Google Develops New Flash Flood Prediction Method Using Old News and Gemini Large Model

Flash floods are one of the world's deadliest meteorological disasters, but they are hard to predict due to their sudden occurrence and small scale. Traditional meteorological data lacks sufficient flash flood observation records, making it difficult to train AI models.

Google's team used the Gemini large model to extract 2.6 million flood events from 5 million news articles, building the Groundsource dataset with spatiotemporal markers. The LSTM model trained on this dataset can input weather forecast data and output flash flood probabilities, and it has been providing risk alerts for 150 countries on Flood Hub.

However, the system has limitations: its spatial resolution is about 20 square kilometers, it does not include real-time radar data, and its accuracy is not as good as the U.S. National Weather Service's system.

In the future, this technology is expected to expand to predict heatwaves, mudslides and other disasters, helping data-scarce regions. Industry insiders believe this is an important step in using innovative data collection to advance AI weather forecasting.