Groundsource: Google’s Gemini turns news articles into flood data

Google Research just dropped something that actually made me sit up straight: Groundsource. It’s a framework that uses Gemini to chew through news reports and spit out structured, historical data about natural disasters. And they’re not just talking about it — they released a dataset of 2.6 million flash flood events from over 150 countries, going back to the year 2000.

Let me unpack why this matters.

The data desert problem

If you’ve ever tried to train a model on historical flood data, you know the pain. Earthquakes have global sensor networks. Floods? Not so much. We’ve got satellite-based archives like the Global Flood Database and the Dartmouth Flood Observatory, but they come with real limitations: clouds block the view, satellites only pass so often, and they mostly catch big, slow-moving disasters. The Global Disaster Alert and Coordination System (GDACS) — a UN and EU joint effort — has about 10,000 entries. That sounds like a lot until you realize flash floods happen every damn week somewhere in the world, and most of them never make it into any structured database.

This is what the team calls a “data desert.” And it’s a real problem because you can’t build reliable predictive models without historical ground truth. You can’t validate forecasts against events that were never recorded.

Turning news into signal

Groundsource’s trick is straightforward in concept but hard to pull off at scale: read news articles, government reports, and local bulletins, then extract structured information — location, date, severity — using Gemini. The pipeline processes text from thousands of sources, filters out noise, and geolocates events. The result is a dataset with 2.6 million records, which is two orders of magnitude larger than GDACS.

What I find interesting is the focus on urban flash floods specifically. These are the events that kill people in cities because they happen fast and catch everyone off guard. Traditional databases miss most of them because they’re too localized or too short-lived for satellites to capture. But local news covers them. A newspaper in Jakarta or a regional bulletin in the Philippines will report on a sudden street flood that killed three people — and Groundsource picks that up.

The chart they released shows an exponential growth in captured events from 2020 to 2025, which makes sense: more digitized news, better models, and likely more flooding in urban areas. But I’d be curious about the false positive rate. News reports can be sensational, and one person’s “flash flood” is another’s “heavy rain.” The paper should address this, but I haven’t dug into it yet.

Why this is different

This isn’t the first attempt to mine news for disaster data. Projects like the Global Event Database (GDELT) have been doing similar things for years. But GDELT is broad — it covers everything from protests to terrorist attacks. Groundsource is laser-focused on extraction quality for a specific hazard type. And because it’s built on Gemini, it can handle multilingual sources and nuanced descriptions that keyword-matching would miss.

More importantly, Google is releasing the dataset openly. That’s a big deal. If you’re a researcher in a developing country with no budget for satellite imagery, you can now download 2.6 million flood events and start building your own models. This could democratize flood forecasting in places that need it most.

The methodology is also reusable. They mention it could be adapted for other hazards — wildfires, landslides, heatwaves. If they pull that off, this becomes a platform, not just a dataset.

The catch

I have one reservation. Groundsource relies on news reports, which means its coverage is biased toward events that get reported. A flash flood in a wealthy suburb with a local newspaper is more likely to be captured than one in a rural slum where nobody files a story. The dataset is global, but it’s not uniform. Google acknowledges this in their paper, but it’s worth keeping in mind if you’re using this data for anything sensitive.

Also, the data only goes back to 2000. That’s 25 years, which is decent for climate modeling but not enough to capture multi-decadal cycles. You’d need longer records to separate climate change signals from natural variability.

Still, this is a solid step forward. We’ve been talking about using NLP for disaster data for a decade, but this is the first time I’ve seen it done at this scale with open access. I’ll be watching to see how the community uses it — and whether Google follows through with datasets for other hazards.