AI for urban hydrometeorology: insights into processes, model suitability, and challenges

Artificial Intelligence (AI) is no longer a novelty in hydrometeorology. From computer vision to rainfall nowcasting, AI-based models now routinely outperform traditional approaches in many benchmark comparisons. Yet, as these tools move closer to operational use --particularly in dense and vulnerable urban environments-- it is timely to step back and ask a more fundamental question: what is AI actually good at, and how should we use it wisely?

This talk reflects on recent advances in AI for urban hydrometeorology through three interconnected research regimes: smart environmental monitoring (“eyes on the water”), short-term rainfall nowcasting, and spatial–temporal rainfall reconstruction. Rather than promoting AI as a universal solution, the talk focuses on model suitability, uncertainty, and the alignment between data-driven methods and physical processes.

In environmental monitoring, modern deep-learning computer vision models have reached an impressive level of maturity. Tasks such as object detection, classification, and segmentation can now be performed reliably using images from fixed cameras, mobile devices, CCTVs, and citizen sensors, enabling scalable monitoring of urban rivers, flooding, and water quality indicators. At the same time, these applications reveal a recurring limitation: AI performs extremely well on what it has seen before, but struggles with rare events, or poorly defined labels --often the cases of greatest societal relevance.

In rainfall nowcasting, AI is often positioned as a disruptive replacement for traditional methods. This talk argues instead for a complementary view. While classical extrapolation efficiently handles storm motion, AI’s real strength lies in learning evolution: how rainfall structures grow, decay, and reorganise across spatial and temporal scales. Deep learning models excel at capturing multiscale spatial–temporal patterns that are difficult to encode explicitly, making them particularly valuable when combined with physically informed frameworks.

A central challenge across these applications is overconfidence. Can we teach AI to say “I don’t know”? Recent uncertainty-aware learning approaches demonstrate that AI models can be trained not only to make predictions, but also to indicate when they are operating outside familiar regimes --an essential requirement for trustworthy deployment.

Finally, the talk highlights the Point-to-Image (P2I) model to illustrate AI’s ability to learn spatial–temporal structure from extremely sparse data. By reconstructing realistic rainfall fields from limited point observations, P2I demonstrates that AI can infer coherent spatial patterns and temporal consistency even when traditional methods fail. This capability challenges long-held assumptions about data density requirements and opens new possibilities for urban hydrometeorology in data-limited environments.

Overall, this talk argues that the most effective use of AI in urban hydrometeorology arises not from replacing physical insight, but from combining process understanding with models that are well matched to the questions they are asked to answer.