Elsherbiny et al. (2026) Assessment of evapotranspiration across diverse arid settings in Saudi Arabia: A meta-learning analysis of multimodal satellite data (2003–2024)

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-04-09
• Authors: Osama Elsherbiny, Obaid Aldosari
• DOI: 10.1016/j.ejrh.2026.103420

Research Groups

• Agricultural Engineering Department, Faculty of Agriculture, Mansoura University, Mansoura, Egypt
• Interdisciplinary Research Center for Membranes and Water Security, King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia
• Department of Electrical Engineering, College of Engineering, Prince Sattam Bin Abdulaziz University, Wadi Ad-Dawasir, Saudi Arabia

Short Summary

This paper develops a novel meta-learning framework for accurate monthly actual evapotranspiration (AET) estimation using multimodal satellite data in arid Saudi Arabia, finding that a two-stage P-spline_P-spline architecture achieved superior predictive outcomes (R²=0.923, RMSE=5.337 mm).

Objective

• Implement model-driven feature selection techniques (Penalized Spline, Gradient Boosting Regressor, and Partial Least Squares Regression) for identifying optimal predictive variables from TerraClimate factors and surface-reflectance indices.
• Develop a meta-learning architecture coupling base learners and meta-models to advance AET prediction accuracy beyond standalone machine learning approaches.
• Identify the optimal meta-learning configuration for computing precise AET estimates that support strategic irrigation planning and water resource allocation in data-scarce environments.

Study Configuration

• Spatial Scale: Three arid sites in Saudi Arabia (Bisha, Buraydah, Duba) at a spatial resolution of approximately 4 km for TerraClimate data and 500 m for MODIS data.
• Temporal Scale: Monthly data spanning 22 years (2003–2024).

Methodology and Data

• Models used:
  • Machine Learning Algorithms: Penalized Spline (P-spline), Gradient Boosting Regressor (GBR), Partial Least Squares Regression (PLSR).
  • Meta-learning approach: Stacking (stacked generalization) framework utilizing P-spline, GBR, and PLSR interchangeably as base and meta-models.
• Data sources:
  • TerraClimate dataset: Monthly actual evapotranspiration (AET), Palmer Drought Severity Index (PDSI), solar radiation (SR), maximum temperature (Tmax), minimum temperature (Tmin), vapor pressure (VP), vapor pressure deficit (VPD), water deficit (WD), and wind speed (WS).
  • MODIS Aqua remote sensing data (Collection 1 km Daily, 500 m resolution): Normalized Difference Vegetation Index (NDVI), green/shortwave infrared ratio (Gr/SW), near-infrared/shortwave infrared ratio (NIR/SW), and green/near-infrared ratio (Gr/NIR).
  • TerraClimate data itself merges CRU Ts4.0 and Japanese 55-year Reanalysis (JRA55) with WorldClim climatological normals.

Main Results

• The proposed P-spline_P-spline two-stage meta-learning architecture, which uses higher-level variables generated by P-splines as inputs to a P-spline meta-learner, yielded superior predictive outcomes for AET.
• This optimal meta-learning model achieved a coefficient of determination (R²) of 0.923 and a Root Mean Squared Error (RMSE) of 5.337 mm on the unseen test set.
• This performance significantly surpassed the standalone P-spline model, which achieved an R² of 0.914 and an RMSE of 5.671 mm.
• The optimal P-spline_P-spline architecture integrated TerraClimate variables (VP, Tmin, WS, Tmax, WD) with surface-reflectance indices (Gr/NIR and SR).
• Hyperparameter optimization for the base P-spline learner was (n_knots, degree, alpha) = (11, 2, 0.0001), while the meta-learner adopted (7, 2, 0.1).

Contributions

• Introduced a novel meta-learning framework that integrates nonlinear feature transformations directly into a meta-learning layer, creating a self-optimizing pipeline for AET estimation in hyper-arid environments.
• Demonstrated the effective use of multimodal satellite data (TerraClimate factors and MODIS surface-reflectance indices) to eliminate reliance on scarce in-situ AET measurements.
• Validated model-driven feature selection techniques (P-spline, GBR, PLSR) for identifying optimal predictive variables from high-dimensional datasets.
• Significantly advanced AET prediction accuracy beyond standalone machine learning approaches, providing more robust and reliable estimates for strategic water management.
• The developed AET model (P-spline_P-spline) outperformed previous state-of-the-art approaches, such as the artificial neural network with an adaptive meta-model (R²=0.914, RMSE=6.115 mm).

Funding

• Prince Sattam bin Abdulaziz University (project number: PSAU/2025/01/33654).

Citation

@article{Elsherbiny2026Assessment,
  author = {Elsherbiny, Osama and Aldosari, Obaid},
  title = {Assessment of evapotranspiration across diverse arid settings in Saudi Arabia: A meta-learning analysis of multimodal satellite data (2003–2024)},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103420},
  url = {https://doi.org/10.1016/j.ejrh.2026.103420}
}