Li et al. (2025) Surrogate modeling for rapid estimation of spatially-resolved flood damage: Application to a coastal region

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-09
• Authors: Shicheng Li, Can Ding, James Yang
• DOI: 10.1016/j.jhydrol.2025.134763

Research Groups

• Department of Civil and Architectural Engineering, KTH Royal Institute of Technology, Stockholm, Sweden
• Department of Built Environment, Aalto University, Espoo, Finland
• R&D Hydraulic Laboratory, Vattenfall AB, Älvkarleby, Sweden

Short Summary

This study introduces BayFlood, a Bayesian-optimized machine learning surrogate model for rapid, accurate, and spatially resolved flood damage estimation using river discharge and tidal level inputs. The boosting-ensemble-driven BayFlood achieved the best performance (coefficient of determination = 0.92–0.98; root mean square error = 0.04–0.08) and reduced computational time by two orders of magnitude compared to hydraulic modeling.

Objective

• To develop a machine learning (ML)-based surrogate model for flood damage assessment (FDA) that provides rapid, accurate, and spatially resolved damage estimation, addressing the computational cost limitations of conventional hydraulic model-based FDA approaches and the lack of spatially resolved quantification in existing ML-based methods.

Study Configuration

• Spatial Scale: A coastal region of the Jianjiang River, Guangdong Province, China, extending approximately 15 km upstream from the estuary, with a computational domain of approximately 76 km².
• Temporal Scale: Focus on static peak flood damage assessment for 60 distinct flood events (30 river flow-dominant, 30 storm surge-dominant, and 4 compound flood scenarios), disregarding temporal flood evolution.

Methodology and Data

• Models used:
  • BayFlood: A Bayesian-optimized machine learning surrogate model.
  • ML Engines: Regression Tree (RT), Boosting Ensemble (BE), Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN).
  • Optimization: Bayesian Optimization (BO) using a Gaussian Process (GP) surrogate model and Expected Improvement (EI) acquisition function.
  • Hydraulic Model: TELEMAC (open-source 2D hydraulic simulation program) used to generate training/validation data.
  • Damage Function: Unit Damage Functions (UDFs) developed by Su et al. (2022) for Guangdong Province, quantifying damage rate based on inundation depth and land use.
• Data sources:
  • Flood Event Data: 60 flood events (river flow-dominant, storm surge-dominant, and compound flood scenarios) generated from 2D hydraulic simulations using TELEMAC.
  • Topographic Data: River channel topography measured in 2020 (1:5000 scale) from the Guangdong Provincial Design Institute of Water Conservancy and Electric Power.
  • Land Use Data: 2019 annual land cover dataset.
  • Hydrological Inputs: River discharge (Q) and tidal level (SL).
  • Hydraulic Model Validation Data: Water level measurements from September and October 2023 flood events for 1D model calibration/validation.

Main Results

• The Boosting Ensemble (BE)-driven BayFlood model consistently outperformed other ML engines (RT, LSTM, CNN) in accuracy and efficiency across all flood scenarios.
• For river flow-dominant (RFD) events, the BE-driven BayFlood achieved a coefficient of determination (CD) of 0.92 and a root mean square error (RMSE) of 0.08 for damage rate (DR) forecasting. 93% of predictions had errors below 0.10, and 97% below 0.20. For flood depth (FD) forecasting, it achieved CD = 0.95, RMSE = 0.53 m, and mean absolute error (MAE) = 0.25 m.
• For storm surge-dominant (SSD) events, the BE-driven BayFlood achieved a CD of 0.98 and an RMSE of 0.04 for DR predictions. 97% of forecasts showed errors below 0.10, and 99% below 0.20. For FD estimates, it achieved CD = 0.99, RMSE = 0.24 m, and MAE = 0.10 m.
• Under super-extreme compound flood (CF) conditions, the BE-driven BayFlood maintained stable predictive performance with a mean CD of 0.93 ± 0.01 and RMSE fluctuating by only 0.02.
• Monte Carlo uncertainty analysis (1000 runs, 5% noise level) revealed a mean damage rate uncertainty of 0.18 (18%) and a mean flood depth uncertainty of 1.3 m, confirming model robustness.
• BayFlood significantly reduced computational time by two orders of magnitude (factor of 66–330) compared to the hydraulic model (which took approximately 1 hour per event), generating damage results within minutes (< 5 minutes per event after training).

Contributions

• Development of the first ML-based surrogate model (BayFlood) for rapid and spatially resolved flood damage estimation, bridging the gap between high-accuracy hydraulic simulations and the need for real-time, high-resolution FDA.
• Achieves significant computational efficiency gains (two orders of magnitude faster) over traditional hydraulic models, enabling practical application for real-time emergency response, pre-disaster planning, and post-disaster recovery.
• Demonstrates high predictive accuracy and robustness across diverse flood scenarios (river flow-dominant, storm surge-dominant, and compound floods) and under input uncertainty.
• Provides actionable, spatially explicit damage maps, enhancing localized decision-making and risk management strategies.

Funding

• Swedish Center for Sustainable Hydropower (SVC)

Citation

@article{Li2025Surrogate,
  author = {Li, Shicheng and Ding, Can and Yang, James},
  title = {Surrogate modeling for rapid estimation of spatially-resolved flood damage: Application to a coastal region},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134763},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134763}
}