Listen to this article:
How AI flood protection works
Floods are among the most destructive natural disasters in the world. In the Czech Republic alone, floods in 1997, 2002, 2010, 2013, and most recently in 2024 caused hundreds of billions of crowns in damage. The sooner hydrologists know that high water is coming, the more lives and property can be saved. This is exactly what Google is striving for with its new model.
The model is built on the LSTM (Long Short-Term Memory) architecture — a type of neural network that specializes in working with time series, i.e. data where order and development over time matter. In simple terms: the model "learns" from historical data how rivers behaved under various rainfall situations, and based on current meteorological forecasts, it then estimates what will happen in the coming days.
In the latest version, which currently powers the public service Google Flood Hub, it uses an architecture called ME‑LSTM. This can process weather forecasts from four different global sources simultaneously — including the European IFS model (ECMWF), satellite rainfall estimates from NASA (IMERG), the Graphcast model from Google DeepMind, and US ground measurements from NOAA. Each source has its own "translator" into the neural network's language, and their outputs are then intelligently combined in the model.
The Czech footprint: Why Google approached ČHMÚ
The Czech Hydrometeorological Institute is no unknown player in international hydrology. For years, it has been involved in developing standards within the World Meteorological Organization (WMO) and has extensive experience operating forecasting systems in the complex terrain of Central Europe, where mountain torrents alternate with lowland rivers. This very diversity makes the Czech Republic an ideal testing environment.
According to the official Google Research blog, the partnership with ČHMÚ was crucial in verifying that the AI model achieves comparable quality to traditional, locally calibrated hydrological models. The team led by RNDr. Jan Daňhelka, Ph.D. and Jakub Krejčí additionally prepared a so-called adapter — a software bridge that enables the model to be integrated into the Delft‑FEWS platform.
Delft‑FEWS is an operational system used daily for flood forecasting by most European hydrological services, including ČHMÚ. Thanks to the adapter created by Czech hands, meteorological institutes around the world can now immediately deploy the model into their regular operations without the need to build new infrastructure.
How many extra days do we gain?
According to a benchmark study published by Google Research this month, the new version of the model extends the reliable forecast horizon by six days for gauged basins (where historical flow data is available) and by one day for ungauged basins. In practice, this means that a flood warning can arrive with a lead time that gives crisis management teams substantially more time for evacuation and building flood barriers.
The model runs in Python using the PyTorch library and is trained on data from the open Caravan dataset — an extensive collection of hydrological measurements from around the world, managed by universities and research institutions. Each national service can add its own local data, thereby "fine-tuning" the model for the specific conditions of its territory. For the Czech Republic, this means, for example, more accurate forecasts for mountain tributaries, where traditional models often fail due to rapid runoff from torrential rainfall.
Open source that changes the rules of the game
Google has published the complete model code on GitHub under the Apache 2.0 license, one of the most liberal open-source licenses. This means that anyone can use, modify, and deploy the model — free of charge, without restrictions, even for commercial purposes.
Why is this important? Traditional hydrological models are extremely computationally demanding — they require supercomputers and tens of hours of computation for a single simulation. Google's AI model, on the other hand, can forecast flows for the entire planet in a matter of minutes on a standard server. This opens doors particularly for developing countries that cannot afford expensive infrastructure.
Dr. Hwirin Kim, Head of the Hydrological Modelling and Forecasting Division at the World Meteorological Organization, welcomed the initiative, saying: "Expanding open-source hydrological tools is essential for how societies manage water resources and respond to environmental challenges."
What this means for the Czech Republic and Europe
For Czech readers, there is concrete good news: ČHMÚ is already testing the model and preparing it for live operation. In the coming months, it could become part of daily forecasting workflows — alongside existing models, not as their replacement, but as a supplement that increases the overall reliability of the system.
In the European context, this is a textbook example of how a public institution (ČHMÚ) collaborates with a private technology giant to develop a tool that is subsequently available to the entire world. Moreover, the model is fully compliant with the European Floods Directive (2007/60/EC), which obliges member states to create flood hazard maps and flood risk management plans.
Czech hydrologists from ČHMÚ are also among the first in the world to have practically verified how complex it is to get an AI model into the routine operation of a state institution. Their experience now helps colleagues from other countries — from Slovakia and Poland to more distant regions of Africa and Asia.
Climate change is turning up the heat
According to data from the European Environment Agency, the number of flood events in Europe has doubled over the last 30 years. Extreme rainfall that used to occur once every hundred years now strikes every 20 to 30 years in some areas. In this reality, every extra day of warning is literally priceless.
Google's model is, of course, not a magic wand — its quality depends on the available data and on how well it is "fitted" to a specific basin. This is precisely why it matters that ČHMÚ not only tested the accuracy but directly prepared a complete integration chain. From raw data to a warning on a citizen's mobile phone.
How does the AI model differ from traditional forecasting systems?
Traditional hydrological models simulate physical processes in the landscape (infiltration, runoff, evaporation) step by step, which requires enormous computing power. The AI model, by contrast, finds patterns in historical data — it learns how specific combinations of rainfall, temperature, and soil saturation in the past led to certain flows. The computation is orders of magnitude faster, and the model can be run on a standard server.
Will I see forecasts from this model as an ordinary citizen?
Not directly — the model is intended for hydrologists and meteorologists, who incorporate its outputs into their forecasting systems. As a citizen, however, you will see the result in the form of more accurate and timely ČHMÚ warnings delivered to your mobile phone, the media, or the institute's website. On a global scale, the outputs can be viewed at Google Flood Hub.
Is the model available for Czech companies or local governments?
Yes — thanks to the Apache 2.0 open-source license, anyone can download and modify the code, including commercial entities. A municipality or region could theoretically train the model on their own data from local measuring stations and thus obtain more accurate forecasts for their specific area. In practice, however, deployment requires hydrological know‑how — this is not a "point-and-click" application for laypeople.
