Flash floods kill thousands annually, yet their fleeting, hyper-local nature makes them notoriously hard to forecast. Traditional sensors and models often miss them. In a novel approach, Google Research has turned to an unexpected data source: the global news archive.

The team used its Gemini language model to analyze 5 million news articles, identifying and geotagging 2.6 million flood events reported worldwide. This created a first-of-its-kind dataset, internally called 'Groundsource,' which translates qualitative news reports into a quantitative, time-stamped record of where and when floods actually occurred.

'Groundsource helps rebalance the map,' explained Juliet Rothenberg, a program manager on Google's Resilience team. 'It lets us extrapolate to regions where formal meteorological data is scarce.' This dataset then trained a specialized neural network to process global weather forecasts and generate flash flood probabilities.

The resulting model now provides risk alerts for urban areas in 150 countries via Google's Flood Hub platform, with data shared with emergency agencies. An official with the Southern African Development Community confirmed it hastened response times during trials.

The system isn't perfect. Its resolution is broad, covering 20-square-kilometer zones, and it lacks the precision of systems like the U.S. National Weather Service's, which uses local radar. However, its strength is functioning precisely where those expensive, infrastructure-heavy systems do not exist.

Industry observers see value in the method. 'Data scarcity is a fundamental challenge in geophysics,' said Marshall Moutenot, CEO of Upstream Tech. 'This was a creative approach to get that data.' Google believes the technique—using language models to build datasets from text—could next be applied to phenomena like heat waves and landslides, offering a new lens on hard-to-predict disasters.

Source: TechCrunch