Google unveiled a system that uses artificial intelligence (AI) to predict urban flash floods up to 24 hours in advance. However, due to regulatory issues, Korea was excluded from the service coverage.
On the 12th, local time, Google released a dataset, "Groundsource," based on flood data collected from 150 countries worldwide, along with a "urban flash flood" prediction model that uses it.
Using its AI model "Gemini," Google analyzed more than 5 million news articles and public records, structured the data with flood occurrence, date, and location information, and compiled about 2.6 million flood cases.
By reflecting this data in Google Maps to analyze the geographic boundaries where floods actually occurred, Google built data specialized for urban flash floods and developed a model that can predict floods up to 24 hours in advance.
When Google compared the model with the U.S. National Weather Service (NWS) flood alert system, the recall rate—the share of actual floods that received alerts in advance—was 32% for Google's model, higher than the NWS's 22%. In contrast, the precision—the share of alerts that were followed by an actual flood—was 26% for Google's model, lower than the NWS's 44%.
This means Google's model captures more potential floods but also produces more false alarms. Google said the balance between recall and precision can be improved by adjusting thresholds.
Google also emphasized that, unlike the NWS, which relies heavily on dense radar networks and ground observation equipment, its model achieved a similar level of performance using only satellite and weather data.
Google plans to provide these predictions for free through its disaster information platform Flood Hub and to release the Groundsource dataset as open source so research institutions and developers can use it.
Google said this "urban flash flood" model is the first case in which its large language model (LLM) has produced a result usable in real disaster response. It said it plans to apply the same approach to predicting other natural disasters, such as landslides and heat waves.