Accurate reference evapotranspiration estimation with limited data for sustainable irrigation in eastern Morocco: a machine learning approach

Abstract

Accurate estimation of daily reference evapotranspiration (ETo) is essential for effective irrigation scheduling, improved water-use efficiency, and sustainable crop production in arid and semi-arid regions. Although the FAO-56 Penman–Monteith equation (ETo PM ) is widely accepted as the reference method, its reliance on complete meteorological data limits its applicability in data-scarce agricultural systems such as those in eastern Morocco. This study proposes a practical solution for daily ETo estimation under data-limited conditions by evaluating machine learning approaches against traditional empirical models. Daily climatic data (2001–2025) were collected from airport meteorological stations, NASA POWER reanalysis, and on-farm sensors at four representative locations in Eastern Morocco (Oujda, Berkane, Taourirt, and Figuig). ETo was computed using the ETo PM equation. Three ML algorithms—Support Vector Regression (SVR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost)—were developed under multiple input scenarios guided by feature-importance analysis, which identified solar radiation (Rsn), maximum temperature (Tmax), and relative humidity (RH) as dominant predictors. To ensure realistic evaluation and avoid information leakage, a strict chronological split was applied, using 2001–2018 for training and 2019–2025 for independent testing. The best ML configurations were then compared with widely used empirical ETo equations (Hargreaves–Samani, Jensen–Haise, Priestley–Taylor, Makkink, and Turc). Results showed that XGBoost consistently outperformed SVR and RF, achieving the best balance between accuracy and computational efficiency. The three-input configuration (Rsn + Tmax + RH) produced near-reference performance across all locations ( R 2 = 0.976; RMSE < 0.55 mm day −1 ; training time ≈ 4.48 s; RAM ≈ 0.13 MB), while the reduced two-input configuration (Rsn + Tmax) maintained reliable performance ( R 2 = 0.936; RMSE ≈ 0.60 mm day −1 ; training time ≈ 0.33 s; RAM ≈ 0.13 MB). In contrast, empirical approaches exhibited poor predictive capability and weak transferability, with negative R 2 values and substantially larger errors (RMSE generally > 4 mm day −1 , exceeding 7 mm day −1 in the weakest formulations). Overall, this machine learning–based framework provides a robust and operational alternative for daily ETo estimation in eastern Morocco, supporting efficient irrigation scheduling, improved agricultural water management, enhanced crop productivity, and sustainable agriculture in arid and semi-arid regions.