Rahman et al. (2025) Water level forecasting in coastal cities using a hybrid deep learning approach

Identification

• Journal: The Science of The Total Environment
• Year: 2025
• Date: 2025-10-15
• Authors: Abdur Rahman, M. Hafidz Omar, Tahir Mahmood, Nasir Abbas, Muhammad Riaz, Naeem Ramzan
• DOI: 10.1016/j.scitotenv.2025.180709

Research Groups

• Department of Mathematics and Statistics, King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia
• School of Computing, Engineering and Physical Sciences, University of the West of Scotland, Paisley, United Kingdom

Short Summary

This study introduces a novel hybrid deep learning model, CNN-Transformer-SKANs, for accurate and real-time hourly water level forecasting in coastal cities. The model achieved superior accuracy (NSE ≈0.99, RMSE < 0.03 m) and robustness in Venice, Italy, even under data-scarce and extreme event scenarios, providing an effective early warning tool.

Objective

• To develop a highly accurate and robust hybrid deep learning model (CNN-Transformer-SKANs) for hourly water level forecasting in coastal cities, specifically Venice, to overcome challenges posed by nonlinear hydrological processes and limitations in long-term historical data availability.

Study Configuration

• Spatial Scale: Venice, Italy (northern Adriatic Sea, Venice Lagoon, latitudes 45°26′ to 45°28′ North and longitudes 12°19′ to 12°21′ East). The study focuses on the core observation zone of Venice.
• Temporal Scale: Two years of high-resolution hourly data (January 1, 2022, to January 1, 2024), comprising 17,521 hourly samples. The model performs hourly forecasting using a 12-hour lookback window.

Methodology and Data

• Models used:
  • Proposed: CNN-Transformer-SKANs (hybrid model combining Convolutional Neural Networks, Transformer layers, and Swallow Kolmogorov Arnold Networks).
  • Baseline models for comparison: LSTM, CNN-LSTM, Transformer, LSTM-KANs, CNN-LSTM-KANs, Transformer-GKANs, CNN-Transformer-GKANs, Transformer-SKANs.
  • Statistical method for synthetic data generation: Generalized Extreme Value (GEV) distribution.
• Data sources:
  • ISMAR-CNR Observatory (Institute of Marine Science, National Research Council)
  • Tidal forecasting and reporting center (Comune di Venezia)
  • Punta Della Salute (Canal Grande) station (used as a zero-water level benchmark).
  • Input variables: Tide level, wind speed (m/s), atmospheric pressure (hPa), air temperature (°C), cumulative rainfall (mm), solar radiation (W/m²), relative humidity (%), water temperature (°C), and historical water level (1 hour ago, m).
  • Output feature: Water level (m).

Main Results

• The proposed CNN-Transformer-SKANs model consistently achieved the highest predictive accuracy, with Nash–Sutcliffe Efficiency (NSE) values of approximately 0.99 and Root Mean Square Error (RMSE) values below 0.03 m.
• It significantly outperformed baseline models (e.g., LSTM, CNN-LSTM, Transformer, and their KANs-enhanced variants), which showed RMSE values between 0.04 m and 0.07 m and NSE values between 0.90 and 0.97.
• The model demonstrated high robustness, maintaining accuracy (RMSE as low as 0.02 m to 0.03 m, NSE up to 0.99) even with reduced training data (e.g., 50% of the dataset) and under synthetic extreme value simulations based on the GEV distribution.
• Seasonal analysis confirmed the model's consistent performance across all four meteorological seasons, effectively capturing both seasonal stability and irregular variability in water levels.
• The CNN-Transformer-SKANs model successfully generated stable and physically plausible 365-day extended forecasts without requiring future external meteorological variables, by recursively feeding back predicted values.

Contributions

• Development of a novel hybrid deep learning architecture (CNN-Transformer-SKANs) specifically designed for accurate and robust hourly coastal flood forecasting.
• Introduction of superior temporal modeling by replacing the LSTM component with a Transformer, enabling more efficient capture of long-range dependencies and parallel processing.
• Enhancement of nonlinear learning and model efficiency through the integration of Swallow Kolmogorov Arnold Networks (SKANs), which utilize adaptive radial basis functions, sparsity, and pruning.
• Improvement in model interpretability due to the combination of Transformer attention mechanisms and SKANs’ visualizable basis functions, leading to faster convergence.
• Demonstrated high accuracy and robustness in challenging scenarios, including data scarcity and extreme hydrological events, validated using real-world data from Venice and synthetic extreme value simulations.
• Provides a powerful, AI-driven early warning system tool for climate-sensitive coastal regions facing increasing flood risks.

Funding

Not explicitly mentioned in the provided paper text.

Citation

@article{Rahman2025Water,
  author = {Rahman, Abdur and Omar, M. Hafidz and Mahmood, Tahir and Abbas, Nasir and Riaz, Muhammad and Ramzan, Naeem},
  title = {Water level forecasting in coastal cities using a hybrid deep learning approach},
  journal = {The Science of The Total Environment},
  year = {2025},
  doi = {10.1016/j.scitotenv.2025.180709},
  url = {https://doi.org/10.1016/j.scitotenv.2025.180709}
}