Google Uses News to Forecast Global Floods

Flash floods, a major global hazard, are now being predicted by Google's artificial intelligence using an unconventional data source: news articles. Researchers utilized Google's Gemini large language model to sift through five million global news reports, identifying 2.6 million distinct flood events. This vast dataset, named Groundsource, provides crucial real-world data where traditional meteorological monitoring is insufficient.

This Groundsource data then trained a Long Short-Term Memory neural network, enabling it to ingest global weather forecasts and predict flash flood probabilities. Google's Flood Hub platform now displays these risk assessments for urban areas in 150 countries. The system shares this information with international emergency response agencies, enhancing their ability to react swiftly to developing flood situations.

While the model operates at a 20-square-kilometer resolution and doesn't use local radar data, its strength lies in providing forecasts for regions with limited meteorological infrastructure and data records. This approach helps to "rebalance the map," extending predictive capabilities to underserved areas. Google envisions applying similar LLM techniques to forecast other critical phenomena like heat waves and mudslides.