Flash floods are among the most dangerous and fastest-moving natural hazards, and forecasting them well has long meant choosing between local precision and global reach. A flash-flood model developed by Google Research aims to close that gap, predicting the probability of an urban flash flood within the next 24 hours across much of the world. It draws on the kind of historical event data described in this companion piece on turning news reports into structured flood data.
The scaling challenge: local precision versus global reach
Specialized, hyper-local early-warning systems have been built for flash floods in specific cities, with examples in Florida (USA), Barranquilla (Colombia), Manila (Philippines), Nakhon Si Thammarat (Thailand), Mayagüez (Puerto Rico), and Barcelona (Spain). These systems typically rely on networks of physical sensors that monitor variables such as gauge- and radar-estimated precipitation, water levels, and flow velocity. They are highly accurate for the locations they serve, but they are hard to scale because of the cost of hardware, the site-specific calibration each one requires, and the engineering expertise needed to run them.
At the other end of the spectrum, large-scale initiatives such as the World Meteorological Organization’s Flash Flood Guidance System (FFGS), the European Runoff Index based on Climatology (ERIC) flash-flood indicator, and the U.S. National Weather Service’s flash-flood warning system provide broad coverage through remote sensing and numerical weather models. Their global reach is limited, however, by a reliance on high-resolution hydrological maps and radar-based forecasts that are largely unavailable across much of the Global South, and by the need for professional hydrologists to interpret model output and issue actionable warnings.
A model built for near-global coverage
To reach as much of the world as possible, the model uses only globally available inputs: satellite precipitation products such as NASA IMERG and NOAA CPC, real-time forecasts from the ECMWF Integrated Forecast System (IFS) high-resolution (HRES) atmospheric model, and Google DeepMind’s AI-based medium-range global weather model. A neural network based on long short-term memory (LSTM) ingests a seven-day history of meteorological and geophysical data together with a 24-hour forecast, and outputs the probability of a flash flood in a given area. The system operates at a spatial resolution of about 20 by 20 kilometers — a limit set mainly by the resolution of the global data sources — and focuses on urban areas with population densities above roughly 100 people per square kilometer. Its forecasts are surfaced for cities in around 150 countries through Google’s Flood Hub platform and shared with emergency-response agencies.
Limitations and what to watch
The trade-off for global reach is reduced local detail. At a 20-kilometer resolution the model cannot resolve street-level flooding the way a dedicated local sensor network can, and it does not incorporate local radar, which would allow real-time tracking of moving rainfall. It is designed for floods caused by weather, so it does not anticipate those triggered by infrastructure failures such as dam or levee breaks. As with any probabilistic forecast, outputs indicate elevated risk rather than certainty and are most useful alongside local knowledge and official guidance. Even with these constraints, extending credible flash-flood forecasting to regions that previously had none is a meaningful step for disaster preparedness.
