Incorporation of spatial autocorrelation improves soil–landform modeling at A and B horizons

Abstract

Research has shown that the performance of soil–landform models would improve if the effects of spatial autocorrelation were properly accounted for; however, it remains elusive whether the level of improvement would be predictable, based on the degree of spatial autocorrelation in the model variables. We evaluated this problem using 11 soil variables acquired from the A and B horizons along a hillslope of Žofínský Prales in the Czech Republic. The results showed that, with no exception, there were increases in R2 and decreases in the Akaike information criterion (AIC), residual autocorrelation, and root-mean-square errors (RMSEs), after incorporating the spatial filters extracted by spatial eigenvector mapping into non-spatial regression models. Furthermore, the improvement of the model was positively proportional to the degree of spatial autocorrelation, inherent in the soil variables. That is, there were strikingly linear and significant relationships, in which strongly autocorrelated soil variables (i.e., having a high Moran's I value) exhibited greater increases in R2 and decreases in AIC, residual autocorrelation, and RMSEs than their more weakly autocorrelated counterparts. These findings indicate that the degree of spatial autocorrelation present in soil properties can serve as a direct indicator for how much the performance of a traditional non-spatial soil–landform model would be enhanced, by explicitly taking into consideration the presence of spatial autocorrelation. More generally, our results potentially imply that the need for and benefit from incorporating spatial effects in geopedological modeling proportionally increases as the soil property of interest is more spatially structured (i.e., landform variables alone cannot capture soil spatial variability).