Derivation and predictive performance of pedotransfer functions: an empirical investigation

Summary The development of pedotransfer functions (PTFs) needs to recognize and account appropriately for the underlying structure in data that is imposed by the sampling design adopted to ensure that the data support any scientific inferences. We fitted several potential PTFs for field capacity (FC) and permanent wilting point (PWP), with and without taking account of the sampling design, to data collected with a nested spatial sampling design. Leave-one-paddock-out cross-validation (LOPOCV) and the Akaike information criterion (AIC) were used to identify PTFs with strong predictive performance. The results suggest that cross-validation statistics appear to provide no additional useful information on the predictive performance of PTFs beyond that provided by the AIC and other model fitting statistics that, in contrast to LOPOCV, are based on the entire dataset. Results further suggest that PTFs derived with the sampling design accounted for had consistently smaller AICs and potentially better predictive performance than when no account was taken. We identified eight PTFs for FC and six PTFs for PWP with strong predictive performance for soil used for dryland cropping in Victoria that were derived from routinely measured soil data with the sampling design taken into account. Highlights Derivation of PTFs should be based on appropriate analysis of data that accounts for the underlying structure in data. A linear mixed effects modelling approach to derive PTFs is proposed that takes account of a nested sampling design. The PTFs derived with sampling design accounted for had consistently stronger predictive performance. Derivation of PTFs should account for the underlying structure in data with appropriate linear mixed effects models.