Accounting for the spatial range of soil properties in pedotransfer functions

Pedotransfer functions (PTF) are widely used in soil hydraulic property modelling. Accounting for spatial structures of soil properties for improving the model performance of PTF is increasingly discussed. To understand how model performance varies when PTF are trained with samples of different spatial structure of the input data, we developed 12 ePTF (ensemble PTF) with data input from differently sized spatial domains to predict field capacity (FC) and wilting point (WP) of agriculturally used soils in Austria. The training domains generally had diameters equal to or larger than the spatial range of the explaining variables (bulk density BD, organic carbon content OC, Sand, Silt, and Clay) and the response variable (FC or WP). A stepwise regression technique was used to train the ePTF, and both bootstrap and random sampling were used to evaluate the uncertainties of the various ePTF. We found that, a training domain considerably larger than the spatial range of the input variables did not help develop a roubust ePTF, particularly when applied on relatively larger scales, independent of their performances during the training stage. We conclude that, covering additional heterogeneous samples from outside of the spatial range of the input variables does not ensure an enhanced prediction capability of ePTF. Also, it might be worth paying more attention to the spatial structure of the predicted variable when its spatial range might be expected to be quite different from the predictors. This would have an implication for guiding sampling practices.