source: google research: the next chapter in flood resilience: open sourcing google’s hydrology framework
level: technical
google research has open-sourced its hydrology modeling framework, the same system powering riverine flood forecasts on flood hub. the python package uses pytorch and includes long short-term memory network architectures. it takes in climate, soil, topography, and weather forecast data to predict daily river flow rates. the code comes with a training pipeline that uses the open-source caravan dataset, and users can add their own data to fine-tune models for local watersheds.
the release includes two model versions: the original from a 2024 benchmark and an upgraded v2 model now running live on flood hub. the v2 model uses a multi-encoder lstm architecture that processes diverse weather inputs into a single streamflow probability distribution. benchmarks show it extends reliable forecasts by six days in gauged basins and one day in ungauged basins compared to the previous version. the system integrates global weather products like graphcast, ifs, imerg, and cpc precipitation data.
the czech hydrometeorological institute partnered with google to validate the model and built an adapter for the delft-fews operational platform. this lets agencies worldwide plug the ai model into standard forecasting workflows. the open-source approach gives national meteorological services full control over their data and allows resource-limited regions to access advanced forecasting without costly infrastructure. the world meteorological organization supports the move, noting open tools help save lives and advance early warning systems globally.
why it matters: open-sourcing this framework lets local forecasters train ai models on their own data, improving flood warnings in underserved regions and reducing reliance on expensive traditional systems.
source: google research: the next chapter in flood resilience: open sourcing google’s hydrology framework