Comparative Evaluation of Machine Learning Models for Daily Reference Evapotranspiration Estimation Using Limited Meteorological Data

Abstract

Reference evapotranspiration (ETo) is a fundamental parameter for irrigation scheduling and efficient water resources management. This study aimed to compare the performance of Linear Regression (LR), second-order Polynomial Regression (PR), Decision Tree (DT), Random Forest (RF), Support Vector Regression (SVR), and Artificial Neural Network (ANN) models for daily ETo estimation. Daily reference ETo values were calculated using the FAO-56 Penman–Monteith method based on measured meteorological variables, including air temperature, relative humidity, wind speed, and solar radiation. A total of 246 daily observations were used, with 80% of the dataset allocated for model training and the remaining 20% for testing. To evaluate the contribution of meteorological variables to model performance, three input datasets were developed: (i) a complete dataset including all variables, (ii) a dataset excluding mean air temperature and relative humidity, and (iii) a dataset further excluding solar radiation. Model performance was evaluated using the coefficient of determination (R²), mean squared error (MSE), and mean absolute error (MAE). The results showed that second-order Polynomial Regression consistently outperformed the other models across all input datasets. The highest prediction accuracy was achieved using the complete dataset, with an R² of 0.9394, an MSE of 0.2505, and an MAE of 0.3839 mm day⁻¹. Excluding mean air temperature and relative humidity had only a minor effect on model accuracy, whereas the removal of solar radiation resulted in a substantial decline in the performance of all models. These findings indicate that solar radiation is a critical predictor for daily ETo estimation and that the Polynomial Regression model provides an accurate and computationally efficient alternative to the FAO-56 Penman–Monteith method, even when meteorological data are limited.