Xu et al. (2026) Rapid Prediction of Compound Flood Based on Hydrological-Hydrodynamic Model and Convolution Neural Network

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-02-25
• Authors: Kui Xu, Yong Tian, Lingling Bin, Chengguang Lai, Weichao Yang
• DOI: 10.1007/s11269-026-04546-x

Research Groups

• State Key Laboratory of Hydraulic Engineering Intelligent Construction and Operation, Tianjin University, China
• School of Civil Engineering, Tianjin University, China
• Faculty of Geography, Tianjin Normal University, China
• School of Civil Engineering and Transportation, South China University of Technology, China
• Tianjin Key Laboratory of Soft Soil Characteristics & Engineering Environment, Tianjin Chengjian University, China

Short Summary

This study proposes a hybrid approach coupling a hydrological-hydrodynamic model (PCSWMM) with a Convolutional Neural Network (CNN) for rapid and accurate prediction of compound flood processes in coastal urban areas, demonstrating a significant increase in computational efficiency while maintaining high prediction accuracy for flood depths.

Objective

• To propose a hybrid approach for rapid and accurate prediction of compound flood processes in coastal urban areas by coupling a hydrological-hydrodynamic model with a Convolutional Neural Network (CNN), specifically focusing on the combined effects of rainfall and tidal levels.

Study Configuration

• Spatial Scale: Haidian Island, Haikou City, Hainan Province, China, with a total area of approximately 13.8 km². The urban flood simulation model discretizes the surface into 23,580 two-dimensional hexagonal grids with a resolution of 25 meters.
• Temporal Scale: Flood simulation and prediction duration of 24 hours, with a temporal resolution of 1 hour for data collection and prediction.

Methodology and Data

• Models used:
  • Hydrological-hydrodynamic model: Personal Computer Storm Water Management Model (PCSWMM)
  • Deep learning model: 1D Convolutional Neural Network (CNN)
• Data sources:
  • Elevation data: Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences.
  • Drainage network and river data: Haikou Municipal Water Bureau.
  • Building distributions: Extracted from satellite remote sensing images (Haikou Municipal Water Bureau) using ENVI software.
  • Historical flood events data: Rainfall processes, tide level processes, and flood depths at observation points for model calibration and validation.
  • Design rainfall and tide levels: Generated using the Generalized Extreme Value (GEV) function based on a typical event (Typhoon Rammasun, July 18, 2014).
  • CNN training data: Simulated flood depth data from PCSWMM under 36 rainfall-tide level scenarios, along with rainfall time series, tide level time series, and average flood depth from the preceding time step.

Main Results

• The CNN model reliably reflects flood depth variation trends, achieving an average absolute error (MAE) of 0.012 meters, a root mean square error (RMSE) of 0.021 meters, and a Pearson correlation coefficient (PCC) of 0.977 across test scenarios.
• The CNN model significantly reduces computational time, operating approximately 450 times faster than PCSWMM simulations (11 seconds for CNN vs. 1.4 hours for PCSWMM for a compound flood prediction).
• The CNN model effectively discriminates between flooded and non-flooded grids (using a 0.1 meter flood depth threshold), with Precision, Recall, and F-score all exceeding 0.9.
• The spatial distribution and temporal evolution of flood patterns predicted by the CNN show a high degree of consistency with PCSWMM simulations.

Contributions

• Proposes a novel hybrid approach for the rapid prediction of dynamic compound flood processes in coastal cities, explicitly considering the combined effects of rainfall and tidal levels, which was largely overlooked in previous studies focusing on rainfall-only events or maximum flood depths.
• Demonstrates a substantial improvement in computational efficiency (approximately 450 times faster) compared to traditional physics-based models, making the approach suitable for real-time flood emergency management and decision-making.
• Provides a robust and reliable tool for predicting the spatiotemporal evolution of compound floods, offering valuable insights for disaster prevention and mitigation in coastal urban areas.

Funding

• National Key Research and Development Program of China (2023YFC3209404)
• National Natural Science Foundation of China (52379019, 42477501)

Citation

@article{Xu2026Rapid,
  author = {Xu, Kui and Tian, Yong and Bin, Lingling and Lai, Chengguang and Yang, Weichao},
  title = {Rapid Prediction of Compound Flood Based on Hydrological-Hydrodynamic Model and Convolution Neural Network},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-026-04546-x},
  url = {https://doi.org/10.1007/s11269-026-04546-x}
}