Reference evapotranspiration (ETo) is useful for water management, calculating crop water requirements and irrigation scheduling. ETo was estimated from 5 empirical methods based on temperature, 5 based on solar radiation and on Machine Learning Technique (MLT). The MLT model consisted of Artificial Neural Networks (ANNs) and Support Vector Machine (SVM), with 6 architectures each one. The MLT and empirical methods were tested against Penman Monteith FAO 56 method based on the following statistical parameters: MBE (Mean Bias Error), RMSE (Mean Square Root Error), d (coefficient of Willmott) and R2 (coefficient of determination). The meteorological data used (maximum temperature, minimum and average temperature, relative humidity, wind speed and sunshine hours: n) were obtained from the National Institute of Meteorology of Mozambique. The results obtained from the modeling showed the following: Jones and Ritchie (JRICH) > Makkink, SVM3 > SVM6 > SVM1 > SVM2 = SVM4 > SVM5 > ANN5> Abtew > Hargreaves â€“ Samani > ANN1 = ANN6 > ANN4 > Irmak > ANN3 > Jensen Haise > ANN2 > Blaney Criddle Original > Schendel > Kharrufa > Mc Guinness-Bordne. Global solar radiation, which is one of the variables needed for the JRICH method (MBE = -0.17 mm day-1; RMSE = 0.38 mm day-1; d = 0.98 and R2 = 0.98) is not always measured or calculated. In this case, SVM1 could be used since it only requires measurements of T (MBE = 0.16 mm day-1; RMSE = 0.62 mm day-1; d = 0.94 and R2 = 0.83).
Keywords: Artificial neural networks, evapotranspiration, empirical methods, machine learning technique, support vector machine.