Evapotranspiration evaluation models based on machine learning algorithms—A comparative study

The constant need to increase agricultural production, together with the more and more frequent drought events in many areas of the world, requires a more careful assessment of irrigation needs and, therefore, a more accurate estimation of actual evapotranspiration. In recent years, several water management issues have been addressed by means of models derived from Artificial Intelligence research. When using machine learning based models, the main challenging aspects are represented by the choice of the best possible algorithm, the choice of adequately representative variables and the availability of appropriate data sets. Machine learning algorithms may be a powerful tool for the prediction of actual evapotranspiration, when a time series of few years is available. Starting from the measurements of a sufficient number of climatic parameters it is possible to obtain forecasting models characterized by very high accuracy and precision. Three different evapotranspiration models have been compared in this study. The models differ in the input variables. Four variants of each model were applied, varying the machine learning algorithm: M5P Regression Tree, Bagging, Random Forest and Support Vector Regression. The data refers to an experimental site in Central Florida, characterized by humid subtropical climate. The best outcomes have been provided by Model 1, whose input variables were net solar radiation, sensible-heat flux, moisture content of the soil, wind speed, mean relative humidity and mean temperature. Model 3, built from data only of mean temperature, mean relative humidity and net solar radiation, has provided still satisfactory results. Model 2, which adds the wind speed to the input variables of Model 3, has provided results that are absolutely comparable to those of Model 3 itself.