Groundsource: Google’s Gemini turns news articles into a global flood database

Google Research just dropped something genuinely useful: Groundsource. It’s a framework that uses Gemini to chew through news reports and spit out structured, historical data about natural disasters. No more relying on sparse satellite imagery or manual archives that miss half the picture.

Their first public dataset covers urban flash floods — 2.6 million records across more than 150 countries, stretching from 2000 to the present. That’s not a typo. 2.6 million. Compare that to the Global Disaster Alert and Coordination System (GDACS), which has about 10,000 entries and mostly captures big, high-impact events. The difference is staggering.

Why this matters

Flash floods are nasty. They’re fast, localized, and historically under-documented. Seismic events have global sensor networks. Hurricanes get tracked by satellites. But hydro-meteorological hazards like flash floods? They’ve been stuck in a data desert. Existing archives like the Global Flood Database (GFD) or Dartmouth Flood Observatory (DFO) rely on satellites, which means cloud cover blocks the view, revisit times miss the action, and only large, slow-moving disasters get recorded. Small urban floods? Forgotten.

That’s a problem because you can’t train reliable AI models on sparse data. If your training set only covers 10,000 major events, your model won’t know what a typical Tuesday afternoon downpour in Jakarta looks like. Groundsource fixes this by pulling from a vastly richer source: news articles, government reports, local bulletins — the stuff humans actually write when something happens.

How it works

The core idea is simple but the execution is hard. There’s an enormous amount of unstructured text out there about historical events. No one can read it all manually. So Groundsource uses Gemini to extract structured information — location, date, severity — from news reports. The pipeline processes articles at scale, filtering out noise and keeping only verified signals.

The chart they released shows exponential growth in digitized news and corresponding flood events captured, with a massive density spike between 2020 and 2025. That’s not just because there are more floods or more news; it’s because the methodology scales with available data. The more articles Gemini can digest, the more events get recorded.

The dataset itself

2.6 million records is a lot. But what’s in them? Each record includes geolocation, timestamps, and event details extracted from multiple sources. The dataset covers urban areas specifically, which is where flash floods cause the most damage and where historical data has been weakest. Google is releasing it as open-access, so researchers, urban planners, and emergency responders can all use it.

I’d have liked to see some validation numbers — how accurate is the extraction compared to human-curated datasets? The blog mentions “verified ground truth” but doesn’t give error rates. That’s a gap. Still, even with some noise, 2.6 million records beats 10,000 any day for training robust models.

Potential beyond floods

The framework isn’t tied to floods. Google explicitly says it could be adapted for other hazards — wildfires, landslides, heatwaves. If they scale this to multiple disaster types, it becomes a foundational resource for climate research and disaster preparedness. Insurance companies will love it. Urban planners will actually have data to work with. Forecasters can validate their models against real historical events instead of guesswork.

My take

This is the kind of AI application that doesn’t make flashy headlines but has real impact. No chatbots, no image generators — just a solid methodology that solves a concrete problem. The data scarcity issue has been a bottleneck in hydrometeorological research for years. Groundsource doesn’t eliminate it, but it makes a huge dent.

That said, I’m curious about bias. News coverage isn’t uniform globally. A flood in London gets reported. A flood in rural Bangladesh might not. The dataset likely overrepresents wealthier, more media-saturated regions. Google doesn’t address this in the post, but it’s worth keeping in mind when using the data.

Still, this is a net positive. Open-access, scalable, and focused on a real problem. More of this, please.