Digital soil property mapping and uncertainty estimation using soil class probability rasters

• We developed PROPR, a digital soil property mapping algorithm.
• PROPR calculates weighted-mean soil property maps with probability rasters.
• PROPR estimates the 90% prediction interval around the weighted mean estimate.
• PROPR may be useful where point observations are limiting but legacy maps are not.

The objective of the work presented in this paper was to investigate how rasters of the probabilities of occurrence of soil classes may be used to create digital soil property maps and maps of their associated uncertainties. The approach we present is formalised in an algorithm we developed called “Digital Soil Property Mapping Using Soil Class Probability Rasters” (PROPR).The soil class probability rasters were derived previously from a spatial disaggregation of the 1:250,000-scale Dalrymple Shire legacy soil polygon map from central Queensland, Australia.We created digital soil property maps of soil pH 1:5 H2O and their uncertainties (as indicated by estimates of the limits of the 90% prediction interval) at six depth increments down the soil profile (0–5 cm, 5–15 cm, 15–30 cm, 30–60 cm, 60–100 cm, 100–200 cm). The calculation of the weighted mean soil pH value for each depth increment at each grid cell was based on reference pH values for each soil class and used the probabilities of occurrence at each grid cell as weights.The calculation of the prediction interval limits for each depth increment involved sampling from the triangular distribution of the soil pH of each soil class using the soil class probabilities at each grid cell as weights in order to identify the number of samples to draw from each distribution. The 90% prediction interval limits were then estimated as the 5th and 95th percentiles of the distribution of samples drawn from the soil classes' triangular distributions.The maps of soil pH displayed strong spatial patterns. Soil pH generally increased with depth. Uncertainty generally increased with depth. Validation on 300 randomly-selected soil profiles returned a Lin's concordance correlation coefficient of 0.193 at the surface increasing to 0.266 at depth. RMSE increased with depth from about 0.75 pH units at the surface to 1.15 at depth.Soil class probability rasters are useful for generating digital soil property maps and maps of the associated uncertainties. Validation left room for improvement but the quality of the results is probably strongly affected by the quality of the spatial disaggregation that produced the soil class probability rasters. The PROPR approach may be useful in situations where profile observations are limiting but where legacy soil maps are available.Generation of the soil class probability rasters to use in PROPR is a predictive exercise in itself and so is also subject to uncertainty. The probability rasters likely can be derived by several methods including logistic regression and data mining; the probability rasters we used were derived via spatial disaggregation of a legacy soil polygon map.PROPR may be useful in situations where profile observations are limiting but where legacy soil maps are available. The soil class probability rasters need to be produced separately. Information on the within-soil-class variability of the target soil property at each depth increment must be known in order to establish the triangular distributions for the uncertainty estimation.PROPR may reduce reliance on having sufficient soil profile observations in areas where such data is limiting. We used available profile observations to estimate the within-soil-class variability in order to establish triangular distributions for the soil classes in our study area, but the information required to do so could also be derived from legacy reports or expert knowledge.