The accurate estimation of actual evapotranspiration (AET) is crucial for sustainable water resource management, especially in water-scarce and agriculturally intensive regions like Beijing and Tianjin, China. Traditional methods for AET estimation, whether empirical or physically based, often face limitations due to high data requirements, limited scalability, and sensitivity to input uncertainties. This creates a critical research gap in providing reliable AET predictions under data-limited conditions. To address this, we evaluated the efficacy of integrating four advanced machine learning (ML) models: Support Vector Machine (SVM), Gaussian Process Regression (GPR), Ensemble Tree, and Neural Network, with Bayesian hyperparameter optimization for AET modeling using the high-resolution TerraClimate dataset spanning 1958–2022. Key meteorological variables, including maximum and minimum temperature (Tmax and Tmin), solar radiation (SR), wind speed (WS), vapor pressure deficit (VPD), and precipitation (PPT), were selected through rigorous correlation and multicollinearity analyses. Model performance was assessed using the coefficient of determination (R2), mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE) on a 75:25 train-test split. Results demonstrate that the optimizable GPR model achieved the highest predictive accuracy (RMSE = 5.54, R2 = 0.98 on test data), outperforming other ML approaches and traditional empirical models. PPT, Tmin, and Tmax emerged as the most influential predictors for AET. Our findings reveal that ML models, particularly when optimized via Bayesian techniques, yield a robust, scalable, and data-efficient alternative for AET estimation in regions with limited meteorological records. This study establishes a new benchmark for AET modeling, with significant implications for irrigation scheduling, drought monitoring, and integrated water management in the North China Plain and comparable agro-ecological regions.
