Scholarlink

<p>Accurate estimation of reference evapotranspiration (ETo) is essential for effective irrigation planning and sustainable water resource management, particularly in regions facing climatic variability and limited ground-based meteorological data. The standard FAO Penman-Monteith method relies on in-situ measurements such as solar radiation, which are often unavailable or incomplete in many developing countries. To overcome this limitation, this study investigates the use of machine learning models to estimate daily ETo using satellite-derived solar radiation—specifically from NASA POWER and ERA5—combined with other routinely available meteorological variables. The analysis focuses on two agro-climatic zones in Morocco: the semi-arid Doukkala region and the sub-humid Loukkos region. Eighteen years of data were compiled from three meteorological stations situated at different altitudes. A comprehensive set of machine learning models, including both standalone and hybrid configurations, were trained and optimized using the Optuna framework for hyperparameter tuning. Model performance was assessed under two evaluation strategies: local training (station-wise evaluation) and spatial generalization using a Leave-One-Station-Out (LOSO) approach. In the local evaluation, CatBoost and hybrid ensembles such as CatBoost + LGBM achieved superior accuracy, with R² scores exceeding 0.95 at the Doukkala stations and over 0.87 at M’rissa. Under the LOSO framework, predictive performance declined as expected, particularly at the sub-humid M’rissa station (R² ≈ 0.67), but remained strong at Mettouh (R² = 0.92) and Zemamra (R² = 0.91). Across both evaluation strategies, NASA POWER consistently outperformed ERA5, especially at the low-altitude M’rissa station, while both datasets provided comparable results in the semi-arid Doukkala region. These findings underscore the potential of machine learning—particularly hybrid ensembles—to deliver accurate and transferable ETo estimates using satellite-derived Rs in data-scarce environments, thereby supporting spatial generalization across contrasting climatic zones.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Assessment of machine learning models for ETo estimation and spatial generalization using satellite-derived solar radiation: a case study in two moroccan agro-climatic zones

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Zaid Belarbi,
  • Yacine El Younoussi

摘要

Accurate estimation of reference evapotranspiration (ETo) is essential for effective irrigation planning and sustainable water resource management, particularly in regions facing climatic variability and limited ground-based meteorological data. The standard FAO Penman-Monteith method relies on in-situ measurements such as solar radiation, which are often unavailable or incomplete in many developing countries. To overcome this limitation, this study investigates the use of machine learning models to estimate daily ETo using satellite-derived solar radiation—specifically from NASA POWER and ERA5—combined with other routinely available meteorological variables. The analysis focuses on two agro-climatic zones in Morocco: the semi-arid Doukkala region and the sub-humid Loukkos region. Eighteen years of data were compiled from three meteorological stations situated at different altitudes. A comprehensive set of machine learning models, including both standalone and hybrid configurations, were trained and optimized using the Optuna framework for hyperparameter tuning. Model performance was assessed under two evaluation strategies: local training (station-wise evaluation) and spatial generalization using a Leave-One-Station-Out (LOSO) approach. In the local evaluation, CatBoost and hybrid ensembles such as CatBoost + LGBM achieved superior accuracy, with R² scores exceeding 0.95 at the Doukkala stations and over 0.87 at M’rissa. Under the LOSO framework, predictive performance declined as expected, particularly at the sub-humid M’rissa station (R² ≈ 0.67), but remained strong at Mettouh (R² = 0.92) and Zemamra (R² = 0.91). Across both evaluation strategies, NASA POWER consistently outperformed ERA5, especially at the low-altitude M’rissa station, while both datasets provided comparable results in the semi-arid Doukkala region. These findings underscore the potential of machine learning—particularly hybrid ensembles—to deliver accurate and transferable ETo estimates using satellite-derived Rs in data-scarce environments, thereby supporting spatial generalization across contrasting climatic zones.