Work

Large-scale urban flood modeling and zero-shot high-resolution generalization with LarNO

Year: 2026

Type: article

Abstract: In urban areas, real-time early warning systems help mitigate the severe impacts of pluvial flooding. Although such systems have undergone significant recent development with the use of deep learning methods, urban-scale high-resolution modeling processes remain fundamentally constrained by the massive amount of GPU memory that is required for training neural networks. To overcome this bottleneck, we present LarNO, a memory-efficient, discretization-invariant neural operator that learns continuous-space hydrodynamic mappings to predict the spatiotemporal distributions of water depths based on dynamic rainfall and static topographic and drainage conditions. This approach enables zero-shot generalization to high resolutions when trained solely on low-resolution data. Additionally, existing autoregressive neural operators compress complex dynamics into a single physical variable, leading to an information bottleneck when modeling complex nonlinear spatiotemporal dependencies. To resolve this issue, we embed spatiotemporal feature extractors into a neural operator to implement latent autoregression. Theoretically, we prove that our approach universally approximates continuous operators and enables the generalization of zero-shot super-resolution. An empirical case study of a highly urbanized region of nearly 100 km 2 in Shenzhen, China, subjected to observed rainfall events with spatiotemporal heterogeneity, demonstrates that LarNO achieves large-scale, high-resolution (5 m, 5 min), long-duration (6 h), real-time spatiotemporal urban flood forecasting at mm-level depth accuracy, reducing prediction errors by more than half relative to state-of-the-art neural operators. Simultaneously, it delivers two-orders-of-magnitude faster inference speedup than a traditional hydrodynamic model. Notably, these improvements incur negligible parameter overhead (only 0.024% more than the baseline neural operator), and LarNO uniquely, among the compared neural operators, supports multi-GPU distributed training and TensorRT-accelerated inference. In addition, the results of an ablation study validate the effectiveness of latent autoregression. Moreover, LarNO supports few-shot transfers to unseen catchments via fine-tuning. Furthermore, comprehensive parameter sensitivity analyses demonstrate the robustness and effectiveness of LarNO. Our work provides a novel framework delivering real-time early warnings about large-scale urban flood events and achieving refined management schemes. The source code is available at https://github.com/holmescao/LarNO .

Source: Journal of Hydrology

DOI: 10.1016/j.jhydrol.2026.135686

Authors: Xiaoyan Cao, Yao Yao, Zhi Wang, Zhangxinyue Zhao, Alistair G.L. Borthwick, Huapeng Qin