Chu et al. (2026) Physics-guided multi-task learning for urban flood prediction under drainage sedimentation scenarios

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-07-14
• Authors: Wenhao Chu, Yuqian Hu, Chaosen Jin, Chunxiao Zhang, Suchuang Di, Rongrong Li, Tianbao Wang, Heng Li
• DOI: 10.1016/j.jhydrol.2026.136036

Research Groups

• School of Artificial Intelligence, China University of Geosciences in Beijing
• Hebei Key Laboratory of Geospatial Digital Twin and Collaborative Optimization, China University of Geosciences Beijing
• Ningbo Hydrological Monitoring and Forecasting Center
• Observation and Research Station of Beijing Fangshan Comprehensive Exploration, Ministry of Natural Resources
• Department of Big Data Analysis and Application, Beijing Water Science and Technology Institute
• Institute of Space and Earth Information Science, The Chinese University of Hong Kong
• State Key Laboratory of Remote Sensing Science, Aerospace Information Research Institute, Chinese Academy of Sciences
• Key Laboratory of Virtual Geographic Environment (Ministry of Education of PRC), Nanjing Normal University

Short Summary

This study develops a physics-guided multi-task learning framework for urban flood prediction under drainage sedimentation scenarios, achieving higher accuracy in joint prediction of maximum inundation depth and velocity components.

Objective

• Investigate the effectiveness of physics-guided multi-task learning for urban flood prediction under complex conditions.

Study Configuration

• Spatial Scale: Urban scale (Beijing)
• Temporal Scale: Real-time to short-term forecasting

Methodology and Data

• Models used: Physics-based model, Multi-gate Mixture-of-Experts (MMOE) network, physics-guided MMOE (PG-MMOE)
• Data sources: High-fidelity two-dimensional flood fields generated using physically-based model, satellite and observation data

Main Results

• Multi-task learning outperforms single-task modeling in predicting maximum inundation depth and velocity components.
• Momentum-based constraints enhance depth–velocity coupling and overall physical consistency.
• PG-MMOE exhibits robust predictive accuracy and spatial generalization across unseen rainfall–sedimentation combinations.

Contributions

• This study demonstrates the effectiveness of embedding hydrodynamic priors into multi-task learning through task-aligned physical constraints for improving urban flood surrogate models.
• The proposed PG-MMOE framework provides a trustworthy surrogate for scenario-based flood risk assessment, offering quantitative support for maintenance decisions and emergency response under complex urban conditions.

Funding

• Hebei Key Laboratory of Geospatial Digital Twin and Collaborative Optimization (Grant No. 2022GZ01)
• National Natural Science Foundation of China (Grant No. 52179209)
• Beijing Municipal Science & Technology Commission (Grant No. Z2011000080200)

Citation

@article{Chu2026Physicsguided,
  author = {Chu, Wenhao and Hu, Yuqian and Jin, Chaosen and Zhang, Chunxiao and Di, Suchuang and Li, Rongrong and Wang, Tianbao and Li, Heng},
  title = {Physics-guided multi-task learning for urban flood prediction under drainage sedimentation scenarios},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136036},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136036}
}