A companion to Wu, Zhang & Stouffs · Nature Communications 2026

The Missing
Flood Maps

FEMA's flood maps guide where America builds, insures, and evacuates — but they cover only a third of the country's rivers. We trained a deep learning model on every map FEMA ever made, then let it draw the rest.

Open the Explorer → How it works
0
Million buildings
in 100-yr flood zones
0
Million buildings
newly revealed by AI
0
Million people exposed
(+69% vs FEMA alone)
0
Of 917 metro areas
inadequately mapped
01 The Explorer — FEMA vs AI, Side by Side

Every 100-year flood zone FEMA has ever published, against every zone the model generated — the entire contiguous United States at 30-meter resolution. Drag the divider to compare. Amber is FEMA; cyan is the AI.

SCROLL TO LOAD MAP · OR OPEN FULLSCREEN →
Fullscreen Explorer →
02 Method — Terrain In, Flood Maps Out

1 — Terrain

The only nationwide input is elevation: the USGS 3DEP digital elevation model at 30-meter resolution, rendered with hillshades and color gradients so the terrain's structure — valleys, fans, floodplains — is visually explicit.

2 — Tiling

Terrain and FEMA's existing flood maps are cut into ~300,000 aligned square tiles on the standard web-map grid, at multiple scales. Every place FEMA has ever mapped becomes a training example: this terrain produced this flood map.

3 — The model

A conditional GAN — the image-translation architecture behind pix2pixHD — learns to draw flood zones from terrain alone. Three discriminators judge its output at three zoom levels; training ends when they can no longer tell generated flood maps from FEMA's.

4 — Judged against reality

Where FEMA maps exist, the model's output can be scored directly. Out of sample, the national model reaches 0.81 pixel precision in Texas and New Jersey — and unlike locally trained models, it holds up in terrain it has never seen.

5 — The whole country

The trained model then runs over all of CONUS — including the two-thirds of river miles FEMA never modeled. The tiles are stitched and vectorized into a single continuous flood layer, in the same format as the NFHL itself.

03 What the Completed Map Shows

Laying the AI layer over the NFHL nearly doubles the exposure count. These are people and buildings inside 100-year flood zones that no regulatory map currently shows.

FIG. 01 — EXPOSURE INSIDE 100-YR FLOOD ZONES
FIG. 02 — MAPPING ADEQUACY, 917 US METRO & MICRO AREAS

Validation

Against held-out FEMA maps in three geomorphologically distinct states, the single national model trades a little in-sample precision for markedly better generalization:

Validation metrics by region: precision, recall and mIoU for in-sample local models and the out-of-sample national model.
Region Local model, in-sample National model, out-of-sample mIoU gain
PrecRecmIoUPrecRecmIoUvs local
04 Three Cities, Three Gaps
MAP LOADS ON SCROLL
Notice — Read Before Using These Maps
  • These maps show fluvial and coastal 100-year flood extents. They largely exclude pluvial (rainfall-driven) flooding, because the FEMA maps the model learned from exclude it.
  • They are a research product, not a replacement for regulatory FEMA flood zones. For insurance, lending, or permitting decisions, consult the FEMA Map Service Center.
  • Out-of-sample accuracy varies by terrain (mIoU 0.48–0.54 across validation states); treat any single location as an indication, not a determination.