News

Setting the scene

Gaining a greater understanding of how artificial intelligence (AI) and machine learning could support better prediction, communication, and management of flood risks across Scotland over the coming years is a key priority. Following a project request from SEPA, The Centre of Expertise for Waters (CREW) funded this project (DelugeAI) to conduct a rapid literature review, stakeholder engagement, and deliver a feasibility assessment and 5-year roadmap for future AI integration into flood forecasting. This workshop, which is a key milestone in the project, brings together area experts to share knowledge, identify opportunities and discuss challenges. The project is led by Chris White’s team from the University of Strathclyde.

From Continental Models to Local Insights: Advancing Hydrological Prediction in Europe with AI

The workshop kicked off with a keynote by Ilias Pechlivanidis, Lead Scientist at the Swedish Meteorological and Hydrological Institute and chair of the HEPEX initiative. Ilias discussed research on using AI to improve large-scale hydrological predictions across Europe. By combining traditional models with machine learning techniques, his team has developed hybrid approaches that better simulate river flows, including extreme events. Ilias shared his hope that in the future “no-one should be surprised by a flood”.

Understanding the use of AI for flood forecasting

Expert insights on AI applications – Lightning talks

After a short break, the second part of the workshop started with a series of three five-minute lightning talks. The first was given by Michael Butts, Lead Hydrologist at the Danish Meteorological Institute, who discussed how AI can support, but not replace, human decision making in flood warning systems. This was a sentiment echoed at multiple points throughout the workshop. By combining remote sensing, river data, and model outputs, AI could help experts interpret complex information.

Massimiliano Zappa, senior scientist at the Swiss Federal Institute for Forest, Snow and Landscape Research (WSL) discussed his team’s work applying AI to drought forecasting. Although he expressed that AI showed promise in improving forecasts, he also stressed the importance of cautious use, transparency, and linking predictions to real-world impacts.

The final lightning talk in this session was given by Jonathan Frame, Assistant Professor at The University of Alabama. Jonathan highlighted how AI, particularly convolutional networks, can be an amazing tool for things like large-scale flood forecasting. By combining physical models with AI, researchers can reconstruct flood events even when cloud cover obscures satellite images.

Initial findings from a systematic review of flood forecasting and AI

In this section of the workshop, Vicky Martí from the DelugeAI project team presented some preliminary findings from their literature review exploring how AI is being used in flood forecasting. Her work identified growing global interest, especially since 2018, in applying AI across different stages of flood forecasting, from monitoring to emergency response. Whilst most academic research is concentrated in countries like China, the US, India, the UK and Canada there are also many practical applications in existence. These range from Google's global Flood Hub to small-scale projects such as FloodCast, FloodAI in the UK and FloodWaive from Germany. Vicky explained that AI is most commonly used to complement traditional models and support decision-making, rather than to fully replace existing forecasting systems.

Open discussion: Expert insights on AI applications

The final activity in the second workshop session was an open discussion chaired by Robert Atkinson from the project team. The expert voices in the room agreed that AI is promising, especially for speeding up forecasts, supporting decision-making, and processing large volumes of data. However, there were also concerns about over reliance on AI, lack of transparency, and the ongoing need for human expertise. Speakers stressed the importance of keeping a “human in the loop” to interpret forecasts and make informed decisions. As such, it will be important to consider having clear best practices and sensitive integration to allow trust, accountability, and effective communication. This is especially important for high stakes scenarios such as flooding.

Looking forward

Expert insights on AI applications: Lightning talks

The third and final session of the day commenced with three more five-minute lightning talks. Maria Luisa Taccari, researcher at the European Centre for Medium-Range Weather Forecasts, discussed how deep learning offers faster, more flexible flood forecasting than traditional models, which can be expensive and find the physics too complex to simulate. Deep learning can learn from real-world data and performs well in challenging areas like Spain, which have seen extensive anthropogenic changes. Although results are promising, questions remain around predicting multiple variables, dealing with varying timescales, and applying models across different climates.

The second lightning talk was given by Steven Ramsdale, Chief Forecaster at the UK Met Office. Steven talked about using machine learning to improve severe weather warnings. By focusing on three key atmospheric parameters, their model can help identify when and where warnings may be needed. This approach keeps human expertise central to final decisions with support to highlight priority areas being provided by an AI tool.

Jan Verkade, Senior Hydrometeorologist at Deltares, gave the last lightning talk of the day and highlighted the growing pressure on flood agencies to adopt AI, but warned of a lack of clear best practices. His team is working with the forecasting community to define how AI can be responsibly integrated—considering ethics, costs, legal frameworks, and the need for human oversight.

Open discussion: Identifying realistic applications of AI in flood forecasting

Following the five-minute talks, Doug Bertram from the project team opened the floor up for another discussion session in which the expert attendees discussed the future role of AI in operational settings. Although AI offers promise in speeding up forecasts and handling large data volumes, workshop participants stressed the importance of maintaining human expertise, especially for decision-making and public trust. Key themes included the need for better data, ethical safeguards, explainability and workforce upskilling. The need for communication tools, interdisciplinary collaboration, and improved knowledge on the environmental costs of AI were also highlighted as areas for further focus. Ultimately, whilst AI can enhance forecasting, experts agreed it must complement, not replace, human judgement.

Closing Remarks

The workshop concluded with reflections from the DelugeAI team lead, Chris White, and SEPA requester, Michael Cranston (Lead Flood Forecaster). They expressed the value of bringing together experts from different fields to explore the role of AI in flood forecasting. Despite the rapid pace of AI development, participants agreed that there’s no single solution, highlighting the need for continued collaboration, shared learning, and practical experimentation. Key themes included the importance of keeping humans “in the loop” for decision-making, building trust in AI tools, and ensuring that any implementation is operationally relevant. There was strong support for focusing on low risk, quick win applications in the near future, whilst acknowledging that more strategic, long-term planning is a challenge. The need for skills development, ethical safeguards, and open communication with end users were also recurring points throughout the workshop. As the project moves forward, the team will compile insights into a report and roadmap.

CREW would like to thank the research team (University of Strathclyde) and the Project Steering Group (SEPA, Environment Agency and Scottish Government) for their dedication and support to the project and in addressing this important issue.