How well does Predictive Soil Mapping represent soil geography? An investigation from the USA

Abstract. We present methods to evaluate the spatial patterns of the geographic distribution of soil properties in the USA, as shown in gridded maps produced by Predictive Soil Mapping (PSM) at global (SoilGrids v2), national (Soil Properties and Class 100 m Grids of the USA), and regional (POLARIS soil properties) scales, and compare them to spatial patterns known from detailed field surveys (gSSURGO). The methods are illustrated with an example: topsoil pH for an area in central New York State. A companion report examines other areas, soil properties, and depth slices. A set of R Markdown scripts is referenced so that readers can apply the analysis for areas of their interest. For the test case we discover and discuss substan- tial discrepancies between PSM products, as well as large differences between the PSM products and legacy field surveys. These differences are in whole-map statistics, visually-identifiable landscape features, level of detail, range and strength of spatial autocorrelation, landscape metrics (Shannon diversity and evenness, shape, aggregation, mean fractal dimension, co-occurence vectors), and spatial patterns of property maps classified by histogram equalization. Histograms and variogram analysis revealed the smoothing effect of machine-learning models. Property class maps made by histogram equalization were substantially different, but there was no consistent trend in their landscape metrics. The model using only national points and covariates was not better than the global model, and in some cases introduced artefacts from a lithology covariate. Uncertainty (5–95% confidence intervals) provided by SoilGrids and POLARIS were unrealistically wide compared to gSSURGO low and high estimated values and show substantially different spatial patterns. We discuss the potential use of the PSM products as a (partial) replacement for field-based soil surveys.