Google uses Gemini to predict urban flash floods 24 hours in advance

Training an AI to predict rapid natural disasters usually hits the same wall: the world never kept good records.

Flash floods are among the deadliest weather events on Earth, killing over 5,000 people each year. Because they hit and vanish within hours, satellite passes often miss them or get blocked by heavy clouds. Traditional disaster databases like the UN and European Commission's GDACS track only around 10,000 major global events, which is basically a drop in the bucket for deep learning.

Google Research dropped Groundsource, a clever framework that skips physical sensors and turns to the morning paper instead. By having Gemini read decades of public news reports, Google built a massive historical dataset out of everyday journalism.

Reading 5 million articles with Gemini

Google fed roughly 5 million global news articles across 80 languages through Cloud Translation and engineered prompts for Gemini. The model extracted flood timings and locations from unstructured prose, which were then pinned down to precise geographic boundaries using Google Maps.

That process generated a dataset of 2.6 million historical flood records across more than 150 countries dating back to the year 2000.

In human verification checks, Google found that 60% of the extracted events were spot-on in both timing and location. When expanded to practical utility—like identifying the correct administrative district or hitting within a day of the flood's peak—accuracy reached 82%.

Google then used this ground truth to train an LSTM neural network that ingests global weather forecasts. The resulting model can predict urban flash floods up to 24 hours in advance, and those alerts are now live inside Google's Flood Hub platform.

Bridging the sensor gap

The model is not without rough edges. Its spatial resolution sits at a relatively coarse 20 square kilometers, and it doesn't ingest local precipitation radar feeds, meaning it won't replace high-precision setups like the U.S. National Weather Service anytime soon.

However, cutting-edge radars are luxury infrastructure that most of the world cannot afford. António José Beleza, an emergency response official with the Southern African Development Community who tested the tool, noted that it materially helped his teams mobilize faster.

As Marshall Moutenot, CEO of Upstream Tech, put it, geophysics is plagued by data scarcity where real-world validation data is brutally hard to find. Extracting that ground truth from public media is an exceptionally creative end-run around hardware limits.

Google confirmed that it plans to apply the exact same text-to-data pipeline to other hard-to-track phenomena like landslides and extreme heat waves.

Why it matters

We think Groundsource is one of the most practical deployments of large language models to date, proving LLMs can construct scientific-grade ground truth out of messy human archives rather than just writing summaries.

For communities across the Global South that lack high-density radar grids, getting a 24-hour head start on a flash flood via Flood Hub is a massive leap forward. Even a coarse 20-kilometer warning is infinitely better than waiting until the water reaches the front door.

The clear risk, however, lies in media reporting bias. News outlets naturally cover populated, connected cities far more aggressively than remote rural villages. If your AI learns historical baselines exclusively from journalism, it risks baking geographic neglect straight into its predictive weights.

Still, Google played to its unique strengths here. By tying Gemini directly into Google Maps and its existing Flood Hub infrastructure, it found a fast, low-cost way to ship climate resilience tools that traditional meteorological agencies simply could not build on paper records alone.