Skip to main content
ARS Home » Southeast Area » Booneville, Arkansas » Dale Bumpers Small Farms Research Center » Research » Publications at this Location » Publication #386456
Research Project: Sustainable Small Farm and Organic Grass and Forage Production Systems for Livestock and Agroforestry

Location: Dale Bumpers Small Farms Research Center

Title: How well does Predictive Soil Mapping represent soil geography? An investigation from the USA Part 2: Case Studies

Author
item ROSSITER, DAVID - ISRIC - WORLD SOIL INFORMATION
item BEAUDETTE, DYLAN - NATURAL RESOURCES CONSERVATION SERVICE (NRCS, USDA)
item Libohova, Zamir
item POGGIO, LAURA - ISRIC - WORLD SOIL INFORMATION

Submitted to: Meeting Abstract
Publication Type: Proceedings
Publication Acceptance Date: 11/1/2021
Publication Date: 11/15/2021
Citation: Rossiter, D.G., Beaudette, D., Libohova, Z., Poggio, L. 2021. How well does Predictive Soil Mapping represent soil geography? An investigation from the USA Part 2: Case Studies. Meeting Abstract. https://doi.org/10.17027/isricwdc-202103.
DOI: https://doi.org/10.17027/isricwdc-202103

Interpretive Summary:

Technical Abstract: Evaluation of Predictive Soil Mapping (PSM) products is usually based on summaries of point statistics, but map users often evaluate them based on how well they represent soil geography. We developed methods to evaluate the spatial patterns of the geographic distribution of soil properties in the lower 48 USA, as shown in gridded maps produced by global (SoilGrids v2.0), national (Soil Properties and Class 100m Grids of the US), regional (POLARIS) PSM, and compared them to spatial patterns known from detailed field survey (SSURGO). Over four areas known to the authors (central NY, coastal plain NC, southwestern IN, Sierra foothills CA) we identified large discrepancies in spatial patterns for several properties (pH, particle-size class proportions, SOM) at selected Global Soil Map standard depth slices, both by expert visual comparison of maps and by statistical pattern analysis. Visual comparison with SSURGO showed that PSM missed significant soil-geomorphic features. 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 indices. The model using national points and covariates was not better than the global model, and in some cases introduced artefacts from a lithology covariate. The disaggregation of SSURGO to 30 m resolution grid cells by POLARIS produced artefacts not related to within-polygon landscape features. PSM uncertainty estimates (SoilGrids and POLARIS 5/95% confidence intervals) were unrealistically wide compared to SSURGO low and high estimated values. Given a set of training observations, covariates, and machine-learning method, PSM gives reproducible and consistent results, but these may not satisfactorily represent soil geography. A set of R Markdown documents is available to repeat the analysis for areas and properties of interest.