Google AI Transforms News for Global Flood Prediction

Leveraging its Gemini large language model, Google analyzed approximately five million global news articles, identifying 2.6 million distinct flood incidents. This data was then structured into a geo-tagged time series dataset called Groundsource, marking the first time Google has used an LLM to convert written reports into such a large-scale structured dataset.

This innovative method addresses the critical lack of consistent meteorological data in many regions, which traditionally hinders deep learning models. The resulting forecasting system, built on a Long Short-Term Memory neural network, now identifies flood risks for urban areas across 150 countries via Google’s Flood Hub platform, sharing crucial data with emergency response agencies.

While the model currently has a resolution of about 20 square kilometers and lacks real-time radar data, it is specifically designed to assist regions without expensive weather monitoring infrastructure. This initiative not only enhances disaster preparedness but also demonstrates the potential for LLMs to transform qualitative data into quantitative insights for other phenomena like heatwaves and mudslides, addressing a significant data scarcity challenge in geophysics.