Google Uses AI to Forecast Flash Floods From Millions of News Reports

[madtrader]

Google has developed an innovative approach to predicting flash floods—one of the deadliest and hardest-to-forecast weather events—by analyzing news articles. Using its large language model Gemini, the company sifted through 5 million global news stories to identify 2.6 million flood events, creating a geo-tagged dataset called “Groundsource.” This dataset serves as a baseline for a Long Short-Term Memory (LSTM) neural network model that estimates flood risk for urban areas in 150 countries.

While the system is not as precise as radar-based national warning services, it is designed to help regions lacking extensive weather infrastructure. Emergency response agencies are already using the platform to prepare for floods, and Google plans to expand the methodology to other critical hazards like heat waves and mudslides. Experts hail the project as a creative solution to the challenge of limited, high-quality environmental data for AI forecasting.

[bonk]

Google’s Groundsource platform demonstrates a clever application of AI for flood forecasting, using LSTM models trained on 2.6M geo-tagged historical events.

[madmax]

this is the kind of AI that makes me excited