Large Language Models (LLMs) for transboundary water management
Managing flood risks in international river basins requires insight into the measures taken by upstream countries. In basins such as the Rhine, interventions in upstream countries, such as water storage, dyke reinforcement or weir management, have a direct impact on water safety further downstream, including in the Netherlands.
Exploring potential of technology
These measures are documented in a vast number of often multilingual governmental or institutional documents and plans. However, there is no comprehensive and up-to-date overview of what is happening across the entire basin, even though such an overview is essential for properly assessing impacts and supporting dialogue between countries.
Against this background, Deltares is exploring how hydrological models and data can be complemented with Artificial Intelligence (AI), especially Large Language Models (LLMs). In an experimental study, we are investigating whether this technology can help to more quickly identify and structure relevant measures in international river basins.
This application is still in an exploratory phase and is not yet used in operational decision-making, for example during extreme situations.
LLMs offer the possibility to rapidly search, process and analyse large volumes of policy documents. This can help generate initial insights into measures that may impact water safety and water availability. Work that would otherwise take weeks or even months if carried out manually.
What can LLMs contribute?
Policy documents, treaties and climate adaptation plans from different countries contain valuable information on measures and strategies related to water safety and availability. At the same time, this information is often spread across hundreds of pages, in multiple languages, and presents in largely unstructured formats.
LLMs can help identify patterns, categorise measures and extract relevant information, transforming unstructured information into structured outputs—tasks that are extremely time-consuming when done manually.
However, the verification, interpretation and practical application of the results still require human expertise and will be further refined in subsequent stages.
The importance of transboundary water management
Traditionally, the Netherlands focuses on what enters the country at its borders and how that water is distributed. However, many measures aimed at water safety at upstream, such as water storage and weir management, also influence water availability during dry periods.
Water does not stop at national borders. Policy decisions in other countries have direct consequences for both water safety and availability in the Netherlands.
By gaining better insight into these measures, it becomes possible to understand potential impacts earlier and to initiate timely dialogue between countries.
An example from the Rhine basin
A recent experiment with LLMs illustrates what this approach could yield: an overview of water safety related measures in the Rhine basin, compiled from thousands of documents.
By “measures”, we mean concrete actions that countries are considering to take for water management—for example, constructing higher dykes to reduce flooding, creating additional water storage in floodplains, or retaining water in reservoirs for energy production, flood protection or as a buffer.
Such upstream interventions can affect both the volume and timing of water reaching the Netherlands. By making this information clearer and available earlier, there is greater scope to assess impacts and to engage with other countries.
How does it work technically?
An agentic data mining system is developed to automate the extraction of measure information from massive collections of documents. The system utilizes cloud and LLMs technologies.
Through an agentic data mining process, the system has successfully identified tens of thousands of relevant data entries, cataloguing essential metadata and geographical coordinates. The outputs form the basis for further analysis and interpretation.
Technology and expertise
At this stage of the experiment, the focus is on exploring what automated analysis of policy documents can deliver. In subsequent phases, we will examine how these insights can be further interpreted and applied.
International collaboration and future perspective
Within Deltares, we are gradually investigating where the use of LLMs adds value, what the limitations are, and how the technology can be applied responsibly. This involves experimenting, learning and refining approaches before broader implementation.
Deltares is exploring how this approach can be further developed in the future, not only for the Rhine basin but also for other international river basins. This includes looking at collaboration and knowledge exchange with international partners.
