Estimating reference evapotranspiration using a machine learning approach

Weather parameters that influence evapotranspiration are air temperature, solar radiation, relative humidity and wind speed. Daily weather data during the wheat-growing period were collected from 1970 to 2018 for the Amritsar district of Punjab state in India. To improve evapotranspiration estimation, a well-defined area of artificial intelligence called machine learning is used. To improve the accuracy of evapotranspiration estimation during the wheat-growing period, a model was developed by random forest (RF), support vector machine (SVM) and artificial neural network (ANN) using different weather input combinations. Based on the evaluation done using various standard statistical criteria during calibration and validation performance of RF was found to be best, followed by SVM and ANN. The model developed by (<i>T</i>_max, <i>T</i>_min, RHM, RHE and <i>R</i>_s) weather input combination was ranked first. Two weather input combinations (<i>R</i>_s, <i>T</i>_max) and (<i>R</i>_s, <i>T</i>_min) performed excellently by RF and SVM, while the weather input combination (<i>T</i>_max, <i>T</i>_min) performed excellently by the ANN. Hence, these input combinations can be used in the estimation of evapotranspiration when the availability of data is limited. From this study, it can be concluded that instead of a large amount of weather data, ET₀ estimation can be done with a few data points by the machine learning technique.