Ask a Scientist: How can researchers use AI to predict a flood?

Google’s Flood Forecasting Initiative uses AI to predict riverine floods up to seven days ahead and urban flash floods up to 24 hours before they occur, leveraging global datasets on rainfall, river levels, and ground conditions. These predictions are accessible through the Flood Hub tool and the Floods API, which provide early warnings to communities and organizations worldwide. The initiative began in India in 2018 and has since expanded to cover over 150 countries, where more than 2 billion people reside.

To address gaps in urban flood data, Google developed Groundsource, an AI-powered methodology that analyzes over 5 million historical news reports to create a dataset of 2.6 million flood events across 150 countries. This data was integrated into a new urban flash flood model within Flood Hub, enabling predictions even in areas lacking local sensors. The approach marks a shift from traditional models that rely on local historical water level data for calibration.

Researchers and aid groups, such as Give Directly, have used Flood Hub and the Flood Forecasting API to deliver early cash assistance to flood-prone regions like Kogi, Nigeria. The initiative reported that recipients were better prepared for floods, with incomes more than doubling, food insecurity dropping by 90%, and 93% feeling more secure. The models currently focus on urban flash floods due to data availability, but Google is exploring extensions to rural and coastal flooding.

Google has open-sourced its hydrology framework and made the Groundsource dataset and Flood Forecasting API publicly available to support disaster resilience research. The goal is to improve predictions for other natural disasters, such as heat waves and mudslides, while enabling National Meteorological and Hydrological Services to tailor forecasts for local needs. The initiative reflects Google’s broader effort to use AI for accurate natural disaster prediction and community safety.