Machine Learning as a Tool to Predict Reference Evapotranspiration

Abstract

Reference evapotranspiration (ET) is essential for agricultural planning and water management, especially in semi-arid areas like T & uuml;rkiye's Siirt province. This study predicted monthly ET using machine learning methods (SVM, XgBoost, KNN, RF) and statistical regression models (Quartile, PLS, LASSO, Ridge, Elastic Net, NonParametric, Linear), as well as a Design of Experiments (DoE) approach. Models used climate variables such as temperature, humidity, wind speed, precipitation, sunshine hours, snow thickness, pressure, cloudiness, and solar radiation. RF was the best-performing machine learning model (R-2=0.948), while XgBoost was the lowest (R-2=0.776). Among regression methods, NonParametric regression performed best (R-2=0.948), and PLS performed weakest (R-2=0.901). The DoE model showed the highest accuracy overall (R-2=0.987), identifying average temperature as the key variable affecting ET. The findings improve prediction accuracy and offer new insights for hydrological understanding and practical applications in water management, agriculture, and climate adaptation in Siirt.