Google is using AI to predict flash floods in cities, and it’s about time

I’ve been watching Google’s flood forecasting work for a few years now, and their latest move is actually pretty interesting. They’re rolling out flash flood predictions for urban areas, using an AI method that scrapes news reports for training data. That’s not something you see every day.

According to the World Meteorological Organization, flash floods account for about 85% of flood-related deaths worldwide. They hit within six hours of heavy rain, turn city streets into rivers, and kill over 5,000 people annually. Early warning systems help, but there’s a massive gap between rich and poor countries. Less than half of developing nations have access to multi-hazard early warning systems. A 12-hour lead time can cut flash flood damage by 60%, but billions of people don’t get that luxury.

Google’s Flood Hub already covered riverine floods across 150 countries, serving over 2 billion people. But urban flash floods are a different beast. Rivers overflow slowly over days. Flash floods happen fast, anywhere, and often far from any stream gauge. Traditional machine learning models rely on historical water level data from physical sensors, but that data barely exists for flash floods in cities.

So Google did something clever: they used an AI-powered methodology called Groundsource to extract ground truth from unstructured data. Specifically, they fed Gemini — their large language model — publicly available news reports about floods. Gemini confirmed locations and times, and those entries were aggregated into a dataset of historical flash flood events. That dataset trained their new flash flood prediction model.

This approach has been tried before in research settings, but scaling it globally is new. The team had to balance local precision with global reach. Hyper-local systems exist in places like Florida, Colombia, the Philippines, and Spain, but they rely on expensive physical sensor networks. Google’s model uses publicly available rainfall data and terrain information, which is far cheaper to scale.

The model predicts flash flood risk up to 24 hours in advance. That’s higher than I expected for a first global rollout. The paper they published alongside this announcement goes into the technical details, but the key innovation is the training data pipeline. News reports are noisy and biased toward populated areas, but the team claims their method achieves high precision.

I have some reservations. News data tends to overrepresent dramatic events in wealthy regions, so there’s a risk of blind spots in poorer, less-reported areas. The team acknowledges this and says they’re working on incorporating satellite imagery and social media data. Also, 24-hour lead time for flash floods is ambitious — convective storms can be unpredictable even hours out. I’d like to see real-world validation stats beyond their paper.

Still, this is a solid step. The Flood Hub is free and publicly accessible. If you live in a flood-prone city, it’s worth checking. The climate isn’t getting kinder, and any advance warning is better than none.