Google trains AI flood forecasting model using 5 million news articles

Google

• Google uses news articles to predict flash floods globally
• Groundsource dataset improves AI flood forecasting accuracy
• Flood Hub now shows urban flood risks in 150 countries

Did our AI summary help?

Flash floods are among the world’s deadliest weather events, killing more than 5,000 people each year. They are also notoriously difficult to predict because they occur quickly and in highly localised areas.

Google believes it has found an unusual way to tackle that challenge — by analysing news reports.

Researchers at the company used Gemini, Google’s large language model, to sift through roughly 5 million news articles from around the world. From those reports, the system identified 2.6 million individual flood events, which were then converted into a geo-tagged dataset called Groundsource.

According to Google, this marks the first time the company has used a language model to extract structured environmental data from large volumes of written news coverage.

The dataset helped address a long-standing problem in weather forecasting. Unlike temperature or river levels, flash floods are too short-lived and localised to be measured consistently with traditional monitoring systems. That lack of reliable data has limited the effectiveness of AI forecasting models.

With the Groundsource dataset as a baseline, Google researchers trained a Long Short-Term Memory (LSTM) neural network to analyse global weather forecasts and estimate the probability of flash floods in specific locations.

The resulting system now feeds into Google’s Flood Hub platform, which highlights potential flood risks in urban areas across 150 countries. The company is also sharing the data with emergency response organisations.

António José Beleza, an emergency response official at the Southern African Development Community, said early trials of the system helped his organisation respond to floods more quickly.

The model still has limitations. It currently predicts risk across areas of about 20 square kilometres, which is relatively low resolution. It is also less precise than systems like the U.S. National Weather Service flood alerts, partly because Google’s model does not use local radar data that tracks rainfall in real time.

However, the system was designed specifically for regions that lack expensive weather-monitoring infrastructure or long-term meteorological records.

Google researchers say that by analysing millions of news reports, the dataset helps fill those information gaps. The company also sees broader potential for the approach. Researchers believe large language models could be used to convert written reports into structured datasets for other difficult-to-measure events, including heatwaves and mudslides.