Google has developed a tool that converts decades of public reports and mapping data into a historical flood record, addressing one of the main obstacles to predicting flash floods before they strike.
The tool, called Groundsource, uses Google's Gemini AI models to analyse publicly available documents and extract details of flood events, identifying more than 2.6 million incidents across more than 150 countries.
Google Maps is then used to assign precise geographic boundaries to each event, turning unstructured historical records into a structured dataset that forecasting models can learn from.
That dataset has been used to train a new model aimed at forecasting urban flash floods up to 24 hours in advance, a meaningful step forward in a type of flooding that has proved particularly difficult to predict because it develops rapidly and is heavily shaped by local terrain and drainage infrastructure.
The forecasts are now available through Google's Flood Hub platform, which already provides riverine flood warnings covering 2 billion people across more than 150 countries.
Yossi Matias, vice president and head of Google Research, framed the work in stark terms in a company blog post.
"When disaster strikes, information is a lifeline," Matias wrote.
Related reading
- Google Maps adds AI assistant and 3D navigation in biggest overhaul in a decade
- Google is rolling out Gemini AI built into Chrome browsers in India, New Zealand and Canada
- Google expands AI Max text guidelines globally
Google has released the dataset as an open-source benchmark, making it available to researchers and partner organisations to build on.
The company said the Groundsource approach could be extended beyond flooding to other hazards, including landslides and heat waves, and that the urban flash flood model now sits within its Google Earth AI family of environmental tools.
The recap
- Google releases Groundsource AI to predict urban flash floods.
- Dataset contains 2.6 million historical flood events worldwide.
- Forecasts available in Flood Hub up to 24 hours ahead.