Google has introduced a new system, named Groundsource, designed to tackle one of meteorology's toughest challenges: predicting flash floods. The tool's innovation lies in its source material. Engineers used the Gemini language model to analyze a global archive of 5 million news articles, extracting and mapping historical flood reports into a chronological, geo-tagged database.

This historical record is then fed, alongside standard weather forecasts, into a machine learning model. The system assesses current conditions against patterns from the past to generate flood risk probabilities. While Google hasn't released specific accuracy metrics, an early tester reported it allowed for faster response to local weather threats. The company is currently displaying risk assessments for urban areas in 150 countries on its Flood Hub platform and providing data to emergency agencies.

The approach has constraints. Its resolution is limited to 20-square-kilometer zones, and it lacks the precision of systems like the U.S. National Weather Service's, which uses real-time local radar. Instead, Groundsource is built for regions where such advanced sensor infrastructure is unavailable.

"We're aggregating millions of reports," explained Juliet Rothenberg of Google's Resilience team. "It enables us to extrapolate to other regions where there isn't as much information." She noted the method could later be adapted for other hard-to-predict events like heat waves and mudslides.

This marks the first application of a language model in Google's weather work, distinct from its highly accurate DeepMind WeatherNext 2 AI forecasting system. It represents a different tack: using documented history to inform warnings for the future.

Source: Engadget