Belarbi et al. (2026) Machine learning estimation of reference evapotranspiration using MODIS-Derived and limited ground variables across Moroccan agro-climatic zones

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2026
• Date: 2026-03-23
• Authors: Zaid Belarbi, Yacine El Younoussi
• DOI: 10.1007/s00704-026-06193-z

Research Groups

• Information System and Software Engineering Laboratory, Abdelmalek Essaadi University, Tetouan, Morocco

Short Summary

This study evaluates machine learning models for estimating daily reference evapotranspiration (ETo) in data-scarce Moroccan agro-climatic zones, demonstrating that MODIS remote sensing and limited ground variables can achieve high accuracy and support water management, despite challenges in inter-regional transferability.

Objective

• To apply and evaluate multiple machine-learning algorithms combined with remote-sensing variables for ETo estimation in a data-scarce, water-scarce, and agriculturally dependent country.
• To analyze the contribution of key meteorological and remote-sensing predictors, emphasizing the potential of remote-sensing variables to substitute temperature and solar radiation as primary drivers of ETo.
• To assess model transferability across stations within and between contrasting agro-climatic regions, simulating real-world conditions where site-specific data are incomplete or unavailable.

Study Configuration

• Spatial Scale: Two agro-climatic zones in Morocco: a semi-arid region (Doukkala) and a sub-humid coastal region (Loukkos). Three meteorological stations were used: Khemis Zemamra and Khemis Mettouh (Doukkala), and M’rissa (Loukkos).
• Temporal Scale: Daily observations over multi-year periods: Khemis Zemamra (September 2000 – November 2019), Khemis Mettouh (September 1996 – September 2012), and M’rissa (January 2008 – July 2024).

Methodology and Data

• Models used: Support Vector Regression (SVR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), CatBoost, and Artificial Neural Network (ANN).
• Data sources:
  • Ground-based: Daily meteorological records (maximum/minimum air temperature, maximum/minimum relative humidity, solar radiation, wind speed, rainfall) and reference evapotranspiration (ETo) calculated using the FAO-56 Penman–Monteith method, provided by the Regional Offices for Agricultural Development of Doukkala (ORMVAD) and Loukkos (ORMVAL).
  • Remote sensing: Moderate Resolution Imaging Spectroradiometer (MODIS) products:
    • MOD11A1: Daily Land Surface Temperature (LSTDay, LSTNight) at 1 km spatial resolution.
    • MCD43A3: Broadband albedo (Black-Sky Albedo, White-Sky Albedo) at 500 m resolution.
    • MOD13Q1: Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI) at 250 m resolution.

Main Results

• Ensemble-based models (LightGBM, XGBoost, CatBoost) consistently achieved the most stable and accurate performance across all scenarios.
• Scenario A (excluding ground-measured solar radiation, Rs): Models maintained high predictive accuracy, with mean coefficients of determination (R²) exceeding 0.96 and Root Mean Squared Error (RMSE) of approximately 0.36 mm day⁻¹. LST_Day was identified as the most influential predictor, followed by maximum temperature, Julian day, wind speed, and minimum relative humidity.
• Scenario B (excluding ground-measured temperature): Ensemble models maintained excellent accuracy (R² ≥ 0.95), with solar radiation and LST_Day as the leading predictors.
• Scenario C (excluding both Rs and temperature): Predictive skill decreased but remained strong (R² ranging from approximately 0.85 to 0.93), with LST_Day, Julian day, minimum relative humidity, and wind speed as dominant predictors.
• Transferability Assessment:
  • Within the same agro-climatic region (Doukkala to Doukkala): Models demonstrated strong predictive skill (R² > 0.90, RMSE ≈ 0.47–0.49 mm day⁻¹).
  • Across contrasting agro-climatic regions (Doukkala to Loukkos): Performance declined markedly (R² ≈ 0.63, RMSE ≈ 1.10 mm day⁻¹), attributed to climatic heterogeneity and domain shifts in predictor distributions (e.g., wind speed, humidity, LST, vegetation indices).
• Statistical Significance: Wilcoxon signed-rank tests confirmed that the best-performing gradient-boosting models (XGBoost, LightGBM) significantly outperformed SVR, ANN, and CatBoost in the Rs-free scenario.

Contributions

• First systematic evaluation of remote-sensing-driven machine learning frameworks for ETo estimation across contrasting agro-climatic zones in Morocco.
• Demonstrated the ability of satellite-derived variables, particularly land-surface temperature, to effectively compensate for missing ground-based solar radiation and temperature data, maintaining high accuracy.
• Provided a comprehensive assessment of model transferability, highlighting robustness within similar regions and limitations across heterogeneous climatic zones, emphasizing the need for localized recalibration.
• Offers a cost-effective and scalable solution for operational ETo estimation in data-scarce environments, supporting smart irrigation and precision water management strategies in Morocco and similar regions.

Funding

This research did not receive any specific grant from the funding agencies in the public, commercial, or not-for-profit sectors.

Citation

@article{Belarbi2026Machine,
  author = {Belarbi, Zaid and Younoussi, Yacine El},
  title = {Machine learning estimation of reference evapotranspiration using MODIS-Derived and limited ground variables across Moroccan agro-climatic zones},
  journal = {Theoretical and Applied Climatology},
  year = {2026},
  doi = {10.1007/s00704-026-06193-z},
  url = {https://doi.org/10.1007/s00704-026-06193-z}
}