Machine learning modeling of reference evapotranspiration in Central Luzon, Philippines

Reference evapotranspiration (ETo) is crucial for calculating irrigation requirements. Instruments that directly measure ETo are still costly and limited while the empirical models are data intensive. Meteorological data of Central Luzon, Philippines (1985-2019) were used to estimate ETo using the FAO Penman-Monteith method. The performances of machine learning algorithms in estimating ETo were analyzed using ground-based weather data. Optimal models were determined using decision thresholds (RMSE<0.39 mm day-1, R2>0.75, MSE<0.15 mm day-1, MAE<0.30 mm day-1). The models were further assessed using principal component analysis for finding relevant variables (σ2=0.95) and the Wilcoxon test for comparing two samples (α=0.05). Results show that optimal model required only two or three weather variables depending on the station. In general, the algorithms can be ranked as follows: Gaussian progress regression, Neural network, Support vector machines, Ensemble of trees, Regression trees, and Linear regression. The study reveals that machine learning can accurately predict ETo using ground-based weather data, and it can be a good alternative to data-intensive empirical models.