Google Open-Sources Hydrological AI for Flood Forecasting

Google Research released an open-source hydrology modeling framework for AI-based flood forecasting on GitHub under an Apache 2.0 license. The framework uses PyTorch and LSTM networks, allowing researchers and forecasters to train models with the same architecture used in Google's Flood Hub. It takes input data on climate, soils, topography, land cover, and meteorological forecasts to predict daily river flow rates. The repository includes two model versions: the original from 2024 and an upgraded version currently powering real-time global flood forecasting on Flood Hub, extending the reliable predictive horizon by six days in gauged basins and one day in ungauged basins. The Czech Hydrometeorological Institute (CHMI) integrated the model into the Delft-FEWS platform, demonstrating operational feasibility. The package also includes a training pipeline using historical river data from the open-source Caravan dataset, enabling agencies to add local data and fine-tune models for their watersheds. Google aims to democratize advanced flood forecasting, particularly for resource-constrained regions, and support the World Meteorological Organization's early warning systems.