The products are of the following three kinds: (1) digital products based on traditional soil survey without any statistical modelling, (2) DSM products based on traditional soil survey products and enhanced by statistical modelling using environmental covariates, and (3) DSM products based on statistical modelling using training points and environmental covariates. This latter is the most common DSM method worldwide, especially for areas without extensive traditional soil surveys.

The first kind of product is represented by the reference products from the Natural Resources Conservation Service (NRCS) of the United States Department of Agriculture (USDA), based on extensive field survey, air photo interpretation, thematic maps, and expert evaluation of digital elevation mapping (DEM) derivatives. This is considered to be the most accurate information, despite the occasional presence of artefacts from the overall mapping programme, as explained later in this section. There are two closely related products from the NRCS.

At the national level, the National Soil Geographic Database, gNATSGO , is a composite of the Soil Survey Geographic Database (SSURGO; mostly 1:24000 scale), State Soil Geographic Database (STATSGO2; 1:250000 scale), and the detailed Raster Soil Surveys (RSS) database, according to the most detailed product available for all areas of the USA. It is aimed at users who require multi-state or CONUS (contiguous United States) extent mapping. For each state or equivalent political unit, the SSURGO or STATSGO2 polygon maps of soil map unit (SMU) produced by traditional survey have been rasterized to a grid, with each cell keyed to a SMU . Grid cells link to the best available (i.e. greatest detail STATSGO-SSURGO-RSS) SMU. The digital products are delivered at 30 and 90 m resolutions for the 48 contiguous states and the federal District of Columbia of the USA (abbreviated CONUS).

At the state level, gSSURGO is also available . This has a higher resolution (both 10 and 30 m) to minimize the degradation of the original polygon delineations and is a direct gridding of SSURGO polygons. It neither uses STATSGO2 for infilling nor RSS if available. It thus is a gridded version of the familiar SSURGO product that is used for local applications. gSSURGO is refreshed annually for those users who do not wish to mix STATSGO or the new raster soil surveys into their analysis.

The gridded gNATSGO and gSSURGO maps are derived from the polygons of SSURGO, which is a representation of those delineated by the field surveyors on stereopairs or orthophotos and subsequently converted to vector digital format by manual digitization. Soil surveys conducted in the last 15 years were compiled using on-screen digitization in a geographic information system (GIS). At boundaries between survey areas, polygon lines at survey limits have been matched during digitizing . These polygons are organized in soil map units (SMUs), with one or more components (soil taxonomic units, STUs) usually named for a soil series but more specific than the parent soil series concept. Taxa above the soil series (family or subgroup) are commonly used in soil surveys of national forestland or wilderness areas. Soil series are the lowest level of Soil Taxonomy and are described in the official series descriptions (OSDs) as modal profiles with a set of ranges for the observed morphology and laboratory measurements. The component STU in a mapped SMU varies in the observed field properties from the OSD modal description but usually fits within a soil series range. The observed field properties of soil component units are utilized for developing a set of interpretations for SSURGO polygon map units. These polygons are available from the NRCS as vector GIS layers and in a convenient format on a geographic background as SoilWeb .

The SMUs of the source maps are mappable landscape elements at the survey design scale. These almost always have multiple component STUs, with reported estimated proportion and geomorphic arrangement within the SMU when possible. However, the locations of the STU within the SMU are not mapped due to the design scale. The STUs are linked to database tables of representative or synthetic soil profiles, with field and laboratory measurements of multiple soil properties and interpretations for soil use. To obtain values for soil properties in a gNATSGO or gSSURGO grid cell, properties of the components of the corresponding SMU are combined by area-weighted averaging. To obtain values at coarser resolutions, weighted average properties of groups of grid cells are upscaled by averaging.

There are inherent problems with this product. First, since traditional surveys were carried out over a long time period, series names and mapping concepts may differ between adjacent survey areas. Thus, SSURGO SMU delineations and linked tabular data represent a progressive data collection and correlation effort spanning nearly 100 years. Therefore, there exist many soil survey vintages, each a snapshot in time, tied to specific land use assumptions and technological limitations. Systematic, continuous updates to the entire SSURGO database have been made since 2013 and are ongoing. Second, the transfer from unrectified photos to topographic base and the edge matching between survey areas has not always been flawless, and in addition, polygons may have been incorrectly drawn on the original survey (Fig. S1 in the Supplement). Thus, we cannot take these primary polygon maps as a completely reliable georeference.

The second kind of product is represented by POLARIS soil properties hereafter PSP, which is the result of harmonizing diverse SSURGO and STATSGO2 polygon data with the DSMART algorithm to produce a probabilistic raster soil class or component map (30 m grid resolution) and then extract property information from gSSURGO or gNATSGO grid cells (representing polygons) aggregated by component name. Despite the source data, this is not an NRCS product and was developed independently of the NRCS.

There are two products representing the third kind of product, i.e. one for the world and one for the continental USA only. This allows us to compare globally and nationally consistent products. The global product is SoilGrids v2.0 hereafter SG2;, a further development of SoilGrids1km and SoilGrids250m . This uses a global point dataset and environmental covariates that cover the entire world (except the high Arctic and Antarctica) and global models. It does not use any information derived from SSURGO or STATSGO map units. Its training points are extracted from the freely shareable World Soil Information Service (WoSIS) point dataset from ISRIC–World Soil Information . These include all profiles in the National Soil Survey Center (NSSC) Laboratory Characterization Database. The freely shareable WoSIS points are augmented by several datasets included in WoSIS that cannot be published externally due to restrictions by the original data providers to ISRIC but which can be used in mapping. In total, ≈240000 profiles were used in model building.

The continental product is the Soil Properties and Class 100 m Grids of the United States hereafter SPCG;, which followed the methodology of , with the addition of USA-specific covariates, notably parent material and drainage classes extracted from SSURGO or STATSGO2 map units, and only used the CONUS extent of environmental covariates in model building. SPCG is similar to SG2 in that it is primarily based on point observations, but it has a richer source of these than SG2, i.e. the NSSC Laboratory Characterization Database (34183 pedons comprising 213499 horizons), the National Soil Information System (NASIS), and the Rapid Carbon Assessment (RaCA) dataset (31215 pedons); the latter applies only for organic C, total N, and bulk density. It also uses SSURGO map units to derive parent material (87) and drainage (4) classes as CONUS-specific covariates.

gNATSGO and gSSURGO are based on traditional soil surveys, mostly on unrectified air photo bases until the late 1990s. The many individual survey areas prior to this time have been partially homogenized during a process of digitization and recompilation onto topographic or orthophoto bases during the 1990s and are provided as the polygon SSURGO map. In the early 2000s, for new surveys and updates, a transition was made to on-screen digitization over orthophotos. Field methods are described in successive editions of the Soil Survey Manual and the field book for describing and sampling soils . Mapping is based on conceptual models of soil–landscape relations developed in each survey area and confirmed by purposive auger and full profile descriptions to characterize the map unit composition. Component concepts are refined with any available laboratory characterization data, with (limited) new laboratory characterization performed as needed. Thus, SSURGO provides a local model of soil–landscape relations, developed in each area from the most significant soil-forming factors relevant to that area. SSURGO is progressively updated by field inspection and correlation, as problems are identified by soil surveyors or map users. Since SSURGO has been compiled from diverse surveys over many years, in some areas there are artefacts of that survey process (Fig. S2).

The three DSM products (SG2, PSP, and SPCG) use a large number of gridded GIS coverages as environmental covariates in their predictive models. These represent soil-forming factors and include climate, ecology, geology, land use/cover, terrain, vegetation, and hydrography (Sect. S4). PSP also uses coarse-resolution (≈2 km) estimates of U, Th, and K γ-ray decay products to represent the suspected variation in parent material kind and origin.

PSP uses the DSMART disaggregation algorithm to predict the most probable component (STU), along with their probability of occurrence, at each 30 m resolution grid cell, and from the modal soil properties of the component, a probability-weighted aggregation. Disaggregation is the process of examining a coarser-resolution gridded or smaller-scale polygon product, which is known to have multiple STU, and identify the locations at a finer grid resolution where these components would be found should the original survey have been made at larger scale. This depends on fine-scale covariates that, in theory, relate to the STU within an SMU. It attempts to deal with the problems caused by multiple surveys over time, inconsistencies among mappers, and poor georeference of SMU boundaries by sampling out of mapped SMU polygons according to declared proportions of map unit components (STU) and using these as pseudo-observations to train DSM models of STU occurrence. PSP does not use any point observations; rather, it samples pseudo-points from gSSURGO or gNATSGO and uses these as training points for the DSMART disaggregation algorithm (see below). The model is trained in overlapping tiles, each containing some set of SSURGO primary surveys, and using covariates covering just the tile. Thus, each POLARIS tile is derived from a local model in two senses. PSP provides a fine-scale map equivalent to ≈1:3000 design scale, i.e. from 16 to 64 times finer resolution than the original 1:12000 to 1:24000 surveys included in SSURGO. An obvious question is whether it is possible to map at this resolution from the SSURGO source, even with the fine-resolution covariates used by DSMART, because of the probabilistic nature of selecting pseudo-points to match with components (STU).

The other two methods are representative of the dominant DSM method as implemented, with some differences in detail, in many countries and for many properties e.g..

SG2 uses random forests implemented in the ranger R package, with prior covariate selection by recursive feature elimination and model tuning by the cross-validation of model hyperparameters (number of covariates at each tree split and number of trees in the forest). The model is trained for the whole world, not per country or region; thus, it is a global model. This is based on the homosoil concept for which identical environmental conditions anywhere in the world should result in the same soils. Its use in DSM assumes that all soil-forming factors are fully specified (i.e. over their whole range and with all their possible interactions) in the model and training set. Due to the uneven distribution of training points in covariate space, and portions of covariate space with no observations, this ideal situation is not met. An obvious question is whether or not the additional information from outside the CONUS leads to an improved model for this region.

SPCG is an extension of the original SoilGrids approach but uses an ensemble of two tree-based machine learning methods, namely random forests (as in the original SoilGrids) and gradient boosting. The model is trained for the CONUS and not per region; thus, it is a reduced version of the homosoil concept. It is a global model in the sense of “use all information over a wide area”, although this is not the entire globe, as in SG2.

About 90 % of gSSURGO is derived from polygon maps with a design scale (1:12000 to 1:24000, depending on the original survey) which corresponds to MLDs of 1.44 to 2.3 ha (1:24000) or 0.38 to 0.575 ha (1:12000) polygons, depending on the definition of the MLD (see above). These correspond to single grid cell resolutions of 240 to 384 m (1:24000) or 60 to 96 m (1:12000). gNATSGO includes some areas surveyed at a smaller scale (1:250000). gNATSGO is delivered as gridded coverages at 30 m or 90 m horizontal resolution on an Albers equal area projection covering the CONUS, with standard parallels at 29.5 and 45.5∘ N and the central meridian at -96∘ E on the NAD83 datum, which uses the GRS80 ellipsoid. We have used the 30 m resolution product. gSSURGO is delivered as gridded coverages at 10 m or 30 m horizontal resolution on the same CONUS projection. We have used the 30 m resolution product. Property information is provided per horizon or layer, each with depth limits. Thus, to produce a prediction for a depth interval, these must be aggregated by the depth-weighted average by thickness across the depth interval. PSP predicts at 1 arcsec of longitude and latitude resolution, i.e. 0.0002777778∘ on the WGS84 datum, equivalent to ≈ 32 m latitude, and proportionally smaller longitude depending on latitude. Depth intervals are the standards specified by GlobalSoilMap. SPCG predicts at 100 m resolution for seven point depths (0, 5, 15, 30, 60, 100, and 200 cm) in the same projection as gNATSGO and gSSURGO. Predictions are means of a depth interval. SG2 predicts at 250 m resolution for the standard depth intervals specified by GlobalSoilMap on an equal area Interrupted Goode Homolosine (IGH) projection on the WGS84 datum . Depth interval predictions are in fact point predictions at the centre of the depth interval, considered to represent that interval. Section S3 explains how these products are accessed and made compatible for comparison at regional and local scales.

SG2 and PSP predict the 5 % and 95 % quantiles of the distribution of predictions. SG2 uses quantile regression forests QRFs;, whereas PSP's uncertainty estimates are based on the property data available for each STU predicted by POLARIS. The profile property data are used to create a depth-harmonized profile with uncertainty for each standard depth interval.

These uncertainty limits are specified by the GlobalSoilMap consortium , defined as “the 90 % Prediction Interval (PI) which reports the range of values within which the true value is expected to occur 9 times out of 10 … there is no assumption that this prediction interval is necessarily symmetric around the predicted value” .

gNATSGO and gSSURGO provide representative, upper, and lower limit values of each property of an STU per horizon or layer. The National Soil Survey Handbook, § 618.2 , explains that the representative value approximates the median but that the quantiles corresponding to the low and high values can be adjusted to the percentiles which best show the spread of the property within an STU. If there are sufficient laboratory data of the sampled profiles of the STU in the National Soil Information System NASIS;, then these are used as the basis for establishing the range. In all cases, expert opinion is used to adjust these to represent the range that a map user can expect to find in the field. Thus these are not directly comparable to the results of QRF but do give some idea of how the field mappers, supported by laboratory observations, conceive of the spread of a property. Note that none of these assessments implies a parametric probability distribution but rather the ranges of selected quantiles only. discuss how these estimates can be derived for USA products following the GlobalSoilMap.net specifications.

As pointed out by , “[t]he user community requires training in, and experience with, the new digital soil map products, especially about the use of uncertainties”. It would be hoped that the uncertainties computed by different methods would be similar.

Qualitative methods for comparing maps rely on expert judgement to identify known soil geographic patterns and evaluate to what extent they are represented on the gridded maps. The maps are displayed side by side, along with a map of their pairwise differences. Areas of disagreement are identified and discussed.

The DSM product can be evaluated at selected known points, typically from the field observation of test areas. The following questions are posed: is the correct soil type or property predicted? And, if not, is the error a reasonable approximation? More interesting are the patterns in the DSM product. These can be compared to patterns used in the mental model of traditional soil survey such as, for example, toposequences and sequences of contrasting parent material.

In both cases (points and patterns), the evaluator may be able to infer which DSM covariates would be needed to improve the map.

Numerical methods for comparing gridded maps as a whole include (1) MD, which is the mean difference (also known as the bias), i.e. the average disagreement between maps, (2) RMSD, which is root mean squared difference, and (3) RMSD adjusted for MD, i.e. the RMSD after subtracting the bias from each prediction. These take the first-listed map as reference and the second as the map to evaluate. They can be normalized by the number of grid cells or total area. In addition, all maps can be compared by their Pearson (linear) correlations. These methods are of limited interpretive value. Their main use is to characterize the bias (MD) over the entire map; they do not reveal where any discrepancies occur. For example, there can be no bias overall but a large difference in the amount and values of higher and lower differences will be seen. This will be reflected in the RMSD although not shown on a map.

Soil properties are usually spatially correlated; we expect similar values of properties in nearby grid cells. The degree of local spatial continuity can be assessed by the variogram computed over local neighbourhoods of the gridded map. We computed and modelled the variogram within a local neighbourhood, and it automatically fit with an exponential model, using the fit.variogram function of the gstat R package . The spatial structure is characterized by the range, proportional nugget, and structural sill of the fitted variogram model. The range shows the radius over which the selected property has spatial correlation. The proportional nugget shows the variability at the prediction point at the centre of a grid cell at a scale shorter than the grid spacing. The structural sill shows the overall variability within the range. These metrics show differences in spatial continuity (range), total variability (total sill) and short-range unexplained variability (proportional nugget) between maps.

Numerical methods for comparing patterns include (1) the V-measure method Sect. ;, implemented in the sabre (Spatial Association Between REgionalizations) R package , and (2) landscape-level metrics Sect. ;, as used in ecology and derived from the FRAGSTATS computer program and implemented in the landscapemetrics R package . These include Shannon diversity and evenness, landscape shape index, and fractal dimension. Although the ecological relevance of FRAGSTATS metrics have been criticized , here we use them to characterize spatial patterns of soil properties and not as inputs to landscape ecology models. Most of the metrics used here have also been used by in a study of urban pedodiversity.

These methods must be applied to classified maps, so the continuous soil property maps must first be classified into ranges before analysis. Different choices of class limits and widths will result in different values of these measures. A somewhat objective method to choose classes is histogram equalization. The analyst determines the number of classes, and equal numbers of grid cells are in each class. To compare maps, the combined values of all maps are used to construct the histogram. For the V measure, the gridded maps must be polygonized.

Regional patterns are at the scale of regional trends such as lithologic units, elevation zones in mountains, and repeating patterns (e.g. basin and range and ridge and valley). The gNATSGO maps are taken as the reference, although we are well aware that they may not always correspond to ground truth. We then comment on the differences and speculate on the causes, based on our knowledge of the DSM procedures used to make each product and the nature of the soil landscape.

Local patterns are at the scale of geomorphic features such as hillslope catenas, fluvial terraces, outwash fans, valley trains, and drumlin fields. This evaluation occurs within the test area for the regional patterns but examines a smaller area with a distinctive soil–landscape pattern. The gSSURGO maps are taken as the reference. This was evaluated by two methods, as follows.

A visual inspection of a DSM product over the landscape can be useful to identify anomalies and the degree to which the DSM product captures landscape features. These are over small areas where the soil–landscape relation is known to the evaluator. This cannot be part of a systematic evaluation but can reveal areas of concern or agreement.

As an example, Fig.  shows SSURGO map units, from a 1965 1:20 k design scale soil survey, with a minimum legible delineation 1.6 ha and draped over a ground overlay of pH (0–5 cm) from SG2, produced by the SoilWeb streaming coverage in Google Earth Pro, with a point query showing the SSURGO map unit composition (Fig. ). The map unit is described by its constituent soil series and their estimated proportions. Each series can then be queried for its OSD , which gives a typical profile, a range of properties, and a link to lab data for the series. In this case, the pattern of properties as predicted by SG2 somewhat follows the map unit delineations but at a much coarser resolution. This is especially evident at the transition from the end moraine (map units beginning with H) and the steep slopes with thin till from the local bedrock (map unit LoF).

Table  shows the statistical differences between gNATSGO (reference) and the DSM products. All DSM products underpredict topsoil pH with respect to gNATSGO by about 0.38–0.48 pH units. The RMSD is substantial also, of the order of 0.49–0.67 pH units, and is somewhat less than this when corrected for bias.

Figure  shows whole-map histograms. PSP has a bimodal distribution, and predicts few pH values around pH 5.8. This was unexpected, since this value is well-represented in the gNATSGO map. It may be an artefact of a covariate that is influential over a wider area than this tile and results in two regional distributions from contrasting elevation or climate zones. The other distributions are fairly symmetric, although SG2 and SPCG are more even than gNATSGO, which is strongly concentrated near pH 6.2. This shows the smoothing effect of the machine learning models.

Figure  shows the pairwise Pearson correlations between the products. The products are overall well correlated. SG2 and SPCG are very closely correlated, since they use similar mapping methods, despite the additional covariates used by SPCG. PSP and gNATSGO are also closely correlated. These correlations do not account for bias. They do, however, show that the maps are similar in their overall pattern as they are evaluated per grid cell.

Figure  shows gNATSGO (reference) along with the predictions of pH of the PSP products. Figure  shows these as difference maps. These figures reveal substantial differences between products. The most obvious is in the detail of the spatial pattern. Despite having been upscaled to s regional resolution, gNATSGO shows finer detail than the other products, especially when compared to PSP.

These figures also show the spatial distribution of the bias compared to gNATSGO (as reference). SG2 and SPCG underpredict pH in the higher hills in the NE portion of the map and in the glaciolacustrine sediments along the lakeshores. The disagreement along the lakeshores is because SG2 and SPCG do not use a surficial geology map, which would be especially useful in recently glaciated areas such as this. The disagreement in the higher hills seems to be a direct result of elevation. This is not because of extrapolation in feature space because, at these elevations, SG2 also misses the soils derived from Onondaga limestone glacial till towards the southern end of the till plain. SG2 has no information on parent material and uses global models. SPCG has very similar differences, despite using SSURGO-derived parent material as a covariate.

PSP predictions are closer to gNATSGO than those of SG2 are, which is not surprising since PSP also uses gNATSGO as its primary information source. This product has removed some of the fine variation in gNATSGO. However the disaggregation by DSMART results in some discrepancies with gNATSGO. In particular, the Homer–Tully outwash valley (northeastern side of map) is underpredicted by 1 pH unit, and the surrounding hills are overpredicted by almost as much. Many of the valley trains (southern side of map, running towards the Susquehanna River) are underpredicted. This is likely due to PSP's soil series predictions, which are based on estimated map unit composition and random selection of series locations within map units for DSM calibration.

The 5 %, 50 %, and 95 % prediction quantile maps are shown in Figs.  (SG2) and (PSP). The low, representative, and high values from gNATSGO are shown in Fig. . Each figure has its own stretch. gNATSGO has narrower ranges than the two DSM products and, by design, does not include unrealistic values. SG2 and PSP have unrealistically wide ranges at all locations. In addition, PSP shows a curious feature, i.e. fine patterning at the two extremes that is not present at the median prediction.

Figure  shows the interquartile range (IQR) 5 %–95 % for the two DSM products, along with the low–high range for gNATSGO. SG2 has a fairly consistent IQR, mostly from about 2.5 to 3.5 pH, whereas PSP has a much wider range of uncertainties, mostly from about 1.5 to 4.5 pH, and shows much more spatial pattern. PSP has the widest ranges on the steep valley sides, especially in the Seneca Army Depot at the north interlake area, and the lowest on the broad till plains and through valleys. These are wide ranges and, although they are an honest reflection of the DSM models, should give pause to map users. This suggests that the GlobalSoilMap specifications for uncertainty are unduly pessimistic. Sources for uncertainty assessment (SG2 for training points and global covariates and PSP for mapped soil series and national covariates) and the different machine learning methods lead to greatly different estimates of prediction uncertainty. The gNATSGO low–high range is narrower than the DSM IQR, but these are not comparable because the expert-assigned range is not based on an estimate of a 5 %–95 % IQR.

Figure  shows the differences between the IQR of the DSM products and the low–high range from gNATSGO. Both DSM products have substantially wider ranges than gNATSGO almost everywhere; however, the pattern of differences is not similar. For example, the difference with SG250 is much larger in the north of the study area, whereas PSP has the larger differences in the southern hills.

The local variograms and their fitted exponential models are shown in Fig. . Table  shows their statistics.

gNATSGO has the shortest effective range. This indicates a fine-scale structure at 250 m resolution, which is of the same order as the minimum legible delineation (MLD) as a grid cell (see the Introduction). The mappers who defined the boundaries between soil classes (and thus representative property values) were able to divide the landscape at this high spatial frequency, if appropriate to the soil pattern. The DSM products have longer ranges, likely due to the longer-range spatial continuity in many of the covariates. PSP has a longer range and is lower sill than the gNATSGO from which it is derived due to the harmonization inherent in the DSMART algorithm. It has the highest proportional nugget, due to DSMART randomly assigning pixels within a gNATSGO map unit to its constituents, so that the neighbouring pixels may be contrasting at the shortest separation. The very low proportional nuggets of the other products are due to the coarse resolution.

Figure  shows the topsoil pH classified into eight histogram-equalized classes in a 0.2×0.2∘ subtile. The class limits are approximately 5.01, 5.14, 5.27, 5.40, 5.54, 5.71, and 6.02 pH, with the extreme values of 4.52 and 6.96 pH. The maps show obvious spatial differences in class distribution. gNATSGO shows more areas in the highest pH class than the DSM products, which is consistent with the results from continuous property maps. The pattern of gNATSGO is the coarsest because the classified values come from minimum-area polygons, whereas the DSM products predict per grid cell. PSP shows the finest spatial pattern because of its disaggregation algorithm that randomly divides gNATSGO polygons according to component proportion. If these components are in different pH classes, there will be a fine-scale pattern within the original polygon. This is clearly the case in the large gNATSGO polygon in the northeastern portion of the map (Connecticut Hill). In this example, SPCG shows large homogeneous areas of the lowest pH class, covering the highest hills, whereas SG2 presents a more nuanced view.

Table  shows the statistics from several V-measure comparisons, based on the histogram-equalized class maps. Only SG2 and SPCG have somewhat comparable patterns. gNATSGO is considerably different from the DSM products because of its derivation from minimum-area polygons.

Figure  shows the inhomogeneity and incompleteness of the SG2 pH class map (the second map for the V measure), with respect to the gNATSGO pH class map (the reference map). These values are the inverse of the composite values of Table  because the very low values in the table correspond to high values in the figure. In the homogeneity map, the blue polygons are the most homogeneous areas of the SG2 map, i.e. where an SG2 polygon has the most homogeneous set of gNATSGO classified values and thus comes closest to the reference. In the completeness map, the blue polygons are the most complete areas of the SG2 map, i.e. where the gNATSGO reference map has the most homogeneous set of SG2 classified values. The two maps have no areas with similar patterns.

A contrasting result is shown in Fig. , which compares the SG2 pH class map with respect to the SPCG map. These maps were made with similar methods and at the same resolution. The inhomogeneity and incompleteness are much lower than for the previous map, showing that the pattern of these two classified maps are fairly similar.

Table  shows the statistics from the landscape metrics calculations. The mean fractal dimensions are almost identical. There is quite some range of aggregations, with SPCG being the most aggregated, i.e. the least complex. PSP has the most complex landscape shape due to its fine-scale disaggregation of gSSURGO polygons. The Shannon diversity indices are highest for SG2, indicating that there is the most even areal division into classes. This may be an artefact of the histogram equalization.

Table  shows the Jensen–Shannon distance between co-occurrence vectors of the four products. The co-occurrence patterns of SG2 is quite similar to that of the other DSM products, whereas gNATSGO is quite different to PSP and SPCG and somewhat different to SG2. This shows that, given this histogram equalization and for the selected property and depth interval, none of the DSM products match the pattern from the traditional soil survey well.

Here we use the silt concentration, as it reveals stronger qualitative discrepancies than pH in this test area. Figure  shows the silt concentration of the 0–5 cm layer for the (top) gridded SSURGO overlain on the original polygons from which it was derived and (bottom) the disaggregated PSP grid cells in a hilly landscape near Caroline, NY.

The gSSURGO product follows the SSURGO lines exactly. Some of the sharp boundary lines do correspond with abrupt transitions on the ground, for example, where the steep hillsides are buried by fan alluvium. But others are not, for example, on the hilltops. These differences are because the predicted silt concentrations are taken from the official series descriptions. PSP follows the map unit lines fairly well but is much finer grained and each 30 m pixel is separately predicted. This results in some smoothing of the abrupt boundary lines from gSSURGO on the hilltops. However, within some SSURGO map units, PSP predicts some differences in the topsoil silt concentration. These are map units with contrasting components, which PSP attempts to disaggregate according to their correlation with covariates. For the most part, these do not seem to be related to terrain or land use.

For example, Fig.  shows the detail of the Holly–Papakting map unit within this PSP window. This map unit has two contrasting soils in similar proportions, i.e. a mineral alluvial soil (Holly series) and an organic soil (Papakting series), and the second has a much lower silt concentration.

It is difficult to see the reason for the pattern within this map unit. PSP has placed the component series in their proper proportions but not according to any apparent landscape feature or covariate.

Another example from this same area is shown in Sect. S5.

To see the fine differences at this high resolution, we consider a 0.15×0.15∘ subtile, with the lower-right corner at -76.30∘ E, 42.45∘ N, and evaluate pH as in the regional assessment (Sect. ).

Table  shows the statistical differences between gSSURGO (reference) and the DSM products, along with the predictions of pH. Figure  shows the pairwise Pearson correlations between the maps. These results are comparable to those for the full tile at regional resolution because both SG2 and PSP underpredict pH by about 0.35–0.45 pH. The correlations are fairly strong between PSP and gSSURGO and between SG2 and PSP but weak between SG2 and gSSURGO.

Figure  shows gSSURGO (reference) along with the predictions of pH by the PSP products. Figure  shows these as difference maps. Clearly, gSSURGO has overall higher values than the other two products and, despite the fine resolution, has in general large areas of identical values. The differentiation between map units follows sharp boundaries, even within a single landscape (e.g. the plateau towards the south of the map), and this is likely an artefact of relying on the representative profiles in the official series descriptions for property values. PSP has a finer pattern, due to disaggregation, and shows a smoother local pattern without the sharp boundaries between map units within a landscape. PSP shows large areas of low pH. SG2 does not follow the landscape lines well, especially the sharp boundaries between uplands and valleys, and predicts very low pH (≈4.5) on the plateau. It is difficult to recognize local landscape units in this global product.