Reference evapotranspiration estimation using machine learning approaches for arid and semi-arid regions of India

Abstract

Accurate estimation of reference evapotranspiration (ET0) is vital for hydrological studies and irrigation scheduling. This study aimed to estimate ET0 using 4 machine learning algorithms: random forest, support vector machine, light gradient boosting decision trees and extreme gradient decision trees. Daily data for 2001 to 2020 at 11 (arid and semi-arid) stations were used for modelling. Eighteen scenarios with different input combinations were evaluated using the data of maximum and minimum air temperature, mean relative humidity and wind speed, number of sunshine hours, solar radiation, and extra-terrestrial radiation data at these stations. The ET0 estimated using the FAO 56 Penman-Monteith equation was chosen as the target value for model fitting. The best input combination was found in the models that used all inputs, while the least accurate were the models that used temperature data only. The results showed that the support vector machine models outperformed the other models at most stations. The application of various input combinations indicated that the use of fewer inputs also gave reasonable accuracy in the modelling. In addition, wind speed and solar radiation were found to be important parameters for precise estimation.