COMPARISON OF REGIONAL AND GENERALMACHINE LEARNING MODELS FOR REFERENCEEVAPOTRANSPIRATION PREDICTION IN TURKEY

Excessive freshwater usage in irrigation is an important issue in Turkey, which can be reduced by using evapotranspiration (ET) measurements to schedule irrigation based on crops' exact water need. The infeasibility of direct ET measurement has driven researchers to estimate its value from meteorological variables. In recent years, machine learning (ML) has been widely used for this purpose and showcased its potential in estimating reference evapotranspiration () from limited variables. In this study, data collected from 165 weather stations in Turkey ranging between the years 1967 and 2020 was used to explore the spatial generalizability of ML models used for estimation by comparing two modelling scenarios: (1) One model for the entire country (2) One model for each geographic region in Turkey. Three ML algorithms were used for this purpose: Support Vector Regression (SVR), Random Forests (RF), and Gaussian Process Regression (GPR). Furthermore, stations were split into train and test stations to allow for conducting a cross-station evaluation. The models were tested using 16 input variable combinations. In general, SVR and GPR models performed better than RF models, which were prone to overfit. For SVR and GPR models, the change in mean absolute error (MAE) between the two scenarios ranged between -16.3% and -0.4% and the change in root mean squared error (RMSE) was between -17.0% and 1.1%, where negative values indicate a lower prediction error in the regional models' scenario. Finally, a randomization test conducted on four of the input combinations showed that the reduction in prediction error when using regional models is statistically significant. Keywords: Reference evapotranspiration, machine learning, cross-station evaluation, regional model, support vector regression, Gaussian processes, random forests