Google’s AI now predicts urban flash floods up to 24 hours ahead

Flash floods are nasty. They turn city streets into raging rivers in under six hours, kill over 5,000 people every year, and account for roughly 85% of flood-related deaths globally. The World Meteorological Organization has the numbers, and they’re grim.

Early warning systems help. A 12-hour heads-up can cut flash flood damage by 60%. But there’s a massive gap: rich countries have decent forecasting, while most of the Global South doesn’t. Less than half of developing nations have access to multi-hazard early warning systems at all. That leaves billions of people blind to the next disaster.

Google Research has been working on this for years with their Flood Forecasting Initiative. Until now, they focused on riverine floods—the slow kind where rivers creep over their banks. That system already covers over 2 billion people across 150 countries. But urban flash floods are a different beast entirely.

The invisible flood problem

Riverine flood models are trained on stream gauges—physical sensors that measure water levels and flow. You feed historical gauge data into a machine learning model, and it learns to predict when a river will spill over. Works well, scales decently.

Flash floods don’t cooperate. They happen anywhere, often far from any gauge. In cities, the mess of impermeable concrete, overloaded drainage, and intense rainfall makes traditional physics-based modeling computationally insane at global scale. And without historical records of exactly where and when flash floods occurred, supervised ML models have nothing to learn from.

Google’s solution is clever: they built a training dataset from news reports. Using Gemini, they analyzed publicly available news articles mentioning floods, extracted confirmed locations and times, and aggregated those into a dataset of historical flash flood events. They call this approach Groundsource. It’s not perfect—news coverage is biased toward populated areas and dramatic events—but it’s a hell of a lot better than nothing.

Local precision vs. global reach

There are hyper-local flash flood warning systems out there. Florida has one. Barranquilla, Manila, Nakhon Si Thammarat, Mayaguez, Barcelona—they all have bespoke setups with physical sensors, radar data, and site-specific calibration. These work great for their specific locations. But they’re expensive to deploy, require constant tuning, and don’t scale.

Google’s approach is the opposite: train one global model on sparse but broad data, then run it everywhere. The trade-off is precision. Local systems might give you street-level warnings; Google’s model predicts risk at a neighborhood or district level. For a city planner or emergency manager, that’s still actionable. For an individual trying to decide whether to evacuate, it’s a start.

How the model works

The paper (linked below) goes into detail, but the gist is this: the model takes precipitation forecasts, terrain data, and urban drainage characteristics as inputs. It outputs a flood risk score for each location. The AI was trained on the Groundsource dataset, which labels past events from news reports. The model learns patterns—heavy rain over impervious surfaces in certain drainage basins—and generalizes to unseen locations.

They claim up to 24 hours of advance notice. That’s ambitious. Flash floods typically happen within six hours of heavy rain, so a 24-hour window gives people time to prepare. But the accuracy likely drops significantly beyond 12 hours. I’d want to see independent validation before trusting that number.

What this means

This is a genuine expansion of flood forecasting coverage. Riverine models miss urban flash floods entirely. Now there’s at least something. The Global South gets the biggest benefit, because those regions lack the budget for hyper-local sensor networks.

But let’s be real: this isn’t a silver bullet. The model’s accuracy depends on the quality and coverage of news data. Remote areas with limited media coverage will be blind spots. And 24-hour forecasts for a phenomenon that develops in hours is a stretch. Still, it’s better than nothing.

Google has rolled this out on Flood Hub. If you’re in a flood-prone city, check it out. The data is free. The code? Not so much—this is proprietary research. But the methodology is published, so others can build on it.

The bigger picture

Climate change is making extreme rainfall more common. Urbanization is paving over more ground, reducing natural drainage. The combination is a recipe for more flash floods. AI-driven forecasting won’t stop the water, but it might give people enough time to get to higher ground.

That’s worth something.