Skip to main content
ARS Home » Southeast Area » Booneville, Arkansas » Dale Bumpers Small Farms Research Center » Research » Publications at this Location » Publication #388349

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

item ROSSITER, DAVID - Isric - World Soil Information
item POGGIO, LAURA - Isric - World Soil Information
item BEAUDETTE, DYLAN - Natural Resources Conservation Service (NRCS, USDA)
item Libohova, Zamir

Submitted to: Soil
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 9/13/2021
Publication Date: 9/13/2021
Citation: Rossiter, D.G., Poggio, L., Beaudette, D., Libohova, Z. 2021. How well does Predictive Soil Mapping represent soil geography? An investigation from the USA. Soil.

Interpretive Summary: We present four case studies in the continental USA illustrating the application of methods, developed in a companion paper in the journal SOIL Discussions (, to evaluate the spatial patterns of the geographic distribution of soil properties as shown in gridded maps produced by Predictive Soil Mapping (PSM) at global (SoilGrids v2), national (Soil Properties and Class 100m Grids of the USA), and regional (POLARIS soil properties) scales, and compare them to spatial patterns known from detailed field survey (gSSURGO). These case studies reveal substantial differences in the performance of PSM due to (1) study area and its soil geomorphology; (2) soil property being predicted; (3) depth being predicted. Each case is unique and reveals different aspects of the reference and PSM products. A set of R Markdown scripts is referenced so that readers can apply the analysis for areas of their interest.

Technical Abstract: The above analysis shows clearly that PSM is no substitute for field survey. Soil geography is often subtle, as field surveyors well know. It can be challenging to form a proper mental model of the soil-landscape relations in a survey area, so it is not surprising that using proxies (environmental covariates) rather than direct observation is not as accurate. This is known from cross-validation or other numerical evaluation exercises of point observations and their PSM predictions. Here we show that the spatial patterns are also not well-reproduced. This is especially relevant for earth surface models that use groups of grid cell predictions and their spatial contiguity, for example, watershed hydrology. On the other hand, not all soil surveyors are equally competent, and the actual soil observations (augerings, profiles) are few, so that the consistent PSM approach may be more accurate in areas where surveyors were either less competent or where the soil-landscape relations were complex and difficult to map in the field. PSM can also smooth out sharp boundary lines between mapped polygons, when these are in nature gradual and where the STU included in the SMU on either side of the boundary are not too different. Surprisingly, the inclusion of parent material and drainage class, and the use of only CONUS-wide covariates, did not improve predictive maps in the test areas. This is clear from the comparison of SPCG and SG2. Geomorphology has proven to be a key component of soil survey, and PSM has great difficulties representing geomorphology, as opposed to landforms. If the landform and geomorphology are not congruent, and the PSM data source does not have a covariate to represent the geomorphology, several geomorphic units will be combined into similar predictions. A typical example is recently-glaciated terrain (Rossiter, 2016) where a given landform may have several geomorphic origins. A long linear low hill may be an esker, a lateral moraine, a drumlin, or thin till over a pre-existing rock structure. For example, in the Central New York example the valley trains of glacial outwash are in low positions, suggesting fine textures, but in fact have a large content of coarse fragments and coarser textures. They also have a pH derived from their source rocks carried by the glacier, not the rocks of the surrounding areas. PSM also has problems identifying soil age. For example, in the North Carolina coastal plain example the different ages of the marine terraces are not clearly differentiated by the slight elevation differences separated by scarps, well-known to local soil surveyors. These age differences result in different degrees of development of the WRB Acrisols (USDA Soil Taxonomy Ultisols), especially the horizon thickness and in the oldest positions the development of plinthite gravel Daniels et al. (1999). Perhaps the most important conclusion is that different PSM methods, with different training points and different algorithms, can produce quite different predictive soil maps. Comparing these with point-wise evaluation (“validation”) gives an incomplete picture of how the different methods represent the soil landscape, which is after all what dictates how the soil is used and managed. The obvious limitation of this study is that it only examines a few of the many study areas, and in each one either one or a few of the mapped soil properties. The relative success of different PSM methods vs. field study and among themselves will surely differ greatly as these are changed. We encourage readers to apply the methods to their own study areas within the USA and to their soil properties of interest, to themselves evaluate the utility of the several PSM products, and indeed the utility of PSM in general. For this, we provide our analysis scripts as R Markdown documents (R Studio, 2020), as explained in the Methods paper.