Google uses AI and old news reports to predict flash floods.

Flash floods represent one of the most dangerous natural disasters, causing over 5,000 deaths annually. The difficulty in predicting them arises from the lack of consistent data, as these phenomena develop rapidly and affect small geographic areas.

Google has addressed this issue by using its language model, Gemini, to analyze approximately five million news articles, identifying 2.6 million flood events. This information has been transformed into a structured dataset, called Groundsource, which has been used to train a forecasting model based on LSTM neural networks. The forecasting system is integrated into Google's Flood Hub platform, providing data on flood risks in 150 countries and assisting emergency response organizations. Although it has limitations, such as relatively low resolution, the method is useful in regions without advanced meteorological infrastructure. Experts believe that this approach could also be extended to other hard-to-measure environmental phenomena, demonstrating how AI can transform unstructured human knowledge into useful data for predicting natural disasters.