Suna et al. (2025) Stacked hybridization of deep learning model with grey wolf optimization for accurate and explainable reference evapotranspiration

Identification

• Journal: Journal of Atmospheric and Solar-Terrestrial Physics
• Year: 2025
• Date: 2025-10-09
• Authors: Truptimayee Suna, Bibhuti Bhusan Sahoo, D. Pawar, Nand Lal Kushwaha, Pradosh Kumar Paramaguru, P. S. Brahmanand, Himani Bisht
• DOI: 10.1016/j.jastp.2025.106655

Research Groups

• ICAR-Indian Agricultural Research Institute, Pusa, New Delhi, India
• School of Agriculture & Bio-Engineering, Centurion University, Paralakhemundi, Odisha, India
• Department of Soil and Water Engineering, Punjab Agricultural University, Ludhiana, Punjab, India
• ICAR-National Institute of Secondary Agriculture, Ranchi, Jharkhand, India

Short Summary

This study developed and evaluated a novel hybrid Deep Neural Network-Grey Wolf Optimization (DNN-GWO) model for accurate and explainable monthly reference evapotranspiration (ET0) forecasting in data-scarce regions. The DNN-GWO model significantly improved predictive accuracy, reducing RMSE by nearly 60% compared to the best-performing standalone deep learning model, offering a robust and interpretable solution for agricultural water management.

Objective

• To develop and evaluate a stacked hybridization of a deep learning model with Grey Wolf Optimization (DNN-GWO) for accurate, explainable, and robust forecasting of monthly reference evapotranspiration (ET0) in data-scarce regions.

Study Configuration

• Spatial Scale: Upper Ganga canal command region, Uttar Pradesh, India.
• Temporal Scale: Monthly forecasting of ET0.

Methodology and Data

• Models used:
  • Hybrid: Deep Neural Network-Grey Wolf Optimization (DNN-GWO)
  • Standalone: Random Forest (RF), Support Vector Machine (SVM), Deep Neural Network (DNN), Recurrent Neural Network (RNN), Deep Belief Network (DBN)
  • Explainability: SHapley Additive explanations (SHAP)
• Data sources: Meteorological observations (implied from context) providing:
  • Solar radiation (Rs)
  • Wind speed (U)
  • Maximum temperature (Tmax)
  • Minimum temperature (Tmin)
  • Relative humidity (RH)

Main Results

• Among standalone models, the Deep Neural Network (DNN) showed the best performance with R² = 0.958, Root Mean Square Error (RMSE) = 0.076 mm/day, and Nash-Sutcliffe Efficiency (NSE) = 0.954.
• The developed hybrid DNN-GWO model further improved predictive accuracy, achieving R² = 0.992, RMSE = 0.0317 mm/day, and NSE = 0.99.
• The DNN-GWO model reduced the RMSE by nearly 60 % compared to the best-performing standalone DNN model.
• SHapley Additive explanations (SHAP) analysis revealed that temperature (Tmax, Tmin) and solar radiation (Rs) were the most influential predictors of ET0.
• The hybrid model demonstrated robustness by providing stable predictions across different input scenarios, suitable for data-limited conditions.

Contributions

• Development of a novel and highly accurate hybrid Deep Neural Network-Grey Wolf Optimization (DNN-GWO) model for monthly ET0 forecasting.
• Significant improvement in ET0 prediction accuracy (approximately 60% reduction in RMSE) compared to state-of-the-art standalone deep learning models.
• Integration of explainability through SHAP analysis, providing critical insights into the influence of meteorological variables on ET0 predictions.
• Provision of a robust, accurate, and interpretable solution for agricultural water management and irrigation scheduling, particularly beneficial for data-constrained environments.

Funding

The provided paper text does not explicitly list specific funding projects, programs, or reference codes.

Citation

@article{Suna2025Stacked,
  author = {Suna, Truptimayee and Sahoo, Bibhuti Bhusan and Pawar, D. and Kushwaha, Nand Lal and Paramaguru, Pradosh Kumar and Brahmanand, P. S. and Bisht, Himani},
  title = {Stacked hybridization of deep learning model with grey wolf optimization for accurate and explainable reference evapotranspiration},
  journal = {Journal of Atmospheric and Solar-Terrestrial Physics},
  year = {2025},
  doi = {10.1016/j.jastp.2025.106655},
  url = {https://doi.org/10.1016/j.jastp.2025.106655}
}