Artificial Intelligence in Water Research – Opportunities, Potentials, and Challenges

Report on the AI workshop organized by BfG and WSA

The two-day event "Artificial Intelligence in Water Research – Opportunities, Potentials, and Challenges," organized by the Federal Institute of Hydrology (Bundesanstalt für Gewässerkunde) and the Water Science Alliance, clearly demonstrated how deeply methods of Artificial Intelligence (AI) and Machine Learning (ML) are now embedded in various fields of water science. On the first day, four sessions titled "Water Science," "Water Budget," "Image and Video Processing," and "Generative AI" presented current research work. The presentations highlighted potentials in data validation, hydrological forecasting, automated analysis of remote sensing and video data, and knowledge management. At the same time, requirements for data quality, validation, and process understanding, as well as the importance of hybrid modeling approaches, were emphasized. The second day, held in a World Café format, facilitated discussions on five thematic areas: "Quality Control," "Hydrology," "Ecology," "Water Quality," and "Generative AI." Topics discussed included opportunities such as pollutant detection, improved prediction of runoff processes in small catchments, and automated monitoring of invasive species using remote sensing, as well as challenges like data quality, explainable AI, data privacy, and liability. The workshop highlighted that AI offers significant potential in water research but also requires careful consideration of methodological, technical, and ethical aspects.

Artificial Intelligence (AI) and Machine Learning (ML) are now integral to daily life and shape numerous societal and economic sectors. Applications such as personalized recommendation systems, search engines, automated translations, and digital navigation services illustrate their importance. With the emergence of large language models and generative AI, there is a shift from largely unconscious to increasingly reflective use of AI-based technologies. This developmental step brings significant potentials as well as new technical, ethical, and regulatory challenges.

In water research, AI and ML methods open up opportunities to qualitatively improve established procedures, conduct data- and labor-intensive analyses more efficiently, and explore new research and application fields. At the same time, fundamental questions arise regarding the quality and availability of data and the relationship between data-driven models and traditionally used deterministic approaches. Generative AI tools are also gaining importance as they offer various support functions in scientific and administrative workflows, such as automated text processing, document analysis, or programming assistance. Despite concrete efficiency gains, challenges persist, particularly regarding technical uncertainties, security aspects, and potential biases in automated decision-making processes. Against the backdrop of the federal government's national AI strategy, which emphasizes the societal benefits of AI-based technologies, the AI/ML workshop, provided a platform for interdisciplinary exchange on technical and societal issues related to the responsible use of AI in water research. This summary reflects the key contents of the scientific contributions and the resulting discussions, offering central impulses for further discourse.

The first day of the event was held in a hybrid format and included four thematic sessions focusing on "Water Science," "Water Budget," "Image and Video Processing," and "Generative AI." Over 70 participants attended in person, with approximately the same number joining online. The technical presentations and discussions showed that data-driven procedures not only open up new possibilities for analysis and evaluation but can also significantly extend established hydrological methods. At the same time, it was emphasized that high demands on data quality, process understanding, and validation limit the use of modern AI procedures and therefore represent a central area of development. Results from an international online survey conducted as part of the UNESCO FRIEND-Water initiative, as well as literature meta-analyses, indicate that ML methods are increasingly being used in water science, primarily in research and less in operational practice. Obstacles include limited trust within the expert community and practical challenges such as restricted data availability, lack of methodological expertise, and insufficient guidelines. The applications presented in time series analysis and forecasting demonstrated the potential of these methods for data validation, such as in the operation of water management facilities and for hydrological forecasting, where deep learning models can serve as powerful enhancements to existing forecasting systems. Their use improves, in particular, medium-term forecasts and enables the integration of extensive and heterogeneous data sources into operational forecasts. The discussion on the role of data-driven models compared to established deterministic approaches made it clear that future developments should aim for hybrid modeling approaches that combine the strengths of machine learning with physical process understanding, thereby strengthening the trust of the expert community.

Another thematic focus was image and video processing, where AI methods are gaining practical relevance. Application examples such as the automated optical determination of water levels, the classification of vegetation types, macroplastics, and oil based on remote sensing data, as well as the "Smart Fish Counter" for the automated evaluation of biological video monitoring data, showed that computer vision-based approaches can significantly reduce manual effort while opening up new possibilities for analysis. The topic of Generative AI concluded the first day by focusing on knowledge management, document creation, and quality assurance. The presentation of the development of the AI assistant by the Federal Highway Research Institute (BASt) illustrated both the potential of large language models for structuring extensive text collections and the associated challenges regarding data sovereignty, data privacy, confidentiality, and quality assurance to minimize AI hallucinations. Additionally, the use of large language models in the quality analysis of hydrological data was discussed.

To systematically gather the perspectives of the participants, the World Café on the second day provided an opportunity to discuss the opportunities, potentials, and risks of AI applications in different fields. Approximately 50 participants could choose three out of five thematic areas: "Quality Control of Measurement Data," "Hydrology," "Ecology," "Water Quality," and "Generative AI." Subsequently, the results of each group were presented in the plenary session to provide an overview of the discussion content from all thematic areas. The quality of the data or training data was identified as a central challenge not only at the "Quality Control of Measurement Data" table but also at the "Hydrology" and "Water Quality" tables. It is a crucial factor for the accuracy and reliability of AI models. Ensuring data consistency and completeness is essential for the successful use of AI-based applications. Mentioned examples included the need for accurate and consistent measurement data in hydrology or for monitoring water quality in the field of water quality. From this perspective, AI applications offer various opportunities: they can detect errors in human-labeled data, homogenize input data, and enable faster and more comprehensive segmentation and classification of data. At the thematic tables, the potential of AI to automate processes in such a way that resources can be used more efficiently and processes accelerated was widely emphasized. Furthermore, the use of AI enables the analysis of large, multidimensional datasets that would be difficult to evaluate with conventional methods. Opportunities and potentials identified at individual thematic tables included the identification of new pollutants in water quality assessment and improved prediction of runoff processes in small catchments in hydrology. Another opportunity highlighted was the ability to make the limitations of AI models visible, thereby creating new focal points in quality assurance. Additionally, changes in lakes caused by invasive species can be automatically monitored using remote sensing. Furthermore, the use of generative AI offers potential for targeted communication.

A widely discussed challenge across all thematic areas was the lack of explainability of AI results. A key aspect in the use of AI applications is Explainable AI (XAI). This creates transparency and makes the results understandable for users. It is particularly important for the traceability of AI results to transparently explain the decisions and predictions. Otherwise, the use of AI resembles a black-box process. At the same time, XAI also offers the opportunity to increase the acceptance and trustworthiness of AI, as it is possible to understand how AI models function and why they make certain decisions. The design of XAI should be chosen in such a way that it supports the user without overwhelming them with too much information. Other challenges discussed at individual thematic tables included the need for computing resources and appropriate infrastructure in the quality control of measurement data. In hydrology, it was noted that the training data often does not represent the overall population. In the XAI field, the trade-off between performance and explainability was mentioned in water quality assessment, data privacy in the application of passive acoustic methods in ecological monitoring, and liability and responsibility issues in the use of generative AI.

AI applications offer significant opportunities for water research. At the same time, several challenges were identified, both individually within specific disciplines and generally related to the use of AI applications. The discussions made it clear that the successful use of AI requires a combination of transparent methods, interdisciplinary collaboration, and continuous research, and these aspects need to be developed more intensively in the future. In the concluding summary, it was emphasized that AI and ML in water research represent a growing, interdisciplinary field of innovation. The event showed that data-driven methods are already being productively used in many areas, while fundamental questions about their integration into existing modeling and monitoring approaches remain open. The need for common standards, open and quality-assured datasets, and closer collaboration between authorities, scientific institutions, and technology developers was particularly evident. The announcement of the workshop's continuation in two years underscored the intention to continuously accompany the dynamic development in the AI field and to strengthen long-term collaboration in the discipline.