Google Research publishes open-source hydrological AI for flood forecasting

Google Research announced the open-sourcing of its hydrological machine learning model for predicting river floods. The developed Python package uses the PyTorch library and reproduces the architecture of Flood Hub models, allowing national meteorological and hydrological services to train and tune forecasting systems on their data. The code includes two model versions: the original version from a 2024 paper and an updated one currently used in global Flood Hub forecasting. Tests showed that the new model extends reliable forecasts by six days for calibrated basins and by one day for uncalibrated ones compared to the previous version. Input data include geographic features (climate, soils, terrain, vegetation) and meteorological forecasts (precipitation, temperature), while outputs are daily river discharge. The development involved the Czech Hydrometeorological Institute (CHMI), which integrated the model into the Delft-FEWS platform. The repository is available on GitHub under the Apache 2.0 license, including training notebooks and video tutorials. Google emphasizes that the open model will allow scientists and agencies to add their own data, including traditional local knowledge, and adapt AI to local catchments without needing expensive traditional infrastructure.