BRENNEN

Industries / Energy, water and waste / Environmental services / Stormwater management

Flood models and warnings

Working out where rain will run and pool in a town, with runoff models, rain radar and street sensors, and warning people in time to get out of the way.

7

Milestones achieved, from 1932 to 2024.

1

To watch for: announced, not yet confirmed outside.

3

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1932

    Turning rainfall into flow

    L. K. Sherman set out the "unit graph" in Engineering News-Record (volume 108, pages 501 to 505): the flow that one unit of runoff, such as one inch spread over an area in a set time, produces at a point downstream. Once an area's unit graph is known, the flow from any storm can be found by adding up scaled copies of it. The method is the basis of the hydrograph chapter of the US Natural Resources Conservation Service's engineering handbook and the National Weather Service's river forecast software.

    weather.govdamtoolbox.org

    Credit

    L. K. Sherman

  2. 1971

    A model for city runoff

    The US Environmental Protection Agency (EPA) first released its Storm Water Management Model (SWMM) in 1971. It splits a town into small areas that catch rain, works out the runoff and pollution from each, and follows that water through pipes, channels, ponds, pumps and overflows. Version 5, a full rewrite developed with the engineering firm CDM Smith, is free and open source; EPA's latest release, 5.2.4, came out in August 2023, and the manual says the model is used "throughout the world".

    epa.govepa.gov

    Credit

    US EPA; CDM Smith (version 5)

  3. 2017

    Warnings by pooling water

    From the 2017 rainy season, the Japan Meteorological Agency (JMA) stopped deciding its heavy rain warnings for flooded streets and buildings on rainfall totals alone. It uses a "surface rainfall index", a number for how much rain is collecting on the ground, computed every 250 m from elevation and land-use data, with separate runoff models for built-up and open land. Its "Kikikuru" maps show, at 1 km detail, where the index is forecast to reach warning level.

    jma.go.jpjma.go.jp

    Credit

    Japan Meteorological Agency

  4. 2021

    AI rain forecasts judged

    DeepMind and the UK Met Office built a generative AI model that forecasts rain from radar 5 to 90 minutes ahead (a "nowcast") over an area of 1,536 by 1,280 km. More than 50 expert meteorologists ranked it first for accuracy and usefulness in 89% of cases against two competing methods (Nature, September 2021). Earlier AI nowcasts had blurred at longer lead times and did poorly on medium to heavy rain, the kind that floods streets.

    Lessona forecast can score well on average and still miss the heavy downpours that matter most, so the team asked forecasters to judge usefulness as well.

    ebi.ac.uk

    Credit

    Suman Ravuri, Shakir Mohamed and colleagues at DeepMind, with Met Office scientists

  5. 2021

    Zhengzhou's warnings go unheeded

    On 20 July 2021 a station in Zhengzhou, China, recorded 624.1 mm of rain, almost a year's worth, in one day. China's State Council investigation, published in January 2022, counted 398 people dead or missing in Henan province, 380 of them in Zhengzhou and its county-level cities. Fourteen died on metro Line 5 after floodwater broke through temporary fences at a rail yard, and six drowned in the Jingguang North Road tunnel, which was not closed until 4:16 p.m. The report found that city officials ignored a series of early warnings from the city's weather bureau, and named weak links in "the mechanism connecting early warning and response".

    Lessona warning saves lives only when someone is bound to act on it, by closing tunnels and stopping trains.

    english.www.gov.cnsixthtone.com

    Credit

    State Council investigation team

  6. 2021

    Ida floods basement homes

    On 1 September 2021 the remnants of Hurricane Ida dropped intense rain on New York City. The National Weather Service issued the city's first flash flood emergency at 9:28 p.m., and the city declared a state of emergency at 11:45 p.m. A 2024 study by the city's health department found 14 deaths tied to the storm; 11 people drowned in flooded basement apartments, 10 of them in basements that could not lawfully be lived in. In the two cases where the time of injury was recorded, it came before the city's emergency declaration, which had no advice for basement residents. The authors recommended basement flood alarms and warnings in more languages.

    Lessona citywide warning did not reach people asleep below street level.

    cambridge.org

    Credit

    Ariel Yuan, Ariel Spira-Cohen, Carolyn Olson and Kathryn Lane, NYC Department of Health and Mental Hygiene (measurement)

  7. 2024

    England counts surface water risk

    The Environment Agency's new National Flood Risk Assessment, its first full update since 2018, drew on its own data and local councils' models. Published on 17 December 2024, it puts about 4.6 million properties in England in areas at risk of flooding from surface water (rain that cannot drain away), 43% more than the previous assessment, against 2.4 million at risk from rivers and the sea. About 1.1 million are at high risk from surface water, three times the previous estimate of 344,100.

    Lessonin England, rain falling on towns now threatens more homes and businesses than rivers and the sea combined.

    gov.uk

    Credit

    Environment Agency, with lead local flood authorities

Watch for

  1. Flash flood forecasts from news

    Google Research says Flood Hub now forecasts the chance of flash floods in urban areas up to 24 hours ahead, on 20 by 20 km cells wherever more than 100 people live per square kilometer. The model was trained on "Groundsource", a set of past floods that Google's Gemini model pulled from public news reports, which Google has released on Zenodo. In Google's own comparison, with warnings mapped onto the same 20 km cells and 24-hour windows as its model, US National Weather Service flash flood warnings had 22% recall (the share of floods that were warned) and 44% precision (the share of warnings followed by a flood); Google calls the precision figure an underestimate and does not say which flood records it scored the warnings against.

    Achieved whena national weather service or independent researchers score the forecasts against local flood records in named cities and publish hit and false-alarm rates.

    research.google

    Claimed by

    Google Research (Oleg Zlydenko and Deborah Cohen), March 2026

Open

  1. 01

    Score warnings against street sensors

    A public scorecard for New York City that matches every National Weather Service flash flood warning since late 2020 with the street flooding that FloodNet's sensors measured inside the warned area. It counts warnings followed by measured flooding, measured floods that had no warning, and the minutes of notice each flood got.

    The brief
    Credit · unclaimed
  2. 02

    Flood maps from traffic cameras

    A model that reads New York City's public traffic camera stills every few minutes during storms, marks roads under water on a public map, and is checked against FloodNet sensors within sight of each camera.

    The brief
    Credit · unclaimed
  3. 03

    Test synthetic drain models

    Drain-network models of 10 New York City neighborhoods generated from open data, run with radar rainfall for past storms, and scored on whether they flood where FloodNet measured flooding.

    The brief
    Credit · unclaimed

Open 01 · the brief

Score warnings against street sensors

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public scorecard for New York City that matches every National Weather Service flash flood warning since late 2020 with the street flooding that FloodNet's sensors measured inside the warned area. It counts warnings followed by measured flooding, measured floods that had no warning, and the minutes of notice each flood got.

Why it is still open

FloodNet, run by New York University and the City University of New York, has logged more than 2,400 flood events since its first sensor went in during October 2020 and reached 400 ultrasonic sensors in July 2026, and the Weather Service's New York office uses its data in warning decisions. A search found no published check of the warnings against the sensors. Google's figures for US warnings were worked out on 20 km cells, not against street measurements.

engineering.nyu.eduresearch.google

Built with what exists

FloodNet's flood records on NYC Open Data and its archive on Zenodo; the Iowa Environmental Mesonet's archive of Weather Service warning areas; Python with GeoPandas.

  1. First step

    Download one summer of FloodNet flood records and the matching warning areas, and score the warnings for a single storm by hand.

  2. Done looks like

    A published table covering every warning and every sensor-recorded flood from October 2020 to the latest storm season, with the hit rate, the false-alarm rate and the median minutes of notice, and code that reruns it each month.

Open 02 · the brief

Flood maps from traffic cameras

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A model that reads New York City's public traffic camera stills every few minutes during storms, marks roads under water on a public map, and is checked against FloodNet sensors within sight of each camera.

Why it is still open

A 2023 NYU student project mentored by FloodNet set out to build a labeled flood set from the city's traffic cameras and a real-time alert prototype for open release on Zenodo, but a search of Zenodo found no such set. A Syracuse University team wrote in 2026 that "there is no existing technology that allows for real-time spatially continuous monitoring of urban pluvial flooding". Research models such as V-FloodNet, which outlines floodwater in video, exist, but none found was scored against street sensors.

engineering.nyu.edumeetingorganizer.copernicus.org

Built with what exists

The city transportation department's public traffic camera feeds; a vision-language model such as Claude or Gemini for first-pass labels checked by a person; an open segmentation model such as V-FloodNet, or a small vision model fine-tuned on one graphics chip (see the Note "One GPU to fine-tune"); FloodNet depths as ground truth.

  1. First step

    List the traffic cameras within 100 m of a FloodNet sensor and save their stills through the next forecast heavy rain.

  2. Done looks like

    At least 5,000 labeled camera frames released under an open license, and precision and recall for "road flooded" on frames from at least 20 cameras within 100 m of a FloodNet sensor, across at least 10 storms.

Open 03 · the brief

Test synthetic drain models

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Drain-network models of 10 New York City neighborhoods generated from open data, run with radar rainfall for past storms, and scored on whether they flood where FloodNet measured flooding.

Why it is still open

SWMManywhere, from Imperial College London, generates a SWMM model of a town's drains from open map and elevation data, and Agentic SWMM, an open AI-agent workflow released in 2026, ran it for about 1 km² of Midtown Manhattan from a single sentence. Its authors call that result "not a calibrated or validated network" and list calibration as the next step.

github.comgithub.com

Built with what exists

SWMManywhere; Agentic SWMM (MIT license); EPA SWMM 5.2 and PySWMM; NOAA's Multi-Radar Multi-Sensor rainfall; FloodNet records; an AI coding agent such as Claude Code.

  1. First step

    Generate a model for one neighborhood with at least three FloodNet sensors and run it for the wettest storm in their record.

  2. Done looks like

    For 10 neighborhoods and 20 storms, a published table of predicted against measured flooding at each sensor, with hits, misses and false alarms, and code that others can rerun.