The impact of direct and indirect digital soil mapping approaches on spatial uncertainty

Abstract

While numerous studies have compared different mapping approaches (or inference trajectories) in digital soil mapping (DSM), their impact on uncertainty quantification and propagation has received less attention. The objective of this study was to investigate key questions related to uncertainty quantification and propagation, which may influence the applicability of DSM products. Including, how do different mapping approaches and assumptions made in DSM affect uncertainty quantification and propagation? What are the pros and cons of the mapping approaches from the perspective of uncertainty? Such questions were examined on the example of mapping soil organic carbon (SOC) in the Great Hungarian Plain, Hungary, by combining machine learning with univariate and multivariate geostatistics. Two cases were investigated: in Case 1, the goal was to map SOC stock for the year 2016 using both direct and indirect mapping approaches and to quantify and compare prediction uncertainty at various spatial aggregation levels. In Case 2, the objective was similar, but focused on mapping SOC stock change between 1992 and 2016. A wide range of inference trajectories (e.g., “calculate then model”, “model then calculate”), prior data transformations (e.g., square root, standard normalization), and uncertainty propagation techniques (e.g., Taylor method, analytical solution) were applied and compared from the perspective of both prediction accuracy (e.g., mean error, root mean square error) and uncertainty (e.g., prediction interval coverage probability plot, interval scores). The results showed that both the chosen inference trajectories and the assumptions made in DSM significantly impact uncertainty estimates not only at point support but also at larger supports. It also highlighted the importance of accounting for the correlation of interpolation errors when conducting uncertainty propagation. Additionally, this research emphasized the need to identify and quantify the contribution of different error sources in uncertainty propagation, as this can be the key to reduce the overall uncertainty associated with the given soil property or function of interest.