Google’s AI extends flash-flood warnings to cities up to 24 hours ahead

Flash floods are among the world’s deadliest natural disasters, and their cruelty lies in their speed: water can rise in minutes, leaving almost no time to warn people or move them to safety. They account for roughly 85% of flood-related fatalities worldwide and take more than 5,000 lives each year. On March 12, 2026, Google Research described progress on an artificial-intelligence method built to give people in cities something they rarely have with flash floods — a few crucial hours of warning.

The new approach forecasts urban flash-flood risk up to 24 hours in advance. Its core innovation is a methodology the team calls Groundsource, which uses the Gemini AI model to extract historical flood information from news reports — building a picture of where and under what conditions floods have struck in places that lack the dense sensor networks and long hydrological records available in wealthy regions. Under the hood, the forecasting model uses a recurrent neural network with a long short-term memory (LSTM) unit, and draws on weather forecasts from NASA, NOAA, the European Centre for Medium-Range Weather Forecasts and Google DeepMind.

“They account for roughly 85% of flood-related fatalities worldwide and take more than 5,000 lives each year.”

This urban work extends a broader effort. Google’s Flood Forecasting Initiative already provides riverine flood warnings covering more than two billion people across 150 countries, but flash floods in dense urban centers had remained a stubborn gap because they unfold too fast and too locally for river-based models. Targeting cities — defined as areas with population densities above 100 people per square kilometer — aims to reach the majority of the world’s population where they are most concentrated and most exposed.

The value of lead time is not abstract. Even a modest head start dramatically reduces harm: researchers note that a 12-hour warning can cut flash-flood damage by about 60%, giving families time to protect belongings, move vehicles and, above all, get to higher ground. For the many low- and middle-income cities that have historically lacked reliable flood warning, an AI system that learns from news reports rather than requiring costly new infrastructure is a pragmatic and equitable way to close the gap.

As with any forecasting system, accuracy will vary by place, and warnings only save lives if they reach people and are trusted and acted upon. Google frames this as ongoing research, with work continuing to sharpen spatial resolution and to extend coverage toward rural populations. But the aim is unambiguously humane: turning the hardest floods to predict into ones people can, increasingly, see coming.