Google AI predicts river floods 7 days out, urban flash floods 24 hours out

The numbers come with conditions worth noting. According to Google's post detailing the global rollout of the model, only 100 countries have forecasts validated against real ground-truth data — the rest rely on "virtual gauges," figures the model calculates without an actual monitoring station on site. At the time of that post, Flood Hub carried nearly 250,000 forecasting points, and the population able to receive forecasts had grown from 460 million to 700 million. The figure of 150 countries and 2 billion people refers to everyone living in the forecast coverage area — a mix of validated and unvalidated forecasts combined.

It started as a small experiment in India in 2018

The project wasn't global from day one. Google first piloted a flood forecasting model using real-time river data in India back in 2018, then expanded its reach as the underlying AI for processing data improved. The scale of that expansion shows up in the number of gauges used for training. The early model trained on streamflow records from 5,680 stations sourced from the Global Runoff Data Centre (GRDC); the updated version nearly tripled that, mainly by adding data from the public Caravan dataset, reaching 15,980 gauges.

What made this expansion possible is a model architecture designed to forecast using global weather, precipitation, and land-surface data — even in regions with no accumulated local measurements. Traditional flood models typically had to be calibrated against a region's own historical water-level records, but the places that need alerts most often lack exactly that kind of record, Cohen explains.

What actually happens inside the model

What Google calls "the model" is really two models with different jobs. The hydrologic model takes in weather data and predicts how much water will flow through a river; the inundation model then takes that streamflow output and maps which areas will actually end up underwater. The first model answers "how much will the river rise?" The second answers "so, will where I live flood?"

The backbone of the hydrologic model is an LSTM neural network built for time-series data. What's notable is that inputs aren't just dumped in together. Each weather data product gets its own embedding network, and their outputs are combined before feeding into the LSTM. Precipitation and temperature time series from NASA IMERG, NOAA CPC, and ECMWF ERA5-Land are each produced differently, so the design routes each through its own channel rather than treating them as interchangeable. On top of that, static watershed attributes — terrain, soil, and climate indices from HydroSHEDS — are added in, along with input from Google DeepMind's medium-range weather forecasting model. Training used reanalysis data from 1980 to 2023, with hindcast validation run over 2016–2023.

The output format is also worth noting. Rather than a single value like "tomorrow's flow will be X," the model produces the parameters of a probability distribution for streamflow. It expresses predictions as a mixture of asymmetric Laplace distributions (CMAL) — and because the output is a distribution rather than a point estimate, you can calculate the probability of exceeding any given water level. When to sound an alert then becomes a matter of where you set the threshold on that distribution. Model performance is evaluated using the Nash-Sutcliffe Efficiency (NSE), a standard metric in hydrology.

What "a 5-day forecast as accurate as today's forecast" actually means

The significance of the lead-time figures only becomes clear when you compare them against a baseline. In a 2024 Nature paper on extreme flood prediction in ungauged basins, the Google team pitted its model against GloFAS, an established European system. For 2-year-return-period floods, the AI model outperformed GloFAS at same-day forecasting at 70% of gauge locations (N=3,673); at the 5-year return period that figure was 60%, and at the 10-year return period, 49%.

The more striking number is on the time axis. The paper reports that the AI model's 5-day-ahead forecast was either better than or statistically indistinguishable from GloFAS's same-day forecast for 1-year, 2-year, and 5-year return-period floods. At the 1-year return period, the AI model performed significantly better; for 2-year events, the difference was essentially zero (N=2,162, P=0.98). In other words, the same level of accuracy was pushed five days earlier — which translates directly into more time for evacuation and resource deployment.

Lead times have kept growing since. The November 2024 update stated that its 7-day forecast accuracy matched the previous model's 5-day accuracy, and v2, open-sourced in June 2026, extended the reliable forecast window by 6 more days in gauged basins and 1 more day in ungauged basins compared to the prior version.

Rivers overflowing and cities flooding turned out to be different problems

When Google first built Flood Hub, it could only predict when rivers would overflow. River flooding has a vast amount of accumulated global data behind it, but sudden urban flash floods have almost no comparable historical record to draw on, Cohen says — which is why integrating urban flash flooding into Flood Hub took longer.

The urban model also uses an LSTM, like the river model, but it's looking at different inputs. Instead of watershed terrain, it takes in static attributes like urbanization density and soil absorption rate, and pulls real-time forecasts from ECMWF's Integrated Forecasting System (HRES) and Google DeepMind's weather model. Its spatial resolution is set at 20×20 kilometers — a grid size chosen to match data that's available anywhere in the world. It applies to urban areas with population density above 100 people per square kilometer.

The process of turning news articles into data

Groundsource was built to fill that gap. Released in March 2026, it's an AI methodology that analyzes public disaster-related data — historical news reports, in particular — and turns it into a high-quality data archive. According to Cohen, Gemini read more than 5 million flood-related news reports spanning 20 years to build a dataset of 2.6 million historical flood events.

According to Google Research's published methodology, the process breaks down into four steps. First, news articles primarily about flooding are filtered out, and Google's "Read Aloud" user agent extracts just the article body from text in 80 languages. That text is then standardized into English via the Cloud Translation API, and Gemini makes three judgment calls.

• Classification: distinguishing articles about actual, ongoing floods from articles discussing future warnings, policy meetings, or risk modeling.
• Temporal inference: anchoring relative time expressions like "last Tuesday" to the article's publication date to pin down when the event actually occurred.
• Spatial identification: extracting place names mentioned in the article and mapping them to standard polygons, a step that uses the Google Maps Platform.

The result is a record of 2.6 million events spanning more than 150 countries from the year 2000 to the present. Google also published quality metrics: 60% of events had both location and timing correct, and 82% were accurate enough for practical analytical use. It also reported capturing 85–100% of major flood events logged in the Global Disaster Alert and Coordination System (GDACS) between 2020 and 2026.

Category	River flood forecasting	Urban flash flood forecasting (Groundsource)
Forecast horizon	Up to 7 days ahead	Up to 24 hours ahead
Model architecture	Per-product weather embeddings + LSTM, probability distribution output	Static urban attributes + LSTM
Key data	Weather, precipitation, land-surface conditions; 15,980 gauge stations	2.6 million public disaster records, including historical news reports
Spatial unit	~250,000 forecast points (as of November 2024)	20×20 km grid
Release	Expanded from 2018 India pilot	March 2026

What this approach still can't do

Google is upfront about its limits too. The 20×20 km resolution used for urban forecasting isn't fine-grained enough to flag "the low-lying spot in your neighborhood" — it's closer to a citywide alert. There's a caveat on the accuracy figures as well. Google notes that because unreported real floods exist, correct alerts can get misclassified as false positives, which likely means the published precision figures underestimate true performance. Put another way, the entire validation baseline is built on news coverage itself.

Regional disparities remain, too. In many African countries, there isn't enough independent ground-truth data outside of Groundsource to even estimate model accuracy against, and countries with fewer than 10 verified ground-truth events were excluded from performance reporting altogether. As a comparison baseline, Google recalculated the U.S. National Weather Service's flash flood warnings using the same 20×20 km, 24-hour criteria, arriving at a recall of 22% and precision of 44% — and noted that this figure, too, is likely an underestimate for the same reason.

Where to check this today

The forecasts are available through a few channels. The Flood Hub page displays forecast alerts on a map, built from global weather data, and the same forecasts also surface in Google Search when you search for local flooding. There's also a dedicated channel for relief and disaster-response organizations, not just individual users: the Floods API lets organizations pull forecast data to issue alerts and prepare support before serious flooding hits. Cohen cited GiveDirectly, a charity that used the API to distribute cash in Nigeria's Kogi State last year before water levels actually rose. GiveDirectly's own follow-up found recipient household incomes more than doubled, food insecurity dropped by 90%, and 93% of recipients said they felt better prepared for the next flood.

For anyone who wants to look under the hood, the code is public. In June 2026, Google open-sourced its hydrology modeling framework — a PyTorch package on GitHub under the Apache 2.0 license. It includes both the original model used in the 2024 benchmarking paper and the v2 currently running on Flood Hub, a pipeline for training on historical river data, and tutorial notebooks, with training data drawn from the public Caravan dataset. Google says the goal is to let national weather and hydrology agencies plug AI flood forecasting into their existing workflows without having to send their own data outside their systems.

Editor's take

Treating this announcement as a single feature misses the point. What Google is actually proud of here isn't the accuracy numbers — it's the fact that the model forecasts even where there's no data. The places that need flood warnings most are usually low-income countries or rural regions without gauges or historical water-level records. Traditional hydrologic models had to be calibrated with local ground-truth data; this approach flips that design philosophy entirely, training on global data so the model can fill in the gaps itself.

It's also worth paying attention to how Groundsource sources its data — unstructured public records like news reports. Rather than building new weather-monitoring infrastructure, this approach recycles text records that already exist to close data gaps, and that logic could plausibly extend beyond floods to other disasters with weak local measurement infrastructure, like landslides or heat waves. That said, using news as the data source means regions and events the press covers less get thinner data too — meaning gaps in media coverage can carry straight through into gaps in the model.

From a domestic perspective, the approach itself is more instructive than the tool. It's a working example of how disaster-response agencies or local governments, even without their own sensor networks, can build forecasting models using public weather and news data as raw material. The open-sourced hydrology modeling framework in particular could be a starting point for local research teams to dig into and adapt to local conditions. The Apache 2.0 license leaves few restrictions on commercial use, and the fact that it's built to accept additional data — the way Caravan does — matters a lot in practice.

In the coming months, it's likely the team will announce further expansion into other natural disasters they've already mentioned, like heat waves or landslides. If the underlying method — building forecasting models starting from data-scarce regions — has already been validated, applying it to the next disaster type is largely a matter of reusing the methodology.