Google’s AI now forecasts urban flash floods 24 hours ahead — here’s how it works

Google’s Flood Hub has been quietly saving lives for years with riverine flood predictions, but flash floods are a different beast. They kill more people — roughly 85% of flood fatalities globally — and they happen fast. Within six hours of heavy rain, streets turn into rivers. 5,000 people die every year from them.

Now Google is rolling out urban flash flood forecasts on Flood Hub, using a new AI methodology that gives up to 24 hours of advance notice. That’s a big deal. Even 12 hours of warning can cut damage by 60%, according to the World Meteorological Organization.

The invisible flood problem

Riverine floods are relatively easy to predict. You have stream gauges measuring water levels, historical data to train models, and the whole thing unfolds over days. Flash floods? They can happen anywhere, often far from any gauge. In cities, the mix of intense rain, concrete, and drainage systems makes traditional physics-based modeling computationally impossible at a global scale.

The real kicker: there’s no historical record of exactly where and when flash floods occurred. Without that, you can’t train a supervised ML model to predict them. This is where Google’s approach gets interesting.

Training AI on news articles

Instead of waiting for governments to install more sensors, Google built a dataset called Groundsource. They used Gemini to crawl through publicly available news reports about floods, extracting location and timing details with high precision. Then they aggregated those entries into a training dataset of past flash flood events.

Is it perfect? No. News coverage is biased toward populated areas and dramatic events. But it’s better than nothing, and apparently good enough to train a model that can predict flash floods up to 24 hours ahead. The paper is worth a read if you want the technical details.

Why this matters

The “warning gap” between rich and poor countries is stark. Developed nations have radar networks, stream gauges, and dedicated forecasting teams. In the Global South, less than half of developing countries have access to multi-hazard early warning systems. Billions of people get no advance notice at all.

Google’s approach is scalable because it doesn’t require hardware deployment. The model runs on publicly available weather data and the Groundsource dataset. It’s not as precise as a hyper-local system with physical sensors, but it covers places that previously had zero warning capability.

What’s next

The rollout is live on Flood Hub now. Google says this is an expansion of their existing flood forecasting initiative, which already covers over 2 billion people in 150 countries for riverine floods. Adding urban flash floods fills a critical gap.

I’m curious to see how this performs in real-world scenarios. The model was trained on news data, which means it might miss events in areas with limited media coverage. But it’s a start, and it’s better than the alternative — which for many people was nothing at all.