Table 1.
List of soil properties of interest.
AfSP = Africa Soil Profiles database, AfSS = AfSIS Sentinel Site database. Range was derived as the symmetric 99% quantile range based on observed data. Number of depths column indicates number of output prediction depths e.g.: 6 depths (0–5 cm, 5–15 cm, 15–30 cm, 30–60 cm, 60–100 cm and 100–200 cm) or 2 depths (0–20 cm, 20–50 cm).
Fig 1.
Temporal and soil-depth coverage of the Africa Soil Profiles and AfSIS Sentinel Site databases.
See also Table 1.
Fig 2.
Distribution of soil samples in Africa used to build spatial predictive models.
(left) legacy soil profile observations (Africa Soil Profiles database) showing ca. 18.5 thousand locations [40], and (right) AfSIS Sentinel Sites showing ca. 9.5 thousand locations, but which are clustered at 60 sentinel sites [41]. Zoom-in on the example area (100 by 400 km) shown further in Figs 6 and 9. Coordinates in the Lambert Azimuthal Equal Area projection (WGS84 ellipsoid) with latitude at projection center = 5°, longitude at projection center = 20°.
Fig 3.
Density of soil observations in Africa showing distinct spatial clustering and an example of residual variography.
(A) relative density of soil observations in Africa determined using a kernel smoother displayed in log-scale (input locations shown in Fig 2), (B) 3D sampling locations scheme, (C) example of an exponential variogram fitted for soil organic carbon residuals. In this case the maximum distance of interest for kriging has been set at 60 km.
Table 2.
Summary statistics for mapping accuracy assessed using 5–fold cross-validation.
ME is the mean error, RMSE the root mean squared error, sg1km are the SoilGrids1km map, rf represents random forest model predictions and lm the linear model predictions (trend model predictions only). The t-test evaluates the difference between the mean errors of the rf and lm models with alternative hypothesis that the difference is greater than 0. The F-test evaluates the ratio between the residual variances of the rf and lm models with alternative hypothesis that the difference is greater than 1. Σ% indicates amount of variation explained by the prediction models and ΔRMSE% indicates improvement in RMSE in percentages compared to the lm model. The ‘⋆⋆⋆’ indicates significance at the 99% probability level. For all soil properties except PHIHOX, SNDPPT, SLTPPT, CLYPPT and BLD, the Σ%, the t-test, and the F-test have been calculated in log-transformed space. SP-SS are the predictions at Sentinel Sites produced using models fitted from AfSP data, SS-SP are the predictions at legacy soil profiles produced using AfSS data. See Table 1 for more details.
Fig 4.
Scatter plots of 5-fold cross-validation errors for soil organic carbon (ORCDRC), pH in water (PHIHOX), bulk density (BLD), soil texture fractions (SNDPPT, SLTPPT and CLYPPT), Cation Exchange Capacity (CEC), total nitrogen (NTO), exchangeable bases (EXBX), total Aluminium (ALUM3S), exchangeable acidity (EACKCL), exchangeable Potassium (EXKX), exchangeable Calcium (ECAX) and exchangeable Magnesium (EMGX).
See also Table 2.
Fig 5.
Importance plots for soil organic carbon and exchangeable Magnesium.
derived using the varImpPlot function available in the randomForest package [45]. m_TAXOUSDA_x are the predicted SoilGrids1km Soil Taxonomy suborders class probabilities, af_DEMSRE5a, af_SLPSRE5a, af_TWISRE5a are the elevation, slope and SAGA GIS Topographic Wetness Index derived at 250 m resolution, af_M13EVIAxx and af_M13RB7Axx are the long-term (2001–2013) standardized values of MODIS EVI and mid-infra red (band 7) products for months January to December, and af_PC1–4EVI5a are the first four principal components derived from annual MODIS EVI images (2001–2013).
Fig 6.
Predicted exchangeable Mg, Ca and K (in cmol+/kg) and Al concentration (ppm) using random forests RK model: zoom-in on the example area from Fig 2.
Vector lines data source: OpenStreetMap.
Fig 7.
Soil organic carbon content in permilles predicted using 3D random forests RK at six standard depths.
White pixels indicate excluded areas (water bodies and deserts).
Fig 8.
Spatial patterns of predicted total nitrogen (in permilles), exchangeable Mg (cmol+/kg) and CEC (cmol+/kg) for topsoil for Kenya.
Legends were set using equally spaced quantiles. White pixels indicate excluded areas (water bodies and deserts). Each soil property is modelled independently and can thus show quite different spatial patterns.
Fig 9.
Predicted soil organic carbon content in permilles (depth: 0–5 cm) with a zoom-in on the area around the town of Arusha (Tanzania).
(left) original SoilGrids1km layer at 1 km resolution vs (right) downscaled spatial predictions at 250 m resolution. Vector lines data source: OpenStreetMap.
Fig 10.
Spatial prediction scheme used to produce AfSoilGrids250m data.
Spatial predictions in the case of an automated soil mapping system can be continuously updated by adding new soil field observations and new covariates.
Fig 11.
High values of exchangeable bases in Africa coincide with the predicted distribution of Alfisols and Mollisols.
distribution of Aquolls from SoilGrids1km [21] (right) and the predicted exchangeable Mg (left). See also Fig 5.
Fig 12.
Global distribution of profiles with observed USDA Soil Taxonomy class.
observations used as calibration data for producing SoilGrids1km predictions [21]. The majority of observations (> 80%) come from the USA National Cooperative Soil Survey Soil Characterization database.