Lou et al. (2026) A highly generalizable data-driven model for spatiotemporal urban flood dynamics real-time forecasting based on coupled CNN and ConvLSTM

Identification

• Journal: Hydrology and earth system sciences
• Year: 2026
• Date: 2026-03-30
• Authors: Wangqi Lou, Xichao Gao, Joseph Hun Wei Lee, Jiahong Liu, Lirong Dong, Kai Gao
• DOI: 10.5194/hess-30-1625-2026

Research Groups

• State Key Laboratory of Water Cycle and Water Security, Beijing, China
• China Institute of Water Resources and Hydropower Research, Beijing, China
• Macau University of Science and Technology, Macau, China

Short Summary

This study proposes a novel data-driven model, coupling CNN and ConvLSTM, for real-time spatiotemporal urban flood inundation depth forecasting. The model effectively captures inundation dynamics and demonstrates robust spatial generalization with significantly higher computational efficiency compared to physics-based models.

Objective

• To propose a novel data-driven model to predict the spatiotemporal distribution dynamics of urban inundation depths by integrating ConvLSTM and CNN components.
• To enhance the model's generalization capability for urban flood forecasting by concurrently extracting information from temporal sequences and static geospatial features and employing a tiling approach during training.

Study Configuration

• Spatial Scale: A 4.06 km² flood-prone urban area in the western sector of the Macao Peninsula. The study area was divided into 100 × 100 grid cells, each with a 2 × 2 m spatial resolution.
• Temporal Scale: Hourly rainfall records (2000–2022) and designed rainfall events (6-hour duration) were used. The model uses rainfall and tide level from the three previous hours to the following 1 hour, and inundation depth from the last three hours, to predict inundation for the following hour.

Methodology and Data

• Models used:
  • Proposed Model: A coupled CNN-ConvLSTM deep learning model. The ConvLSTM component processes dynamic spatiotemporal data (rainfall, tide level, historical inundation depth), and the CNN component processes static geospatial features (DEM, ASP, CURV, SLOPE, RDEM, MSLOPE, MANHOLE, NETWORK, MASK).
  • Physics-based Hydrodynamic Model (for data generation and comparison): A coupled model simulating two-dimensional surface flows (Saint-Venant equations with finite-volume methods) and one-dimensional pipe drainage network flows (SWMM EXTRAN module), with interactions simulated using weir flow formulas and rainfall-runoff modeled by the Horton infiltration method.
• Data sources:
  • Geospatial Data: Digital Elevation Model (DEM) with 2 m resolution, drainage network, and building distribution information from the Macao Cartography and Cadastre Bureau.
  • Rainfall Data: Hourly historical observed data (Dapaotai station, 2000–2022) and designed rainfall (three typical patterns, four return periods: 10, 20, 50, 100 years, 6-hour duration).
  • Storm Tide Data: Designed tidal process lines of 5 warning levels.
  • Synthetic Compound Scenarios: Integration of rainfall events and tidal process lines, with 6-hour intervals randomly selected and combined.
  • Training/Testing Data for DL Model: Simulated inundation depths generated by the physics-based hydrodynamic model.

Main Results

• The proposed model effectively captured inundation processes at specific flood-prone sites (LHK, IHS, LPM), with Nash–Sutcliffe Efficiency (NSE) values exceeding 0.80 for most events, and Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) values below 0.20 m.
• The model demonstrated robust generalization performance across the study area, with mean NSE of 0.83, RMSE of 0.08 m, and MAE of 0.05 m. The Critical Success Index (CSI) for flood detection was 0.83.
• Maximum inundation depth predictions showed high accuracy, with the majority of regions exhibiting an absolute error of less than 0.10 m and a relative error under 5%.
• The data-driven model achieved a 4000 × speed advantage in inference time (4 s per prediction) compared to the physics-based model (16 200 s per simulation) on the specified hardware.
• K-fold cross-validation confirmed the model's robustness, showing consistent performance metrics (median NSE > 0.8, median RMSE < 0.15 m, median MAE < 0.10 m) across different training–validation splits.
• The hybrid CNN-ConvLSTM architecture outperformed CNN-only and ConvLSTM-only models, exhibiting superior and more stable predictive performance with improved lower-quartile NSE values and reduced median RMSE/MAE.

Contributions

• Development of a novel deep learning model that integrates ConvLSTM and CNN architectures in parallel to effectively capture both spatiotemporal dynamic information and static geospatial features for urban flood inundation forecasting.
• Introduction of a tiling strategy during model training, treating spatial sub-regions as independent samples, which significantly enhances the model's spatial generalization capability and reduces GPU memory requirements.
• Implementation of an auto-regressive prediction framework, allowing the model to use its current prediction as input for the next timestep, enabling real-time forecasting and potential data assimilation.
• Demonstration of a substantial increase in computational efficiency (4000x speedup) for real-time prediction compared to traditional physics-based hydrodynamic models, making it highly suitable for operational use.
• Comprehensive evaluation demonstrating robust performance in capturing both temporal dynamics at specific sites and spatiotemporal distribution across the entire urban area, even under complex hydrometeorological conditions.

Funding

• National Key Research and Development Program of China (grant no. 2022YFE0205200)
• National Natural Science Foundation of China (grant no. 52209044)

Citation

@article{Lou2026highly,
  author = {Lou, Wangqi and Gao, Xichao and Lee, Joseph Hun Wei and Liu, Jiahong and Dong, Lirong and Gao, Kai},
  title = {A highly generalizable data-driven model for spatiotemporal urban flood dynamics real-time forecasting based on coupled CNN and ConvLSTM},
  journal = {Hydrology and earth system sciences},
  year = {2026},
  doi = {10.5194/hess-30-1625-2026},
  url = {https://doi.org/10.5194/hess-30-1625-2026}
}