Groundsource: Using Gemini to Turn News Reports into a Flood Database

Google Research just dropped something I’ve been waiting to see for a while: a practical, large-scale use of AI to solve a real data problem. It’s called Groundsource, and the idea is refreshingly straightforward.

We all know news articles are full of valuable information about what’s happening in the world. But extracting that info at scale? That’s always been the bottleneck. Google’s team figured out a way to use Gemini to turn unstructured news reports into structured, historical data — specifically for urban flash floods.

The data desert problem

Here’s the thing about flash floods: they’re fast, localized, and often deadly. But unlike earthquakes, which have a global network of sensors, there’s no unified observation system for these hydrometeorological events. We’re essentially flying blind when it comes to historical data.

Existing databases like the Global Flood Database (GFD) or the Dartmouth Flood Observatory (DFO) are useful, but they rely heavily on satellite imagery. That means they miss a lot — clouds block the view, satellites only pass over so often, and smaller, quick-moving floods just never get captured. The UN’s GDACS system has about 10,000 entries, but it focuses on high-impact disasters.

Ten thousand records sounds like a lot until you realize you’re trying to train global-scale AI models. For flash floods, that’s a drop in the bucket. This “data desert” has been a major roadblock for accurate forecasting, and it’s been frustrating to watch because the data is out there — it’s just locked away in news articles and local bulletins.

How Groundsource works

The core innovation here isn’t a new model architecture or some fancy transformer. It’s about using Gemini to extract signals from global news media at scale. Think about it: every time a local paper in Bangladesh reports on a sudden flood, or a government agency in Kenya issues a flash flood warning, there’s a record. Groundsource analyzes these reports and transforms them into a structured event archive.

Right now, the first Groundsource dataset covers flash floods from 2000 to the present, across more than 150 countries. We’re talking about 2.6 million records. That’s not just an incremental improvement — it’s orders of magnitude more data than anything we’ve had before.

I was skeptical about the quality at first. News reports can be noisy, biased, or just plain wrong. But the team seems to have built in verification steps. They’re not just blindly scraping headlines; they’re extracting specific, localized details — dates, locations, severity — and cross-referencing where possible.

Why this matters

This isn’t just an academic exercise. Having this kind of historical baseline is critical for three things:

First, it lets scientists train better hydrological models. Machine learning needs data, and now there’s data to work with. Second, it validates forward-looking climate projections. We can check if our predictions match what actually happened. Third, it has practical applications — urban planning, insurance risk assessment, and emergency response all depend on knowing where floods have historically occurred.

The fact that Google is making this dataset openly available is a big deal. Open-access data like this can level the playing field for researchers and organizations that don’t have Google’s resources. A university in Southeast Asia or a local government in Africa can now use this data to improve their own forecasting.

The bigger picture

What excites me most about Groundsource is that it’s a methodology, not just a one-off dataset. The same approach could be applied to other hazards — wildfires, landslides, heatwaves. If they can extract flash flood data from news, they can do it for anything.

I’ve seen a lot of AI projects that promise the moon and deliver a crater. This one feels different. It’s solving a specific, painful problem with a practical tool. No hype, no grandiose claims about saving the world — just a solid piece of engineering that turns unstructured text into structured knowledge.

Of course, there are limitations. News coverage is uneven globally — a flood in a wealthy country gets more coverage than one in a poorer region. And the data only goes back to 2000, when digital news became more prevalent. But even with those caveats, this is a massive step forward.

I’ll be watching to see what happens when researchers start digging into this dataset. My bet is we’ll see some interesting findings about flood patterns that were previously invisible. And if Google extends this methodology to other disasters, we might finally start closing the data gap that’s been holding back climate resilience efforts for years.

The paper and dataset are linked in the original announcement. Go check them out if you’re into this stuff.