Google’s new flash flood AI gives cities a fighting chance

Flash floods are brutal. The World Meteorological Organization says they account for about 85% of flood-related deaths globally. They turn city streets into rivers in under six hours, kill more than 5,000 people every year, and they’re notoriously hard to predict. Early warning systems help—a 12-hour heads-up can cut damage by 60%—but most of the Global South doesn’t have them. That’s a lot of preventable deaths.

Google Research has been working on flood forecasting for years, but mostly for riverine floods—the kind where a river slowly creeps over its banks. Those models are trained on physical stream gauges that measure water levels. They work, covering over 2 billion people across 150 countries. But flash floods are a different beast. They happen fast, anywhere, often far from a gauge. In cities, the mix of intense rain, concrete, and drainage systems makes traditional physics-based modeling computationally insane at global scale. And without historical records of where flash floods actually happened, you can’t train a supervised ML model to predict them.

So Google did something clever: they built a dataset from scratch using news reports. They call it Groundsource. They fed Gemini—their LLM—publicly available news articles about floods, and had it extract precise locations and timestamps of events. Then they aggregated those into a training set for a new flash flood model. It’s a neat hack: instead of waiting for governments to install gauges everywhere, they just read the news. The paper is out, and they’ve rolled out the predictions on Flood Hub for urban areas. Up to 24 hours advance notice.

The scaling challenge is real. Hyper-local systems exist in places like Florida, Barranquilla, Manila, and Barcelona. They rely on networks of physical sensors—radar, precipitation gauges, flow meters—and they’re accurate. But they’re expensive and require site-specific calibration. You can’t drop that in every city in the developing world. Google’s approach is different: it’s purely AI-driven, trained on the news dataset, and it runs globally without hardware. The trade-off is precision—it won’t tell you exactly which street will flood, but it’ll tell you the risk zone. For a city that had nothing before, that’s a huge step.

I’ve been watching Google’s Flood Forecasting Initiative for a few years. The riverine stuff is solid. But flash floods always felt like the harder problem. The fact that they’re using LLMs to extract training data from news is interesting—it’s a practical application of generative AI that actually solves a real data scarcity problem, not just another chatbot wrapper. The model itself is a spatiotemporal transformer, which is standard for this kind of work, but the training data pipeline is the real innovation.

Is it perfect? No. The dataset is only as good as the news coverage it’s drawn from. If a flood happens in a remote area without media reporting, it won’t be in the training set. And the model’s urban focus means it might miss flash floods in rural or suburban areas where drainage patterns are different. But for cities in the Global South that currently have zero warning capability, this is a massive improvement. 24 hours might not sound like much, but when you’re dealing with a six-hour onset event, that’s a full day to move people and resources.

Google’s also releasing the Groundsource dataset publicly. That’s a nice move—it lets other researchers build on it. The paper describes the methodology in detail, so it’s reproducible. I’d like to see independent validation, but the initial results look promising. If this scales, it could save thousands of lives a year. That’s not bad for a project that started with reading the news.