Google open-sources the AI hydrology model behind Flood Hub

What did Google just release for flood forecasting, and who can use it?

Google Research has published its hydrology modeling framework as open-source software on GitHub. The release targets National Meteorological and Hydrological Services (NMHSs), researchers, and operational forecasters who want to train AI-based river flood prediction models using the same underlying architecture that drives Google's Flood Hub product.

What the framework actually is

The package is written in Python and depends on PyTorch. It implements model architectures based on Long Short-Term Memory (LSTM) networks, which take in geographical features covering climate, soils, topography, and land cover, together with meteorological forecasts for rainfall, temperature, and other conditions, to predict daily river flow rates.

A built-in training pipeline lets users work directly with the Caravan dataset, an open-source collection of historical river discharge data. Agencies can supplement that baseline with their own local observations and specialized datasets, giving regional forecasters a path to incorporate knowledge that a globally trained model would otherwise miss.

Why agencies keeping data control matters

One explicit design goal is data sovereignty. Because the framework runs on an agency's own infrastructure, NMHSs and other government bodies can train and operate models without sending sensitive or proprietary hydrological records to a third party. Google's announcement frames this as empowering local experts rather than replacing them.

Google says it tested the framework in partnership with the Czech Hydrometeorological Institute (CHMI) before publishing it publicly. That collaboration appears to be the primary validation Google cites for the framework's readiness for operational use.

Two different user groups, one codebase

The release is pitched at two distinct audiences with different goals.

First, hydrological researchers can add and test new model architectures, training data sources, and methodological approaches on top of what Google has already built. Second, operational forecasters responsible for issuing actionable flood warnings for specific regions can fine-tune the models with local gauge data and domain knowledge.

The distinction matters because a research team's objective (improving forecast accuracy at global scale) differs considerably from a forecaster's objective (issuing reliable, timely warnings for a particular watershed).

Relationship to Google Earth AI

Google is folding this hydrology framework into what it calls the Google Earth AI family of geospatial models and datasets. That positions the release alongside other Earth observation and environmental modeling work coming out of Google Research, though the announcement does not enumerate specific sibling projects in that family.

What is not included

The source documentation does not describe licensing terms, minimum hardware requirements, or whether pre-trained model weights are bundled with the repository. Prospective users will need to consult the GitHub repository directly for those details.

Component	Details
Language	Python
ML backend	PyTorch (open source)
Model architecture	LSTM-based
Input features	Climate, soils, topography, land cover, meteorological forecasts
Output	Daily river flow rate
Default training data	Caravan dataset (open source)
Local data support	Yes, agencies can add proprietary datasets
Validated with	Czech Hydrometeorological Institute (CHMI)
Parent product	Google Flood Hub

The practical implication is that a hydrology team with existing gauge records and computing resources can now replicate, and potentially improve upon, the model architecture behind one of the more widely deployed AI flood forecasting systems currently in operation.