Flash floods are among the deadliest weather events worldwide, killing more than 5,000 people each year. They’re also notoriously hard to predict because they happen fast, are highly localized, and often occur where continuous monitoring infrastructure is limited. Google believes it has found a creative workaround: using old news reports and AI to build a new kind of forecasting data set.
Why flash floods are so difficult to forecast
Traditional meteorological forecasting depends on dense, continuous observations like river gauges, radars, and long-term records. Flash floods, however, are short-lived and localized — so many events never get captured by official sensors. That data gap makes it difficult to train machine-learning models that require reliable, ground-truth records.
How Google built Groundsource from the news
Google researchers turned to a nontraditional data source: the news. Using Gemini, Google’s large language model, they processed roughly 5 million news articles and extracted details about about 2.6 million flood events. The result is a geo-tagged time series called Groundsource, a structured data set that turns written reports into usable, quantifiable records of past flooding.
According to Google Research product manager Gila Loike, this is the first time Google has used language models in this specific way to generate a large-scale geophysical data set.
Training a forecasting model
With Groundsource as a real-world baseline, the team trained a Long Short-Term Memory (LSTM) neural network to combine global weather forecasts and the historical reports. The model outputs the probability of flash-flood risk in a given area, turning qualitative news reports into a quantitative tool for prediction.
Deployment and real-world use
Google has integrated the forecasts into its Flood Hub platform, highlighting urban risk areas across 150 countries and sharing data with emergency response agencies. Emergency officials who trialed the system say it helped speed response efforts — for example, António José Beleza from the Southern African Development Community reported that the forecasts improved his organization’s ability to react to flooding.
Strengths and limitations
- Strengths ✅
- Creative use of public, historical text to fill observational gaps.
- Globally scalable approach suited to regions without extensive weather infrastructure.
- Groundsource is publicly shared, enabling broader research and collaboration.
- Limitations ⚠️
- Relatively low spatial resolution — risk is identified across ~20 km² grid cells, so it’s not pinpoint-precise.
- Less precise than systems that incorporate local radar and real-time sensors (for example, US National Weather Service alerts).
- Dependent on the quantity and quality of historical reporting in different regions.
What experts are saying
Juliet Rothenberg, a program manager on Google’s Resilience team, explained the broader aim: “Because we’re aggregating millions of reports, the Groundsource data set actually helps rebalance the map. It enables us to extrapolate to other regions where there isn’t as much information.”
Marshall Moutenot, CEO of Upstream Tech and co-founder of dynamical.org, praised the approach as a creative fix to a common problem: “Data scarcity is one of the most difficult challenges in geophysics. Simultaneously, there’s too much Earth data, and then when you want to evaluate against truth, there’s not enough. This was a really creative approach to get that data.”
Why this matters
Groundsource demonstrates how language models can transform qualitative historical records into quantitative data that improves forecasting where traditional sensor networks are sparse. The technique could be extended to other ephemeral hazards — such as heat waves, landslides, or mudslides — to create new data products that support early warning systems and disaster response worldwide.
Bottom line
By reading the archives of global news with AI, Google has converted human reports into a large-scale, geotagged flood data set and used it to power predictive models. The result isn’t a replacement for high-resolution, radar-based alerts, but it is a practical, scalable tool to reduce blind spots in global flood forecasting — especially in regions that need it most. 🌍🌧️