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 (Tmax and Tmin), mean relative humidity (RH), wind speed (WS), and sunshine hours (SSH). 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 Tmax, RH, Ws, SSH, and Tmin. The results of the current study are substantial as it offers an approach to estimate ET0 in semi-arid data-scarce region. Four machine learning models (linear regression [LR], support vector machine [SVM], random forest [RF], and neural network [NN]) were assessed to estimate ET0 using minimal climate data.The combination of two, three, and four input variables showed promising results for ET0 estimation using SVM under full and limited dataset conditions.The SVM model outperformed the other models in terms of accuracy by giving high values of R2 (0.86-0.98) and lowest values for MAE (0.168-0.529), RMSE (0.22-0.70), and MAPE (5.91%-18.13%).The RF model could not predict the ET0 precisely with poorest values of R2 (0.79-0.89), MAE (0.51-0.682), RMSE (0.63-0.85), and MAPE (17.65%-23.04%).