Comparison of Machine Learning and Empirical Methods for Data-Efficient Estimation of Reference Evapotranspiration in a Semiarid Region

Accurate estimation of reference evapotranspiration (ET0) is critical for efficient irrigation planning and water resource management, particularly in semiarid regions where meteorological data are limited. This study compares the performance of empirical models (Hargreaves, Thornthwaite, and Blaney–Criddle) and machine learning (ML) algorithms, i.e., support vector regression (SVR), random forest, decision tree, linear regression, and artificial neural networks, for estimating ET0 in the Çekerek subbasin of Türkiye. Each method was evaluated under multiple input scenarios simulating various data-scarce conditions. Results show that ML models significantly outperformed empirical equations, with SVR achieving the best performance (R=0.997; RMSE=0.145 mm/day; NSE=0.991) even with reduced input parameters. The superior performance of ML approaches is attributed to their ability to capture nonlinear relationships among meteorological variables and to maintain robustness under reduced-input scenarios, unlike empirical equations that are structurally limited. Among empirical approaches, the Hargreaves method yielded the most consistent results but remained sensitive to seasonal variation. The findings demonstrate the potential of ML-based models to provide accurate, data-efficient alternatives for ET0 estimation in operational water resource applications, especially in regions with sparse climate monitoring networks.