Google has launched Urban Flash Flood forecasts on Flood Hub, using AI to predict flash flood risks in cities up to 24 hours before they occur. The system addresses a critical global gap: while developed nations have robust flood forecasting, over 2 billion people in the Global South lack access to early warning systems that could save lives.

Traditional flash flood forecasting faces unique challenges. Unlike riverine floods, flash floods happen rapidly and unpredictably across urban areas without historical gauge data. To overcome this, Google developed Groundsource, an AI-powered methodology that extracts ground truth from news reports using Gemini to identify past flood events. The system now operates at a 20x20 kilometer resolution using only globally available weather data.

The model employs a recurrent neural network with LSTM architecture, trained on meteorological data and geographic factors like urbanization density and soil absorption. Initial evaluations show performance comparable to the U.S. National Weather Service in regions most vulnerable to flooding, with similar precision and recall metrics in South America and Southeast Asia. The system currently predicts impacts in areas with over 100 people per square kilometer, focusing on where populations are densest.