Evaluating Soil Property Maps and Associated Uncertainty Using Prediction Intervals Derived from Conventional and Disaggregated Soil Maps.

Of the different approaches to producing soil property data to meet GlobalSoilMap (GSM) standards, leveraging legacy vector soil surveys has potential to produce much of the initial version of GSM property grids in many countries. Our purpose is to evaluate using different scales of legacy soil map data to produce GSM surfaces with attributed prediction intervals (PI) to address uncertainty. Direct use of the USDA-NRCS State Soil Geographic (STATSGO2) database and the USDA-NRCS Soil Survey Geographic (SSURGO) database were compared with more appropriately scaled disaggregated SSURGO (dSSURGO) data for producing property grids with PI. An equal area spline was used to create estimates at the appropriate GSM depth breaks. The dSSURGO data was created from SSURGO using random forest models that trained satellite imagery, terrain metrics, and lithology maps to typified locations of soil series classes. Underlying probabilities from the resampled classification tree (RCT) model were used to specify fuzzy membership of soil classes for querying actual soil property values from the original SSURGO attribute tables. Property maps and associated PI of soil pH, organic carbon, and clay percentage of fines were produced with all three data sources. High, representative, and low values listed in the soil series property description tables were used to create a property value and PI boundaries. To create a prediction value for each property source data were aggregated as a weighted mean (i) by map unit polygon component percentage of soils in each map unit as a weighted average for the STATSGO2 and SSURGO versions and (ii) by using probabilistic fuzzy membership of soil series for dSSURGO from the underlying RCT model. To create PI the highest high value and lowest low value for all soils included as components in STATSGO and SSURGO, and from the soils included with a minimum membership threshold in the dSSURGO model. Property and PI maps produced by these methods were evaluated in the Eastern Allegheny Plateau and Mountains of southern West Virginia using laboratory characterized soil profile descriptions.