Abstract
Due to the absence of experimental reference evapotranspiration (ETo) records, data-driven models con-sider in most cases calculated ETotargets to train and test the models, usually according to the standardFAO56 Penman Monteith equation (FAO56-PM). This procedure is also adopted for calibrating moreconventional empirical approaches like the well-known Hargreaves (HG) equation. This study aims atassessing the performance implications derived from using calculated targets instead of experimentalmeasurements for training and testing data-driven models or calibrating empirical methods. Thereforean application of a gene expression programming (GEP) based approach for estimating ETois presentedconsidering calculated and lysimetric targets fed with two different input combinations and assessedthrough k-fold testing. The same procedure is adopted to evaluate the calibration of the HG equation.Finally, the FAO56-PM and the HG equations are compared with their corresponding GEP models bearingin mind the type of targets used. The locally trained GEP4 and GEP6 models trained using the experimen-tal lysimetric targets are more accurate than the corresponding HG and FAO56-PM equations, assessedusing lysimetric benchmarks. The external performance accuracy of GEP4 and GEP6 models decreasesconsiderably in the cross-station approach using experimental targets. In this case, the FAO56-PM andthe HG equations might be preferable. The accuracy of the GEP models trained with calculated targetsdecreases considerably when the performance is assessed using experimental benchmarks. The conclu-sions drawn when only calculated benchmarks are used might be not sound or even false. The sameapplies for empirical equations calibrated with calculated targets. Four new GEP-based equations (oneper input combination and station) are provided to estimate ETo.