Overview
- Google used its Gemini model to analyze about 5 million news articles and extract roughly 2.6 million geo-tagged flood events, creating the Groundsource dataset.
- An LSTM-based model trained on this dataset forecasts urban flash-flood risk up to 24 hours in advance, now displayed in Flood Hub across approximately 150 countries.
- Google published the research and dataset as an open benchmark and added the work to its Google Earth AI portfolio for broader scientific reuse.
- Emergency agencies are receiving the forecasts, and a Southern African Development Community official reported faster response in trials, though public accuracy metrics are still pending.
- The system currently flags risk over areas of about 20 square kilometers and does not integrate local radar, limiting precision compared with some national alert services.