Google Uses Gemini to Build Massive Flood Prediction Dataset

Google Research just dropped Groundsource — a geo-tagged time series dataset built to help predict one of nature's nastiest killers. Flash floods claim more than 5,000 lives annually and remain notoriously hard to forecast.

Here's how they built it: Google pointed Gemini at 5 million historical news articles and had the AI extract 2.6 million flood events, each tagged with location and time data. The result is a massive structured dataset that didn't exist before.

The approach is clever. Instead of relying on patchy sensor networks or satellite imagery alone, Google essentially turned decades of global journalism into usable climate data. Gemini did the heavy lifting of reading, parsing, and geolocating millions of articles at a scale no human team could match.

Groundsource could become a critical tool for disaster preparedness researchers working to get ahead of deadly flash floods.