Google uses AI to predict urban floods

Google has created Groundsource, an artificial intelligence methodology that turns old public reports into useful data for anticipating flash floods in cities up to 24 hours in advance.

The project starts with a specific problem: for years, there were not enough detailed records of this type of flooding. Without reliable historical data, it was difficult to train models capable of detecting when and where these events might occur.

How Groundsource works

Groundsource uses Gemini to analyze decades of public disaster reports. The system identified more than 2.6 million historical floods across more than 150 countries.

Google then used Google Maps to calculate the precise geographic boundaries of each event. The result is a dataset focused on urban flash floods, an especially difficult type of emergency to monitor because it can develop quickly on streets, in tunnels and in densely populated areas.

Using this data, Google trained a new prediction model. The company says the system is a tangible step toward anticipating urban floods up to a day in advance, but that does not mean every flood can be predicted accurately or that every alert will correspond to an actual event.

Alerts reach Flood Hub

Predictions for urban flash floods are now available in Google Flood Hub, alongside the forecasts Google already offered for major river floods.

Those river forecasts cover 2 billion people across more than 150 countries for the most significant events. The new feature expands the system's reach to a different risk: water that accumulates rapidly in urban environments, where even a localized storm can disrupt transportation, damage homes or put lives at risk.

For you, the most direct impact will depend on whether your community, local authorities or emergency organizations use these predictions. An earlier alert can help officials close a road, move vehicles, protect equipment or evacuate a vulnerable area before the water rises.

A data bank for other researchers

Google has also added the urban flood model and dataset to its Google Earth AI family. The material is available as an open reference for scientists and collaborators working in regions with few historical records.

The idea could extend to other natural disasters. The same approach could be used to turn verified reports into data about:

• Landslides.
• Heat waves.
• Other events with few structured records.

The important point is not only that AI can analyze the past. By organizing millions of reports and locating them precisely, it can help create the tools authorities need to prepare more effectively. What remains to be seen is how the model performs in real-world situations, what coverage it achieves in each location and how its predictions are turned into useful alerts for the public.