Flash floods are notoriously difficult to predict, but Google has proposed a completely new solution, through...A flood forecasting tool called "Groundsource"By using the Gemini language model to extract relevant data from past news reports, Google can predict the likelihood of sudden floods. This is not only the first time Google has used language models for this type of forecasting, but it also brings hope for early disaster preparedness to regions lacking meteorological infrastructure.
Transforming millions of news articles into meteorological big data
In the past, training AI models capable of predicting regional disasters was often limited by a lack of sufficient local historical data. Groundsource, however, uses the Gemini model to sift through up to 500 million news reports from around the world, separating out reports related to floods.
Next, the system converts this textual data into geographically tagged records, chronologically listing each event. Researchers then use this data to train a new model capable of receiving current weather forecasts and leveraging historical data from Groundsource to estimate the likelihood of sudden flooding in specific regions.
Currently, Google has already...Its "Flood Hub" platformGoogle has identified flood risks in urban areas of 150 countries. It is also providing this forecast data to emergency response agencies in these regions to assist local personnel in disaster preparedness.
While there is currently no specific information on the accuracy of the prediction model (Google says it will take time to verify), one test user has stated that the system has indeed helped his organization respond more quickly to localized weather disasters.
Flash flood prediction models need historical data and model training that often doesn't exist. Our solution: Groundsource, a new AI-powered methodology that uses Gemini to transform 5M+ global reports into a precise dataset of 2.6M+ flood events.
This provides a massive,…
— Google Research (@GoogleResearch) March 12, 2026
Solutions specifically designed for areas lacking meteorological infrastructure
Of course, as an emerging technology, Groundsource's forecasting model still has some inherent limitations. For example, it can currently only identify risk areas covering a 20-square-kilometer area. And because Google's model does not integrate local radar data, it cannot incorporate real-time rainfall tracking data, so its accuracy is still not as good as the US National Weather Service's flood warning system.
However, the platform was originally designed to function in areas that typically lack weather sensing infrastructure.
Juliet Rothenberg, project manager for Google's Resilience team, pointed out that this technology could be further applied to predicting other unpredictable natural phenomena such as heat waves and landslides.Further explanationJuliet Rothenberg stated, "We are compiling millions of reports, which allows us to infer the situation in other regions where information is lacking."
While Groundsource marks Google's first application of language models to weather forecasting, it's not the first time Google has relied on artificial intelligence in this area. Previously, Google DeepMind's WeatherNext 2 forecasting model had already proven its extremely high accuracy.
