Estimation of Reference Evapotranspiration (ET0) using Intelligent Methods based on Lysimetric Measurements with different soils and grass species

Abstract

Abstract Four intelligent methods of Gene Expression Programming (GEP), Support Vector Machines (SVM), Multivariate Adaptive Regression Splines (MARS), and Random Forests (RF) were employed and run to estimate the reference evapotranspiration (ET0). Based on the four categories of radiation, mass transmission, temperature, and combination models, weather variables are used in four groups as inputs of the intelligent models. To involve all data in the training and testing process of the models, the K-fold cross-validation technique was employed. Finally, evaluation of the presented intelligent models was performed with measured data of the six sets of lysimetric data (two grass species × three soil textures), also the effect of soil texture and species on intelligent models was investigated. The results obtained indicate that the RF models performed the best among the four groups, which is followed by SVM, MARS, and GEP models, respectively. The best performance belonged to the RF model from the Combination group (RF1) with the values of coefficient of determination, R2 = 0.833; mean bias error, MBE = -0.007 mm/day; mean absolute error, MAE = 1.07 mm/day; root mean square error, RMSE = 1.357 mm/day, and scatter index, SI = 0.184. Based on the ranks of accuracy, the order of the models, from the best to the worst, were: RF2, MARS1, SVM1, RF3, MARS2, GEP1, SVM2, GEP2, RF4, SVM3, MARS3, SVM4, GEP3, GEP4, and MARS4, respectively. The estimated ET0 values by the models indicated higher accuracy for the lysimeters with Sandy Loam Soil compared to other lysimeters. The performance comparison of the intelligent models for the planted grass species in lysimeters yielded different results, as the accuracy was higher for the Festuca grass in Sandy Loam soils compared to the Lolium grass, in contrary in clay and Silt Loam soils the performance of Lolium grass was better.