as areas around West Virginia, Louisiana, and Florida, having a larger proportion of populations at risk.

Figure 3c shows the additional flooded population revealed by the AI in all census tracts. Due to the continuous nature of the AI flood zones, we see that the AI flood map has revealed additional population at risk, both in tracts already mapped by FEMA and also in tracts that are overlooked by FEMA.

Adding the data from Fig. 3b, c together, we obtain Fig. 3a, which shows the total percentage of the population at risk in each census tract. This map depicts a much more comprehensive assessment of the exposed population in the country as it combines the flood zones mapped by FEMA and the additional flood zones mapped by the AI model.

In essence, the AI-generated flood zones effectively fill the gaps present in FEMA maps. With identical spatial resolution, we can measure and contrast unmapped populations and structures at risk, highlighting limitations in FEMA maps through a national spatial analysis. Furthermore, merging the AI flood maps with the FEMA flood maps can create a spatially continuous map that approximates a comprehensive national flood hazard map for the contiguous United States.

Socioeconomic disparity between mapped and unmapped cities

In the previous section, we analyzed the spatial heterogeneity in FEMA’s mapping efforts around the country and quantified the populations that are left out of the official flood map. With higher population densities in urban areas, omissions from the FEMA flood map in these areas leaves more population at risk per unit area. This is prominently shown in Fig. 2B, C, where large clusters of urban populations remain unmapped.

In this section, we want to quantify the degree of completeness of FEMA maps for all CBSAs in the lower 48 states to focus impact assessment in densely populated areas, and investigate whether more vulnerable populations are being protected by the FEMA maps. Here, demographic vulnerability is defined by age—specifically, communities with high percentages of elderly (≥ 65 years) and children (< 18 years)—as well as by CBSA-level economic indicators: household median income, income growth rate, and the Gini index.

We restrict the area of interest only to 917 CBSAs, where each has at least one urban core area with more than 10,000 population (hereafter referred to as cities or CBSAs interchangeably), and group each area based on the ratio between the percentage of exposed population mapped by NFHL and AI as indicated in Fig. 4a:

• Group 1: No NFHL data—Flooded CBSAs that have more than 1% of population mapped by AI and less than 1% mapped by NFHL.

• Group 2: Largely unmapped—CBSAs with more population mapped by AI than NFHL, and have a minimum mapped population of 1%.

• Group 3: Slightly unmapped—CBSAs with more population mapped by NFHL than AI, and have a minimum mapped population of 1%.

• Group 4: Completed by NFHL—Flooded CBSAs with more than 1% population mapped by NFHL and less than 1% mapped by AI.

Fig. 4. Demographic and spatial distributions of 917 core-based statistical areas (CBSAs) across varying degrees of FEMA mapping completeness.

a Classification of CBSAs into four groups based on the ratio of predicted exposure mapping between the AI model and the NFHL database. b Population histograms of cities falling into each mapping completeness category. c Comparison of total exposed population percentages across the four CBSA groups. d Geographic distribution revealing regional clustering of completely unmapped vs. highly mapped cities. e Socioeconomic profiles demonstrating that while completed NFHL areas (Group 4) align with lower household incomes and higher wealth disparities, entirely unmapped cities (Group 1) disproportionately harbor the highest percentages of highly vulnerable demographics, such as the elderly and children. Demographic data source: US Census ACS ⁴⁶.

Each of the four groups represents varying degrees of the NFHL mapped/unmapped ratio as revealed by the generated AI flood maps. The higher the ratio of the population mapped by AI, such as in Group 1—No NFHL data and Group 2—Largely unmapped, the more incomplete the current NFHL is to those areas and vice versa. The 1% threshold is chosen to remove cities without any flood risks.

Figure 4b plots the populations of the cities in the four groups on a histogram. Cities in Group 3, where NFHL has mapped more population than the additional population revealed by AI, are the majority, with 441 cities. Cities in Group 2, where there are more unmapped populations than those mapped by NFHL, are the second most common group with 266 cities, which contribute to a large number of total populations that are unmapped by NFHL. Alarmingly, while the total number of cities is fewer, there are still 19 cities with more than 1 million inhabitants, such as St. Louis and Pittsburgh in Group 2, suggesting deficiencies in NFHL’s efforts to cover major metropolitan areas. With the high population base of the metropolitan areas, a small change in flood map completeness will exclude a large number of people. This warrants further investigation, which we will explore in the next section.

Groups 1 and 4, representing each end of the mapped/unmapped spectrum, are the minorities with Group 1 at 103 cities and Group 4 at 89 cities. These cities are mainly micropolitan areas with a population of less than 10,000 people with a few large metro areas as exceptions. Figure 4d maps the geographic distribution of these cities with the other groups. Although Groups 2 and 3 are scattered throughout the country, we can observe more visible clusters for Groups 1 and 4. For example, Group 1 cities, where there are no NFHL data at all, are more common in the Northwest such as Spokane Valley, WA and Idaho Falls, ID, and in the Northeast such as South Portland, ME and Glens Falls, NY; cities in Group 4, on the other hand, are more common in the Southwest such as Phoenix, AZ and San Jose, CA and in the South such as Corpus Christi, TX and Cape Coral, FL.

Interesting insights can be found by comparing the total exposed population of cities between these two groups, as shown in Figure 4c. We see that most of the Group 1 cities without NFHL data have lower percentages of population at risk, with the median value of 3.5% and the mean value of 6.6%. Group 4 cities, where FEMA has a high degree of completion, have a much higher median value of 8.5% and mean value of 13.2%. This shows that cities that are unmapped by FEMA (Group 1) generally have a lower percentage of population at risk. However, according to Fig. 4e, the Group 1 cities also have the highest percentage of vulnerable population, namely elderly and children, who are more susceptible to hazards. At the same time, there are certain outlier cities in Group 1 with a high percentage of population unmapped by FEMA, and these should be prioritized as soon as possible. Such CBSAs include Houma-Bayou Cane-Thibodaux, LA, Alamosa, CO, and Alexandria, LA.

Looking again at the socioeconomic profiles of the four groups shown in Fig. 4e, we observe that the cities that are completed by FEMA (Group 4) score markedly lower in household median income, income growth rate, and have a higher Gini index. As populations residing in CBSAs with much lower median incomes are often the most fragile to acute flood hazards, this finding recognizes FEMA’s effort to provide public flood insurance in the NFIP to build flood resilience in economically challenged communities. A similar pattern can be found in Group 3 (Slightly Unmapped), where the cities have much lower household median income and a higher Gini index than Groups 1 and 2, suggesting FEMA’s focus on lower-income regions.

In summary, the AI-generated flood map has revealed that 40% of all metropolitan and micropolitan areas are inadequately mapped by the NFHL (Groups 1 and 2), with 103 CBSAs out of the total 917 cities having little to no FEMA maps, even though there are at least more than 1% of the population at risk. More alarmingly, demographic analyses have revealed that these cities also have the highest percentage of elderly and children populations, who are more vulnerable to crises as shown in Fig. 4e. That being said, despite the presence of large, unmapped regions, FEMA’s effort in mapping densely populated areas is to be acknowledged. More than 57% of the CBSAs that are exposed to flooding (Groups 3 and 4) are fully mapped or completed to a high degree by the NFHL. We also observe a clear leaning towards cities with lower household income and higher wealth disparities, indicating a conscious mapping effort in regions that are weaker economically.

Visual pattern analysis reveal inadequacies in NFHL mapping

Through spatial analyses, we have investigated how the AI-generated flood map can effectively highlight the inadequacies and spatial inequity in the FEMA flood map at the Census Tract and CBSA levels. At the community and building level, this spatial inequity means that homeowners living in exposed areas but are excluded from flood zones suffer disproportional risks compared to those living within identified flood zones. Therefore, another important use case for the AI-generated map is to enhance the coverage of the NFHL and aid homeowners to better assess flood risks.

We selected three metropolitan areas (MSA) with a total population of over 1,000,000 for a close-up analysis to compare the AI-generated flood map with the FEMA flood map. These results are presented in Fig. 5. In general, the various patterns shown in the figure can be categorized into three different typologies of undermapping at the local level:

• Underestimation of flooded areas within flood channels in areas recorded by FEMA as shown in (Fig. 5a1, a2, c1, and c2)

• Abrupt cut-offs of flood zones due to FEMA zoning or administrative region border as shown in (Fig. 5b1)

• and completely missing regions with no FEMA presence in the entire county/city as shown in (Fig. 5b2)

Fig. 5. Multi-scale analysis of NFHL undermapping typologies highlighted by the AI model.

The figure illustrates three distinct forms of FEMA flood map inadequacies across three selected metropolitan areas: a Underestimation of flooded areas within established flood channels, such as missed commercial properties and public infrastructure in the St. Louis MSA (a1 Bridgeton, a2 Harrisonville). b Missing FEMA coverage in the Grand Rapids MSA, including abrupt, unnatural cut-offs of documented flood zones artificially bounded by administrative borders (b1 Hudsonville), and entirely unmapped jurisdictions leaving dense urban centers completely oblivious to flood risks (b2 Downtown Grand Rapids). c Overly conservative predictions in the Charlotte MSA (c1 West Midtown, c2 Lake Norman). In all cases, the AI map successfully identifies and corrects these spatial gaps using modern high-resolution topography. Building footprint from OpenStreetMap. OpenStreetMap (https://www.openstreetmap.org/copyright). OpenStreetMap content is published under a CC-BY-SA 2.0 Attribution-ShareAlike 2.0 License (https://creativecommons.org/licenses/by-sa/2.0/). Basemap from Mapbox Satellite. Flood layers: FEMA NFHL. Figure created using QGIS. The geographic center coordinates are indicated within the subfigures where applicable.

In the St. Louis metropolitan area, with a total population of 2,814,000, we compared the generated flood map with the NFHL data, and found that the NFHL has mapped 76,000 exposed population in the urban (rounded to the nearest thousand) with an additional 73,000 exposed population unmapped by NFHL. Taking the Bridgeton area along the Missouri River as an example (Fig. 5a1), it is evident that a large number of commercial and public buildings are not covered by the NFHL data. The original NFHL data does not include the above-ground portions of the highway, but the entrances and exits are still affected by flooding and become impassable, which does not reflect the actual situation. Our model, however, predicted this scenario and included the affected sections of the highway and surrounding buildings in the flood map. In Harrisonville (Fig. 5a2), due to the absence of Illinois data in the NFHL, a large portion of farmland and some residential areas are not mapped.

In the Grand Rapids-Wyoming-Kentwood metropolitan area, the data reveals that 37,000 exposed populations are at risk that are unmapped by the NFHL, which is 2.84 times the mapped population. In Hudsonville (Fig. 5b1), we observed that the NFHL flood map breaks abruptly in a vertical manner along the administrative border between the two counties, resulting in large unmapped areas to the east that do not reflect the actual flood conditions. In downtown Grand Rapids (Fig. 5b2), the entire urban area has not been mapped by the NFHL, leaving a large number of buildings and population at risk.

In the Charlotte-Concord-Gastonia metropolitan area with a total population of 2,529,000, 50% of the entire exposed population is unmapped by the NFHL. In West Midtown and Wilson Heights (Fig. 5c1), due to the overly conservative predictions of the original flood map, we discovered that commercial buildings, urban public buildings, and some residential areas were not included in the flood-affected zones. Around Lake Norman (Fig. 5c2), the original map was also too conservative, showing that residences located in the lake’s center and along the shoreline were not affected, which does not align with the actual conditions during a flood event, as revealed by the additional flooded areas generated by the AI.

In summary, FEMA flood maps exhibit inconsistent local coverage and often suffer from various forms of undermapping. In areas without FEMA maps, the AI model can highlight areas that are most likely to be flooded. In areas with FEMA maps, flood zones can be abruptly cut off due to administrative limits, while AI-generated maps have demonstrated a wider scope by extending flood regions beyond these administrative boundaries. In areas with continuous flood zones, the FEMA maps might underestimate the extent of flooded regions, while the AI-generated maps extend the flooded areas based on the terrain characteristics. This observation highlights a key strength of our model: even though it was trained on a massive and noisy dataset containing a mix of both high-quality, updated maps and older, outdated ones, the model is able to generalize the underlying relationship between the terrain context from modern 3DEP terrain data and the FEMA flood inundation patterns. As shown in the figure, the model, having learned from correct examples elsewhere, disregards erroneous boundaries of the FEMA flood zones and can generate hydrologically consistent flood zones that follow the terrain. Thus, the model is not merely replicating outdated maps elsewhere; it is a robust system that can identify and, in many cases, correct these historical or administrative inconsistencies based on current topographical data.

Discussion

In this research, we have introduced an efficient flood mapping method that tackles the decade-long problem of the inequitable access to accurate flood information in the US. With the AI-generated maps filling in the missing flood zones in the FEMA NFHL dataset, we can pinpoint and quantify the extent of spatial disparity with a resolution of 30 m for all the lower 48 states.

At the national scale, our findings highlight risk areas not covered by the NFHL and emphasize the unseen economic and social costs behind its gaps. We found that there are at least 11 million and 4.1 million excluded populations and buildings, respectively, that are highly likely to be located within 100-year flood zones. Despite decades of effort in mapping the nation, this represents a 69% and 81% increase in exposure compared to the baseline currently registered in the database. Spatially, there are large areas at risk that remain completely ignored by the NFHL, particularly in the West, Midwest, and Northeast (Fig. 2). On a brighter note, the analysis suggests that the current FEMA mapping efforts have been effective in targeting high-impact areas, already accounting for 59% of exposed populations while only mapping 33% of river channels¹³. As such, to justify the high investments in future flood mapping, efforts must be prioritized in the highly vulnerable but sparsely mapped regions (Fig. 3c, d).

At the municipal level, we found that the degree of data completion varies based on the total percentage of the population at risk and the size of the city. Alarmingly, 40% of all metropolitan and micropolitan areas remain inadequately mapped. Even more concerning, the 103 CBSAs with little to no FEMA coverage contain the highest proportions of vulnerable demographics, such as the elderly and children (Fig. 4e). That being said, despite the presence of large unmapped regions, FEMA’s effort in mapping densely populated areas is to be acknowledged, and their focus on lower-income areas suggests a conscious effort to support economically weaker regions. However, at the neighborhood and building level, risk-exposed populations still suffer from information asymmetry. Our analysis has shown that entire communities can be abruptly cut off from the database due to administrative boundaries like FIRMs, while FEMA’s documented conservatism often underestimates actual flood channels^31,32 (Fig. 5). Encouragingly, the results demonstrate that our model’s output can be blended together with the current NFHL to create a more complete flood map, helping to guide vulnerable communities in making educated decisions today.

Placed within the context of recent literature, our results provide a complementary view on the scale and nature of unmapped flood risk. Our identified exposure of 11.04 million people and 4.1 million buildings provides a conservative, national-scale baseline of unmapped fluvial and coastal risk. This is lower than the 26 million residents estimated to live in federally-overlooked floodplains by ref. ¹⁵, primarily because their advanced hydrodynamic models explicitly include pluvial risk, whereas our NFHL-trained model largely omits it²⁵.

Our approach offers distinct advantages in transparency and verifiability, directly addressing the accuracy-scale-cost trilemma in flood mapping²³. While advanced hydrodynamic models are invaluable for scientific research, their high-resolution results are often closed-access and not easily verifiable^26,27. In contrast, our 30-m, building-level methodology is fully open-source and reproducible, providing a publicly-accessible assessment of gaps in the current regulatory map.

Finally, our finding that median household incomes are lower in cities with more complete FEMA mapping aligns with the broader socio-demographic literature. Studies by refs. ^3,33,34 consistently demonstrate that flood risk is disproportionately concentrated in economically disadvantaged and socially vulnerable communities. Refs. ^15,26 further corroborate these vulnerability patterns across different flood types and historical claims. Collectively, these studies support our conclusion that FEMA’s mapping prioritization, despite its spatial gaps, has been partially effective in initially targeting economically vulnerable areas.

Our results carry implications beyond flood mapping. The core contribution is a demonstration that conditional generative models can serve as scalable data completion engines for authoritative but spatially incomplete geospatial datasets. First, the model learns an implicit physical prior; without explicit encoding of hydrological equations, it captures terrain-to-inundation relationships that produce hydrologically plausible outputs. Second, we have shown that a single model can generalize across highly heterogeneous geographic contexts—from the arid Southwest to densely vegetated Southeastern floodplains—when trained on sufficiently diverse data. This methodology is directly transferable to other geospatial data completion problems where high-quality local data exists, but national coverage is lacking, offering valuable insights for the ongoing development of geospatial foundation models^35–37.

While this generative approach provides a continuous proxy for national flood risk, it is important to acknowledge its limitations. A key characteristic of our maps is that they inherit the specific scenarios present in the training data. Because the FEMA NFHL primarily delineates the 100-year fluvial and coastal floodplain, our maps do not account for pluvial (rainfall-driven) flooding, an increasingly critical hazard due to climate change^25,34. However, we hypothesize that the model could be trained to generate maps for other return periods (e.g., 500-year) or flood types if the corresponding training datasets were available. This presents considerable opportunities for future research, positioning our methodology as a low-cost means to scale up and interpolate specialized regional flood studies.

Crucially, these AI-generated maps should be contextualized appropriately. They serve as a public guide to reveal flood-prone areas in unmapped regions, but they are not intended as a direct replacement for official, regulatory FEMA flood zones. Their primary utility lies in quantifying exposure and directing limited mapping resources toward the most vulnerable unmapped communities so they can begin resilience planning immediately.

As climate change introduces additional volatility to natural hazards worldwide, access to high-quality risk information must be universally available. The open-source, transparent methodology proposed in this study makes significant progress toward this goal by tackling the accuracy-scale-cost trilemma. By drastically lowering the barrier to creating these maps, this approach can help bridge the information gap that has historically led to disparate financial impacts^21,38. Furthermore, our findings provide actionable insights for federal resource allocation, aiding policymakers in targeting high-impact areas. For cities, this complete albeit synthetic risk map enables the reassessment of zoning and flood management policies, while its building-level granularity empowers individual homeowners to make educated decisions and avoid high-risk development.

Moving forward, synthesizing these AI-derived outputs with emerging geospatial foundation models or diffusion architectures will likely refine their predictive capabilities. Ultimately, integrating open-source, high-fidelity AI models into national mapping strategies can democratize risk information, ensuring that the most vulnerable communities are not left in the dark against escalating climate threats.

Methods

As mentioned in the introduction, the incompleteness of the FEMA NFHL dataset and the lack of public access to alternative flood maps has led to various social and economic problems in the federally-overlooked regions. To address this issue, we have created a method that is able to complete the FEMA flood maps at a national scale without the high cost of computation of the current state-of-the-art flood models. The proposed method takes inspiration from GDT models^28,29 that have been shown to be effective in generating vector spatial data in missing regions by learning from high-quality data from available regions, and applying the learned parameters to new regions. In the context of the NFHL data, this method is able to learn from areas mapped by FEMA and translate its learned parameters to regions that are undermapped or completely unmapped. Although our study focuses on flood maps, this process can be applied to any other hazard map that can be inferred with contextual information. Here are some key terms used throughout the process:

• Tileset—a set of images tiles with uniform resolution, which results from the slicing/tiling of a larger geospatial image set to fit into the deep generative algorithm

• Target Tileset—the expected ground truth for the training of the model, i.e., the flood maps

• Context Tileset—an input dataset that contains contextual information to help the model to learn the interpolation to the target, e.g., digital elevation models.

• Blended Tileset—a merged tileset combining the target tileset with the context tileset. Used as training sample to the model.

• Generated data—the output as generated by the trained model i.e., the flood maps generated by the model.

• In-sample dataset—data regions selected from within the geographic boundaries of the training dataset, used to evaluate the model’s ability to interpolate within known geographical contexts.

• Out-of-sample dataset—data regions located entirely outside the geographic boundaries of the training dataset, used to evaluate the model’s extrapolation performance in new, unseen geographical contexts.

Data preparation

The 1 arc-second (approximately 30 m) Digital Elevation Model (DEM) from USGS 3DEP³⁹ serves as the Context Data. Its nationwide contiguous availability allows us to synthesize flood maps in regions missing from the FEMA NFHL dataset. For training and inference, the target dataset (NFHL 100-year flood maps from 2023⁴⁰) and context dataset (terrain) are processed into uniform raster tiles (e.g., 1024 × 1024 or 512 × 512). Overlaying the flood regions onto the DEM creates a Blended Tileset, yielding better visual and statistical performance than using the flood tileset alone.

Additional visual enhancements, such as color gradients and hill shades, are applied to the context (terrain) tiles to help with model training and metric computation. It is important to note that these enhancements, particularly the hill shades, create a high-contrast embossed effect that visually exaggerates small shifts in elevation and does not mean that the absolute elevation of the terrain is substantial.

This tiling process is ran at different zoom levels according to the Web Map Tile Service (WMTS) standard, which allows the model to learn the flood pattern at different zoom levels. At higher zoom levels, a single region can be split into more tiles, generating more samples for the model training. This helps to avoid certain problems that arise from the lack of data, such as mode collapse or the occurrence of visual artifacts in the generated images. At lower zoom levels, each tile captures more information, which allows the model to be more contextually aware, such that the generated images can be stitched together seamlessly. This multi-scale training approach reduces the prevalence of edge artifacts; by learning the continuity of features at multiple scales, the model can generate more seamless maps without requiring a post-processing overlap or blending strategy. We found that the best training set should contain tiles from multiple zoom levels. For the case study conducted in this paper, we used zoom levels from 10 to 13 according to the WMTS standard for training and inference.

Model architecture

The generative model that enables data completion is adapted from the architecture suggested by ref. ²⁸, a type of Conditional Generative Adversarial Network^41,42 that is optimized for tiled geospatial data. As shown in Fig. 6a, this architecture uses a coarse-to-fine generator (composed of a global generator G1 and a local enhancer G2) and a set of three multi-scale discriminators. This approach allows the model to effectively learn and generate flood features across different scales, from broad river valleys to fine-channel details.

Fig. 6. Overview of the generative flood mapping framework.

a Model architecture: the coarse-to-fine generator consists of a Global Generator (G1) operating at 512 × 512 and a Local Enhancer (G2) at 1024 × 1024 resolution. Three multi-scale PatchGAN discriminators compare generated and ground truth tiles at progressively downscaled resolutions to compute the total adversarial loss. b Flood data generation pipeline: curated context data (terrain) is tiled, passed through the trained generator to produce synthetic flood tiles, stitched into a seamless raster, and vectorized into a continuous flood map.

Our model is capable of generating high-resolution images at 1024 × 1024 pixel resolution. This capability is achieved using a Coarse-to-Fine generator⁴³, which produces target images at different scales (Fig. 6a). To attain this performance, two residual networks, G1 and G2, are constructed. The training process involves initially training the residual network G1 on lower resolutions. Subsequently, G2 is appended to G1, and the combined network is trained jointly on high-resolution images. The overall structure of the generator is an encoder-decoder pair⁴⁴, where the first half of the model encodes inputs into latent vectors, and the latter part decodes these vectors into visual representations. During training, with each forward pass, the components of the generator aim to synthesize outputs that can fool the discriminators into classifying the generated images as real. Concurrently, the discriminators learn to distinguish between the generated images (classifying them as fake) and the ground-truth targets (classifying them as real).

There are three discriminators at different resolutions to determine the quality of the output image at different resolutions. In each discriminator, Convolutional Kernels are applied to extract and compare the features from both the generated images and the ground truths. The final classification is the average value across 3 discriminators that indicates whether the generated image is real or fake. At the end of each forward pass, the loss for both the generator and discriminator is calculated, and their weights are updated. With each epoch, both the generator and discriminator improve in their respective roles. The training continues until the generator’s output is able to fool the discriminator as real images approximately 50% of the time, indicating that both the generator and the discriminator have converged to optimal performance, achieving a balance where neither outperforms the other consistently.

Model training and inference

During training, areas with existing FEMA flood zones are selected and processed into Context and Target tiles, and are fed into the unpaired translation model described above. The model was optimized using the Adam optimizer with an initial learning rate of 0.0002 and a momentum parameter β₁ = 0.5. The batch size was set to 4 due to the high resolution of the tiles. The models were trained for a total of 100 epochs, maintaining the initial learning rate for the first 60 epochs and linearly decaying it to zero over the remaining 40 epochs. The overall objective function combined a Least Squares Generative Adversarial Network (LSGAN) loss for realistic map generation, a feature matching loss from the 2 discriminators (weighted at λ = 10.0) to stabilize training, and a perceptual loss calculated with a pre-trained VGG19 network to ensure visual and structural consistency of the flood maps. The training time and epochs required to reach convergence scales proportionally with the number of tiles. We found that a minimum of 500 tiles are required for stable training to avoid mode collapse.

During inference, the trained model is able to take in any context data (formatted with same tiling and augmentation techniques as during training), including regions without any FEMA data, and generate a new Blended Tileset. The Blended Tileset is then stitched and vectorized to create a continuous, vectorized flood zone, similar to the original format of the NFHL dataset. Figure 6b demonstrates the steps of the inference process.

In our experiments, we found that a single model, trained on all available FEMA data, provides the best validation result that balances local precision and national scale generalization. On the other hand, training localized model with similar terrain patterns leads to better local performance when applied to areas with similar terrain patterns, but struggles when applied to further, morphologically different areas. This trade off between specialized models versus general models is discussed in detail in Section 3.5. The national model used for analysis in the Results Section is trained with 300,000 tiled images, covering all regions that are mapped in the NFHL dataset. Then, all available context information (the entire CONUS 3DEP terrain data) was fed into the model to create the AI-generated flood maps used in the “Results” section.

Evaluation metrics

The results from the model can be measured against the ground truth tilesets to understand their strengths and weaknesses. We selected three metrics, Recall, Precision, and mean Intersection-over-Union (mIOU), for this purpose. Given the unsupervised nature of the model, the selected metrics are not optimized during training. Rather, these numbers offer us insight into how the generated flood regions compare to the ground truth in different geographic contexts. It is important to note that the interpretation of these metrics will differ when evaluating in-sample versus out-of-sample data, as ground truth mappings themselves may exhibit systematic biases or omissions in previously unseen areas. The recall and precision are calculated from the components of the confusion matrix:

• True Positives (TP) (pixels correctly classified as flood),

• False Positives (FP) (pixels incorrectly classified as flood),

• True Negatives (TN) (pixels correctly classified as non-flood), and

• False Negatives (FN) (pixels incorrectly classified as non-flood).

• Recall reports the True Positive Rate of the prediction. It measures the fraction of positives that are correctly classified. The higher the Recall, the higher proportion of pixels in the GT flooded region are being generated by the model. It is defined by TP/(TP + FN).

• Precision reports the number of true positives divided by the total number of positive predictions. There are instances where the GT is undermapped, or the model hallucinates in non-flooded regions. Therefore, Precision is a good metric to understand how much of the positive predictions are actually true positives. It is defined by TP/(TP + FP).

• Mean Intersection-over-Union (mIoU) calculates the average Intersection-over-Union between the flood layers of the ground truth and synthesized tileset. This metric measures the similarity between the two sets in the most stringent way, and heavily penalizes false positives.

Model validation

To evaluate the generative capability of the model, we benchmarked its performance across three regions with distinct geomorphological features: New Jersey, Texas, and Michigan. Assessing model performance involves navigating the tradeoff between accurately mapping localized hydrology and extrapolating patterns into novel topographies. The integrated results of these experiments are compiled in Table 2.

Table 2.

Integrated performance metrics across in-sample, out-of-sample (local), and national models

	In-sample (Local)			Out-of-sample (Local)			National			National
Region	Prec.	Rec.	mIoU	Prec.	Rec.	mIoU	Prec.	Rec.	mIoU	Gain (mIoU)
Texas	0.95	0.86	0.83	0.74	0.59	0.47	0.81	0.61	0.53	+0.06
Michigan	0.89	0.80	0.73	0.52	0.60	0.38	0.63	0.65	0.48	+0.10
New Jersey	0.95	0.82	0.79	0.65	0.48	0.37	0.81	0.61	0.54	+0.17

In-sample validation (Interpolation)

When predicting within their training geolocations, local models reliably replicated known flood patterns, achieving high precision (≥0.89), recall (≥0.80), and mIoU (≥0.73). Visually, the model effectively recreates established corridors and adeptly extends predictive mapping into locally unmapped counties, such as Shiawassee, MI, and Morris, NJ (Fig. 7a), confirming its capability to mend acute jurisdictional gaps.

Fig. 7. Spatial inference and national model generalization.

a Generated maps seamlessly continuing the flood footprint from known NFHL territories into previously unmapped jurisdictions, mitigating sudden administrative cutoffs. b Extrapolating the learned baseline relationships to the completely unmapped surroundings of Grand Rapids, MI. Despite the lack of any local ground-truth data, the model successfully infers hydrologically viable floodplains with high morphological consistency. c Completing the missing flood map using the definitive national model in fundamentally disparate terrain contexts. For regions natively mapped in the NFHL dataset (light blue), the model’s generated results (dark blue) exhibit overwhelming overlap and intersection with established ground truth. Crucially, across all regions entirely unmapped by the NFHL dataset, the model flawlessly extends predictive mapping into data-poor channels, guaranteeing unprecedented resolution continuity. All images are normalized to identical geographic scales. Sources of the map samples: OpenStreetMap contributors. Basemap for (b): OpenStreetMap (https://www.openstreetmap.org/copyright). OpenStreetMap content is published under a CC-BY-SA 2.0 Attribution-ShareAlike 2.0 License (https://creativecommons.org/licenses/by-sa/2.0/). Terrain images: U.S. Geological Survey, 2024, 3D Elevation Program (3DEP) Seamless 1 Arc-Second Digital Elevation Model, accessed April 11, 2025 at URL https://www.usgs.gov/3d-elevation-program/about-3dep-products-services. Flood layers: FEMA NFHL. The geographic center coordinates are indicated within the subfigures where applicable.

Out-of-sample validation (extrapolation)

Extrapolating model parameters into morphologically dissimilar, unseen regions degrades quantitative metrics, with mIoU strictly dropping below 0.50 across all naive local models due to inherent structural differences (Fig. 8a). However, this quantitative decline belies the model’s practical utility in generalizing fundamental hydrological relationships. For instance, despite numerical degradation, the model successfully infers hydrologically viable floodplains across the entirety of the previously unmapped Grand Rapids, MI region (Fig. 7b), proving highly effective at providing a baseline assessment in true data-poor environments.

Fig. 8. Quantitative performance metrics.

a Box plot illustrating model performance degradation and increased inference variability when extrapolating locally specialized parameters to morphologically dissimilar out-of-sample regions. b Variance reduction and performance stabilization achieved by the national model across all regions compared to out-of-sample local models. The national model uniformly demonstrates a higher mean, a higher median, and a narrower interquartile range across disparate environments.

National model superiority

To mitigate the localized overfitting that constrains extrapolation, a comprehensive national model was trained on the complete, spatially heterogeneous repository of available NFHL regions (comprising approximately 300,000 tiles). When deployed in extrapolated regions, the national model systematically outperforms isolated local models, demonstrating concrete performance gains (+0.06 to +0.17 mIoU) alongside reduced prediction variance (Table 2 and Fig. 8b). Because this national model ensures superior generalization across diverse geographic contexts (Fig. 7c), it serves as the foundational flood hazard layer for all contiguous US analyses in this study.

Author contributions

A.N. Wu designed the research, developed the methodology and software, performed the data processing and analysis, created the visualizations, and wrote the manuscript. R. Stouffs provided research funding and project administration, and supervised the work during A.N.Wu’s doctoral studies. Y. Zhang contributed to the research design, structured data analysis, supervised the work, and provided research funding and personnel support. All authors reviewed and edited the manuscript.

Peer review

Peer review information

Nature Communications thanks Zeda Yin and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This research was conducted at the Singapore-ETH Centre, which was established collaboratively between ETH Zürich and the National Research Foundation Singapore (Grant No. FCLG-B2-SEA). This research is also supported by the National Natural Science Foundation of China (Grant No. 20241710038).

Data availability

All datasets used in this study are derived from publicly available sources. The complete AI-generated flood hazard layer for the contiguous United States is released as an open dataset. Available at 10.6084/m9.figshare.29163797.

Code availability

The training and inference code, built on PyTorch, is available at https://github.com/Iceofsky/FloodMapper, along with pretrained model weights and sample data for reproducibility⁴⁵. Analysis notebooks for reproducing all figures and statistical results are available at https://github.com/Iceofsky/national-flood-analysis.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-74336-x.