Uncertainty analysis of sample locations within digital soil mapping approaches

Abstract

In digital soil mapping the spatial distribution of soil classes or properties is quantified by formulating empirical spatial or non-spatial soil inference systems between soil observations and spatially referenced environmental covariates. Uncertainty about the location of soil samples, however, will inflate the uncertainty in these predictive relationships. In this study we demonstrate the influence of location error on digital soil mapping prediction results by spatially corrupting samples within a fixed neighbourhood range around their original location. The corruption range was estimated from the maximum expected spatial uncertainty. Two corruption models were tested. In the systematic corruption model only one sample was shifted at one time, while the others were kept constant at their original locations. This approach therefore allows for the examination of single sample locations, and hence the identification of data values for which location error will have a marked effect on final predictions. In the random corruption model all samples were simultaneously shifted at the same time. This allows to calculate the upper and lower boundary of the range of prediction accuracy by combining either the best or the worst averaged prediction results within each corruption location. The introduced corruption models were applied to 167 soil organic carbon content samples for an area of 150 km 2 located in Rhineland–Palatinate, Germany. Soil organic carbon contents were compared for the A-horizon as well as the 0–10 cm depth interval. The prediction results (root mean square error; RMSE) obtained from the original locations is 2.72 for the A-horizon and 1.51 for the 0–10 cm depth interval. The minimum and maximum RMSE values based on spatially corrupting the original locations are 2.36 and 3.68 as well as 1.24 and 1.82 for the A-horizon and the 0–10 cm depth interval respectively. Moving or removing single samples with a strong negative influence on prediction accuracy decreases the prediction error by a magnitude of ~ 0.1 RMSE. We conclude that the introduced corruption approaches should be applied to areas where soil scientists have to deal with legacy data that is afflicted with spatial uncertainty and/or environmental predictors that are spatially imprecise to assess and analyze spatial uncertainty for error propagation and the detection of locations with high potential uncertainty.