Developing and Assessing Prediction Intervals for Soil Property Maps Derived From Legacy Databases.

The GlobalSoilMap project aims to create a global grid of a variety of soil functional properties at a fine resolution. Uncertainty surrounding these property estimates is of utmost importance when utilizing soil maps for predictive purposes. For the initial version of the map being produced of the United States, property values were estimated from the U.S. General Soil Map (STATSGO2) database, which is a broad-based inventory of soil data recorded across the United States. Prediction intervals were developed from the low and high estimated property values provided in the database. For each map unit, this method provided a unique prediction interval that was likely to encompass property values of soils typically found in that map unit. We empirically evaluated these soil property prediction intervals derived from STATSGO2 for three soil properties: organic carbon content, pH, and clay content. Using measured property data from up to 722 pedons from the National Cooperative Soil Survey database, prediction intervals were assessed by modeling their coverage accuracy over a set of external validation data. The effects of soil depth, soil order, temperature regime, and moisture regime on prediction interval coverage were analyzed, and coverage was found to be 87.6% for organic carbon, 90.6% for pH, and 86.4% for clay. It is shown that legacy data from the United States that includes low and high property methods can be used to represent uncertainty in the form of prediction intervals. Coverage based on these methods closely approximates the nominal level of 90% specified for GlobalSoilMap products. Consistency of these intervals was demonstrated across a variety of soil orders, temperature regimes, and moisture regimes.