Predicting reference evapotranspiration in semi-arid regions using optimized machine learning models

ABSTRACT Accurate estimation of reference evapotranspiration (ETo) is essential for efficient agricultural water management, especially in arid and semi-arid regions where climate variability limits traditional methods. This study evaluates the FAO Penman–Monteith (PM) equation and five machine learning (ML) models using a 5-year (2019–2023) dataset from two California stations: Parlier (Mediterranean) and Meloland (desert). Among 31 input combinations, CatBoost outperformed others, achieving an R2 of 0.995 and a root mean squared error (RMSE) of 0.161 mm/day at Parlier, and an R2 of 0.991 and an RMSE of 0.211 mm/day at Meloland. Using only mean temperature and net radiation, CatBoost maintained strong performance (R2: 0.932–0.948; RMSE: 0.507–0.589 mm/day). Bayesian optimization enhanced model tuning and reduced overfitting. Compared with random forest and XGBoost, CatBoost reduced RMSE by 5–12% and generalized well across sites. Violin boxplots confirmed its stability and accuracy. A user-friendly graphical interface was also developed for real-time ETo estimation. These results highlight the promise of ML-based ETo prediction tools for practical, scalable use in data-limited environments.