Google Uses Gemini to Predict Flash Floods from Old News

Flash floods are notoriously hard to predict, especially in regions without dense weather‑sensing networks. Google Research revealed Groundsource, a new methodology that leans on its Gemini language model to mine more than five million old news reports and extract over 2.6 million geo‑tagged flood events.

The project repurposes archival reporting into a chronological dataset that researchers can use alongside live weather forecasts. Gemini was asked to identify and contextualize flood mentions in news coverage worldwide, then produce a structured timeline of events. That historical backbone helps a trained model estimate the likelihood of flash flooding for a roughly 20‑square‑kilometer area.

Google is showcasing the system via a Flood Hub platform that highlights urban risks across 150 countries and is sharing its data with local emergency response agencies. One early trial user reported faster response times to localized weather threats, suggesting the approach can provide practical benefits where traditional radar and sensor networks are scarce.

There are caveats: Groundsource’s spatial resolution is coarser than systems that ingest local radar, such as the US National Weather Service’s alerts, and Google hasn’t yet published comprehensive accuracy metrics. The model also depends on the historical record captured in news archives, which can vary by region and time.

Juliet Rothenberg, program manager on Google’s Resilience team, says the same technique could eventually be adapted for other hard‑to‑predict hazards like heat waves or mudslides. For now, Groundsource stands out as Google’s first use of a large language model explicitly to help create weather‑relevant datasets, adding to the company’s broader portfolio of forecasting tools.

It’s an interesting step: leveraging textual historical records to fill observational gaps. If accuracy benchmarks follow, this kind of hybrid approach might become a useful complement to conventional meteorological data, particularly in underserved areas.