Zeroualı et al. (2025) Next-generation runoff prediction: Merging RFE, SHAP insights, and satellite data with innovative deep learning techniques

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2025
• Date: 2025-10-30
• Authors: Bilel Zeroualı, Abdullah Alodah, Zaki Abda, Faten Nahas, Nadjem Bailek, Richarde Marques da Silva, Youssef M. Youssef
• DOI: 10.1016/j.ejrh.2025.102870

Research Groups

• Laboratory of Architecture, Cities and Environment, Faculty of Civil Engineering and Architecture, Department of Hydraulic, Hassiba Benbouali University of Chlef, Algeria
• Department of Civil and Environmental Engineering, Federal University of Paraíba, Brazil
• Stokes School of Marine and Environmental Sciences, University of South Alabama, USA
• Department of Civil Engineering, College of Engineering, Qassim University, Saudi Arabia
• Department of Civil Engineering and Hydraulic, Laboratory LGCH, 8 May 1945 University, Guelma, Algeria
• Department of Geography, College of Humanities and Social Sciences, King Saud University, Saudi Arabia
• Laboratory of Mathematics Modeling and Applications, Department of Mathematics and Computer Science, Faculty of Sciences and Technology, Ahmed Draia University of Adrar, Algeria
• Department of Geosciences, Federal University of Paraíba, Brazil
• Geological and Geophysical Engineering Department, Faculty of Petroleum and Mining Engineering, Suez University, Egypt

Short Summary

This study developed and evaluated three advanced hybrid deep learning models (RFE-GRU-BiLSTM, RFE-GRU-CNN, and RFE-CNN-GRU-BiLSTM) for daily runoff prediction in north-central Algeria. The research found that these models, incorporating Recursive Feature Elimination and SHAP analysis, significantly improve predictive accuracy and interpretability, with lagged discharge identified as the primary driver of runoff.

Objective

• To investigate the effectiveness of three advanced hybrid deep learning models (RFE-GRU-BiLSTM, RFE-GRU-CNN, and RFE-CNN-GRU-BiLSTM) in predicting daily runoff in the Oued Ouahrane Ras basin.

Study Configuration

• Spatial Scale: Oued Ouahrane Ras watershed in the Cheliff Basin, north-central Algeria, covering an area of 270.7 km².
• Temporal Scale: Daily runoff prediction, using data spanning from 1998 to 2012.

Methodology and Data

• Models used:
  • Recursive Feature Elimination (RFE) for dimensionality reduction.
  • SHapley Additive exPlanations (SHAP) for post hoc feature importance analysis.
  • RFE-Gated Recurrent Unit–Bidirectional Long Short-Term Memory (RFE-GRU-BiLSTM).
  • RFE-Gated Recurrent Unit–Convolutional Neural Network (RFE-GRU-CNN).
  • RFE-Convolutional Neural Network–GRU–BiLSTM (RFE-CNN-GRU-BiLSTM).
• Data sources:
  • Satellite-based precipitation data from the Tropical Rainfall Measuring Mission (TRMM).
  • Ground-based in-situ daily hydrological observations (rainfall and runoff) collected from the National Agency of Water Resources (ANRH) of Blida station.

Main Results

• The RFE-GRU-CNN model achieved the best performance on the TRMM dataset, yielding a minimum Root Mean Square Error (RMSE) of 3.61 m³/s (Model M8).
• The RFE-CNN-GRU-BiLSTM model produced superior outcomes for the in-situ dataset, with an RMSE of 3.815 m³/s (Model M7).
• SHAP analysis consistently identified the lagged discharge input (Qt−1, Qt−2) as the key predictor across models, highlighting the catchment’s short-term memory and rapid runoff persistence controlled by active storage components. Precipitation inputs (both in-situ and TRMM) had a secondary influence.
• The implementation of RFE introduced a modest and consistent computational overhead, increasing average training time by approximately 43.37 seconds, which remained within acceptable limits. Training durations ranged from 1106.54 to 2425.52 seconds for TRMM data and 1224.71 to 1955.10 seconds for in-situ data.
• TRMM-based models generally exhibited slightly lower RMSE and Mean Absolute Error (MAE) values, and higher Correlation Coefficient (R) values, indicating higher precision and consistency in daily runoff estimation.

Contributions

• Development and evaluation of innovative hybrid deep learning models (RFE-GRU-BiLSTM, RFE-GRU-CNN, RFE-CNN-GRU-BiLSTM) for daily runoff prediction, demonstrating enhanced accuracy and interpretability.
• Integration of Recursive Feature Elimination (RFE) for effective dimensionality reduction and SHapley Additive exPlanations (SHAP) for robust post hoc feature importance analysis in hydrological forecasting.
• Comparative analysis of satellite-based (TRMM) and in-situ data, providing insights into their respective strengths and limitations in optimizing model performance for runoff prediction.
• Provided new hydrological insights for the Oued Ouahrane Ras basin, identifying lagged discharge as the dominant predictor and characterizing the catchment as a memory-driven system with rapid response dynamics.
• Demonstrated that the proposed hybrid models achieve competitive or superior performance compared to traditional and some modern approaches in runoff prediction.

Funding

• Deanship of Graduate Studies and Scientific Research at Qassim University (QU-APC-2025).

Citation

@article{Zeroualı2025Nextgeneration,
  author = {Zeroualı, Bilel and Santos, Celso Augusto Guimarães and Alodah, Abdullah and Abda, Zaki and Nahas, Faten and Bailek, Nadjem and Silva, Richarde Marques da and Youssef, Youssef M.},
  title = {Next-generation runoff prediction: Merging RFE, SHAP insights, and satellite data with innovative deep learning techniques},
  journal = {Journal of Hydrology Regional Studies},
  year = {2025},
  doi = {10.1016/j.ejrh.2025.102870},
  url = {https://doi.org/10.1016/j.ejrh.2025.102870}
}