Estimation of Daily Reference Evapotranspiration in Humid Climates Using Data-Driven Methods of Gaussian Process Regression, Support Vector Regression and Random Forest

Accurate estimation of reference evapotranspiration has great importance in irrigation scheduling. Moreover, the lack of availability of lysimetric data has led researchers to use indirect methods, including data-driven approaches. In the present study, the ability of Gaussian process regression (GPR), support vector regression (SVR) and random forest (RF) data-driven methods was investigated to estimate the evapotranspiration of the reference plant. For this purpose, meteorological data on average temperature, wind speed, relative humidity and sunny hours in the period 2013-18 were collected in nine northern stations of Iran including Astara, Bandar Anzali, Rasht, Ramsar, Nowshahr, Sari, Turkmen port, Gorgan, and Gonbad Kavous. Evapotranspiration calculated using FAO-Penman-Montith method was considered as the target output and four combined scenarios of meteorological parameters were considered to calibrate and validate the studied methods. The accuracy of the mentioned methods was compared using the statistical parameters of correlation coefficient, scatter index, and Wilmott’s coefficient. The results showed that GPR4 model with scatter index in the range of 0.132 to 0.179 in Astara, Bandar Anzali, Rasht, Ramsar, Nowshahr and Sari stations, SVR4 model with dispersion index of 0.116 to 0.120 in Turkmen and Gonbad Kavous stations and the Hargreaves-Samani method with a scatter index of 0.509 at Gorgan station had much more accurate estimates of the evapotranspiration of the reference plant.