Reference evapotranspiration prediction using machine learning models: An empirical study from minimal climate data

The scarcity of climatic data is the biggest challenge for developing countries, and the development of models for reference evapotranspiration (ET₀) estimation with limited datasets is crucial. Therefore, the current investigation assessed the efficacy of four machine learning (ML) models, namely, linear regression (LR), support vector machine (SVM), random forest (RF), and neural networks (NN), to predict ET₀ based on minimal climate data in comparison with the standard FAO-56 Penman-Monteith (PM) method. The data on daily climate parameters were collected for the period 2000−2021, including maximum and minimum temperatures (T_max and T_min), mean relative humidity (R_H), wind speed (W_S), and sunshine hours (S_SH). The performance of the developed models considering different input combinations was evaluated by using several statistical performance measures. The results showed that the SVM model performed better than the other ML models during training (R² = 0.985; mean absolute error [MAE] = 0.170 mm/day; mean square error [MSE] = 0.052 mm/day; root mean square error [RMSE] = 0.229 mm/day; mean absolute percentage error [MAPE] = 5.72%) and testing stages (R² = 0.985; MAE = 0.168 mm/day; MSE = 0.050 mm/day; RMSE = 0.224 mm/day; MAPE = 5.91%) under full dataset scenario. The best performance of the models to estimate was with T_max, R_H, W_s, S_SH, and T_min. The results of the current study are substantial as it offers an approach to estimate ET₀ in semi-arid data-scarce region.