Groundsource: Google’s Gemini Project That Turned 2.6 Million News Reports Into Flood Data

Google Research just dropped something that actually matters for climate science: Groundsource. It’s a methodology that uses Gemini to chew through news reports and spit out structured historical data about natural disasters. The first dataset they’re releasing covers urban flash floods — 2.6 million events across more than 150 countries, going back to the year 2000.

Let me be blunt: this is the kind of AI application that doesn’t make headlines but saves lives. Historical flood data has been a mess for years. We’ve had satellite imagery, sure, but clouds block it, satellites don’t revisit frequently enough, and the whole setup tends to catch only the big, slow-moving disasters. Flash floods — the ones that kill people in urban areas with almost no warning — slip through the cracks.

Existing databases like the Global Flood Database (GFD) and Dartmouth Flood Observatory (DFO) are useful but limited. The UN’s GDACS system has about 10,000 entries, which sounds like a lot until you realize you’re trying to train global-scale AI models. Ten thousand records is a drop in the bucket when you need to validate predictions across hundreds of countries and decades of weather patterns.

Groundsource’s approach is straightforward in concept but hard to execute: scrape news articles, government reports, and local bulletins, then use Gemini to extract structured data — location, date, severity, affected area — from unstructured text. The core innovation isn’t the AI itself; it’s the pipeline that turns messy human-language reporting into something a hydrological model can actually use.

I’ve seen a lot of “AI for good” projects that sound great on paper but produce garbage data. What makes Groundsource different is that they’re releasing the dataset openly. You can download it, poke at it, and decide for yourself if the quality holds up. That transparency is rare in corporate research, and it’s exactly what climate science needs.

The methodology isn’t limited to floods either. Google’s paper outlines how the same framework could be adapted for other hazards — wildfires, landslides, maybe even disease outbreaks. If they pull that off, we’re looking at a fundamental shift in how we build historical baselines for disaster modeling.

There are caveats, of course. News coverage isn’t uniform — rich countries generate more articles than poor ones, which introduces geographic bias. A flash flood in Tokyo gets reported; one in rural Bangladesh might not. The team acknowledges this, and the dataset includes confidence scores to help researchers account for it.

Still, 2.6 million records is orders of magnitude more than what we had before. For anyone working on flood forecasting, urban planning, or climate adaptation, this is Christmas morning. Go grab the dataset, kick the tires, and see what you can build with it.