Shu et al. (2026) High-resolution urban flooding inundation forecasting through hydrodynamic interaction and multimodal deep learning

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-03-08
• Authors: Xinyi Shu, Zongxue Xu, Silong Zhang, Depeng Zuo, Chenlei Ye, Xudong Zhang, Lei Yu
• DOI: 10.1016/j.jhydrol.2026.135271

Research Groups

• College of Water Science, Beijing Normal University, Beijing, China
• Beijing Key Laboratory of Urban Hydrological Cycle and Sponge City Technology, Beijing, China
• Guangdong-Hong Kong Joint Laboratory for Water Security, Beijing Normal University, Zhuhai, China
• Center for Water Research, Advanced Institute of Natural Sciences, Beijing Normal University, Zhuhai, China
• School of National Safety and Emergency Management, Beijing Normal University, Zhuhai, China

Short Summary

This research proposes a multimodal deep learning model (SMDFN) tightly coupled with a hydrological-hydrodynamic model to improve high-resolution urban pluvial flooding inundation forecasting by capturing hydrodynamic interactions and enabling multimodal feature extraction. The model demonstrates superior performance, reducing RMSE by 13.8%, and offers a multi-region collaborative forecasting solution.

Objective

• To develop a multimodal deep learning model that effectively captures the interaction between hydrodynamic processes and inundation prediction, extracts multimodal data features, and accurately maps floods in discrete urban areas for high-resolution urban flooding inundation forecasting.

Study Configuration

• Spatial Scale: Urban areas (applied in Jinan, China), high-resolution inundation mapping.
• Temporal Scale: Rapid urban flood forecasting.

Methodology and Data

• Models used:
  • Tightly coupled 1D-2D hydrological-hydrodynamic model (for generating high-resolution inundation datasets).
  • Surface Multimodal Deep Flood Net (SMDFN) - an encoder-decoder deep learning architecture integrating residual connections, spatial attention, channel attention, and mask filtering.
• Data sources:
  • High-resolution inundation datasets generated from tightly coupled hydrodynamic simulations, capturing the interaction between 1D and 2D processes through manhole discharge.
  • Multimodal data features (processed by SMDFN).

Main Results

• Tightly coupling of 1D and 2D hydrodynamic processes provides a better representation of the interaction between surface and drainage flooding compared to loose coupling.
• Inundation depths in flood-prone areas under loose coupling ranged from 0.76 m to 1.40 m, which were generally higher than those obtained through tightly coupling.
• The SMDFN model effectively enables multimodal feature extraction and parallel prediction across multiple regions, demonstrating strong applicability under both disturbed and idealized rainfall scenarios.
• The incorporation of attention mechanisms and residual blocks significantly improved the model's performance, leading to a 13.8% reduction in Root Mean Square Error (RMSE).

Contributions

• Development of SMDFN, an innovative multimodal deep learning model that integrates hydrodynamic simulations and tightly coupled 1D-2D processes for high-resolution urban flood inundation forecasting.
• Enhanced representation and understanding of the interaction between hydrodynamic processes and inundation prediction through a tightly coupled modeling approach.
• Introduction of multimodal data feature extraction and accurate flood mapping in discrete urban areas using deep learning with attention mechanisms and residual blocks.
• Provides a novel multi-region collaborative forecasting solution for rapid urban flood forecasting and risk management.

Funding

• Not specified in the provided text.

Citation

@article{Shu2026Highresolution,
  author = {Shu, Xinyi and Xu, Zongxue and Zhang, Silong and Zuo, Depeng and Ye, Chenlei and Zhang, Xudong and Yu, Lei},
  title = {High-resolution urban flooding inundation forecasting through hydrodynamic interaction and multimodal deep learning},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135271},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135271}
}