Disaggregating a regional-extent digital soil map using Bayesian area-to-point regression kriging for farm-scale soil carbon assessment

Abstract

Most soil management activities are implemented at farm scale, yet digital soil maps are commonly available at regional or national scale. Disaggregating these regional and/or national maps is applicable for farm-scale tasks, particularly in data-poor or limited situations. Although disaggregation is a frequently discussed topic in recent digital soil mapping literature, the uncertainty of the disaggregation process is not often discussed. Underestimation of inferential or predictive uncertainty in statistical modelling leads to inaccurate statistical summaries and overconfident decisions. The use of Bayesian inference allows for quantifying the uncertainty associated with the disaggregation process. In this study, a framework of Bayesian area-to-point regression kriging (ATPRK) is proposed for downscaling soil attributes, in particular, maps of soil organic carbon. An estimation of point support variograms from block-supported data was carried out using the Monte Carlo integration via the Metropolis–Hastings algorithm. A regional soil carbon map with a resolution of 100 m (block support) was disaggregated to 10 m (point support) information for a farm in northern New South Wales (NSW), Australia. The derived point support variogram has a higher partial sill and nugget, while the range and parameters do not deviate much from the block support data. The disaggregated fine-scale map (point support with a grid spacing of 10 m) using Bayesian ATPRK had an 87 % concordance correlation with the original coarse-scale map. The uncertainty estimates of the disaggregation process were given by a 95 % confidence interval (CI) limit. Narrow CI limits indicate that the disaggregation process gives a fair approximation of the mean soil organic carbon (SOC) content of the study site. The Bayesian ATPRK approach was compared with dissever, which is a regression-based disaggregation algorithm. The disaggregated maps generated by dissever had 96 % concordance correlation with the coarse-scale map. Dissever achieves this higher concordance correlation through an iteration process, while Bayesian ATPRK is a one-step process. The two disaggregated products were validated with 127 independent topsoil carbon observations. The validation concordance correlation coefficient for Bayesian ATPRK disaggregation was 23 %, while downscaled maps generated from dissever had 18 % concordance correlation coefficient (CCC). The advantages and limitations of both disaggregation algorithms are discussed.