Zheng et al. (2026) Physics-guided urban flood modeling: spatial heterogeneity and diagnostic attribution via CMoE-HDAF

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-07-26
• Authors: Shaoxu Zheng, Ziwu Fan, Xuan Huang, Guang Yang, Yipeng Liao, Yunfei Hong, Jingxiu Wu, Guoqing Liu, Chen Xie
• DOI: 10.1016/j.jhydrol.2026.136115

Research Groups

• Southern University of Science and Technology, China
• Nanjing Hydraulic Research Institute, China
• Key Laboratory of Taihu Basin Water Resources Management, Ministry of Water Resources, China

Short Summary

This study develops CMoE-HDAF, a physics-guided surrogate modeling framework, to enable rapid and spatially detailed urban flood forecasting and diagnostic attribution in high-density cities, achieving high accuracy and computational efficiency compared to conventional hydrodynamic models.

Objective

• To develop a computationally efficient, physically structured surrogate modeling framework (CMoE-HDAF) for rapid urban pluvial flood forecasting and diagnostic attribution, capable of representing complex hydrodynamic responses and spatial heterogeneity.

Study Configuration

• Spatial Scale: Longgang River Basin in Shenzhen, China, with simulations performed over 243,944 computational nodes.
• Temporal Scale: Analysis of 45 distinct rainfall scenarios; basin-wide inference time reduced to 32 seconds.

Methodology and Data

• Models used:
  • CMoE-HDAF (Cluster-based Mixture-of-Experts architecture with a Hierarchical Dominance Attribution Framework)
  • InfoWorks ICM (calibrated hydrodynamic model used to generate reference data)
• Data sources:
  • Simulation-based reference inundation data generated by a calibrated InfoWorks ICM model for 45 rainfall scenarios.
  • Static terrain, pipe-network, and hydrographic attributes used for basin partitioning.

Main Results

• The CMoE-HDAF surrogate model achieved a Nash-Sutcliffe Efficiency of 0.9761 when compared to hydrodynamic simulations.
• It significantly reduced basin-wide inference time to 32 seconds, enabling rapid scenario analysis.
• The CMoE component successfully partitioned the basin into 25 distinct hydrological regimes based on static terrain, pipe-network, and hydrographic attributes.
• The HDAF provided diagnostic attribution, characterizing how the relative influence of terrain, pipe-network, and hydrographic conditions varied across different rainfall scenarios and hydrological regimes.

Contributions

• Development of CMoE-HDAF, a novel physics-guided surrogate modeling framework that balances physical fidelity with computational efficiency for urban flood forecasting.
• Integration of a Cluster-based Mixture-of-Experts architecture to effectively capture spatial heterogeneity and regime-specific rainfall–inundation relationships.
• Introduction of a Hierarchical Dominance Attribution Framework for multi-scale, diagnostic interpretation of the controls (terrain, drainage, hydrography) on urban inundation response.
• Provides a calibrated-model emulator and diagnostic attribution tool for rapid analysis and spatially structured interpretation of heterogeneous urban inundation responses.

Funding

• Not explicitly stated in the provided text.

Citation

@article{Zheng2026Physicsguided,
  author = {Zheng, Shaoxu and Fan, Ziwu and Huang, Xuan and Yang, Guang and Liao, Yipeng and Hong, Yunfei and Wu, Jingxiu and Liu, Guoqing and Xie, Chen},
  title = {Physics-guided urban flood modeling: spatial heterogeneity and diagnostic attribution via CMoE-HDAF},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136115},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136115}
}