Spatial dependence and multivariate stratification for improving soil carbon estimates in the Piedmont of Georgia

Abstract

The spatial dependencies and structures of soil C and related soil properties under pine, hardwood, and pasture landcover in the Piedmont area of Georgia are compared, as well as the C estimates and errors of plot averages vs. kriging. The ability of four sampling methods and five different sample sizes to estimate soil C and its distribution based on stratification of landcover and topography variables were also investigated. Many of the soil properties demonstrated dependencies equal to the maximum lag size, suggesting dependencies outside the scale of our plots. Spatial structures were also not consistent between blocks suggesting a covariate other than landcover. Kriged estimates incorporate spatial correlations and resulted in higher standard errors, which may reflect an underestimate of the errors associated with the plot-based samples due to autocorrelation. At 5 and 12% sampling proportions, Latin Hypercube Sampling and stratified random sampling produced the best approximations of population parameters of soil C distribution, and simple random sampling produced better mean estimates with higher precision than systematic random sampling.