Pedotransfer Functions Versus Model Structure: What Drives Variance in Agro‐Hydrological Model Results?

ABSTRACT Pedotransfer functions (PTFs) are widely used empirical relationships to estimate soil hydraulic parameters. PTFs are usually derived from point soil samples analysed in the field or laboratory; thus, they contain uncertainties at different levels (i.e., from sampling and measuring techniques, as well as empirical approaches chosen to quantify relationships). When PTFs are used to parametrize agro‐hydrological models, both the choice of PTF and the choice of the model may influence the simulation results. Both sources of variance (PTF choice and model structural differences) were found to be relevant in previous studies, but how they relate to each other has rarely been investigated. In this study, we addressed this research gap by conducting a systematic analysis of the variance in selected agro‐hydrological model outputs (i.e., seepage water, soil water content, actual evapotranspiration, transpiration, biomass production) based on an ensemble of 18 PTFs applied to four agro‐hydrological models, namely: APEX, CANDY, DAISY and SWAP. The models were calibrated for aboveground biomass and phenology of silage maize and evaluated using data of actual evapotranspiration, seepage water and soil water content obtained from a lysimeter facility in Switzerland. ANOVA‐based variance partitioning was applied to attribute variance in model outputs to two uncertainty sources (PTF choice, model choice). Overall, we found that agro‐hydrological model structural differences had a larger influence on the variance in model outputs than PTF differences. Further analyses undertaken per model showed that the sensitivity of the simulated outputs to the choice of PTF differed between the models; our results showed that the models integrating the Richards equation (SWAP, DAISY) were more sensitive to the choice of PTF than those using a reservoir cascade approach (APEX, CANDY). Our results also showed that simulated outputs using the mean of a PTF ensemble performed better than when using a single PTF, irrespective of the model and output variable. We therefore recommend using PTF ensembles in agro‐hydrological modelling studies. The benefit of using large PTF ensembles is, however, likely to be reduced in larger ensembles of agro‐hydrological models, as structural model uncertainties will dominate over PTF uncertainties, according to the four‐member model ensemble investigated here.