Figure 1.
Spatial resolution and temporal coverage/publication time of some widely used global environmental data layers (global soil layers have been highlighted): GLWD — Global Lakes and Wetlands Database, HWSD — Harmonized World Soil Database, MOD12C1 — MODIS Land Cover Type Yearly L3, MOD13C2 — Vegetation Indices Monthly L3, CHLO/SST — MODIS Aqua Level-3 annual Chlorophyll/mid-IR Sea Surface Temperature, FRA — Forest Resources Assessment, GPW — Gridded Population of the World, DMSP-OLS — Nighttime Lights Time Series, GlobCov — Land Cover classes based on the MERIS FR images, GADM — Global Administrative Areas, TanDEM-X — Germany's topographic radar mission.
Key agenda setters in the terms of production and dissemination of remote sensing and thematic environmental layers at the beginning of the 21st century include: NASA's MODIS (Moderate-resolution Imaging Spectroradiometer) and Landsat products — in terms of thematic content and usability [6]–[8], and Germany's TanDEM-X new global 12 m resolution DEM with ±2 m vertical accuracy [9]. Based on information retrieved on February 15th 2014. was produced using the Global Soil Information Facilities (GSIF), which was recently developed at ISRIC as a framework and platform to support widespread, open collaboration in the assembly, collation and production of global soil information.
Figure 2.
World distribution of soil profiles used to generate the SoilGrids1km product (about 110,000 points).
Courtesy of various national and international agencies (see: Acknowledgments).
Figure 3.
Examples of input layers used to generate SoilGrids1km: (a) long-term day-time MODIS land surface temperature, (b) percent cover Chernozems (based on the HWSD data set), and (c) global soil mask map.
The spatial prediction domain of SoilGrids1km are the areas with vegetation cover and urban areas, while bare soil areas have been masked out. See text for more explanation.
Figure 4.
Standard stratification and designation of a soil profile: (left) soil horizons, solum thickness and depth to bedrock (‘R’ layer), and (right) six standard depths used in the GlobalSoilMap project [3].
Figure 5.
Individual soil profile from the ISRIC soil monolith collection (a) and globally fitted regression model for predicting soil organic carbon using depth only (b).
The individual profile horizons are described by Mokma and Buurman [39]. Adjusted R-square for the model on the right is 0.363. Open circles show measured values for the profile on the left.
Figure 6.
Soil organic carbon stock calculus scheme.
Example of how total soil organic carbon stock (OCS) and its propagated error can be estimated for a given volume of soil using organic carbon content (ORC), bulk density (BLD), thickness of horizon (HOT), and percentage of coarse fragments (CRF). See text for more detail.
Table 1.
Mapping performance of SoilGrids1km — amount of variation explained (from 100%) or purity/kappa for categorical variables — for eight targeted soil properties and two soil classes distributed via SoilGrids1km.
Figure 7.
Example of SoilGrids1km layers: (A) soil organic carbon content in permille, and (B) soil pH for the topsoil (0–5 centimetres).
Boxplots show the sampled distribution of the soil property based on the present compilation of global soil profile data.
Figure 8.
SoilGrids1km-derived soil-depth curves for the profile shown in Figure 5.
Location of the profile: 6.3831°E, 50.479167°N. The shaded background indicates the 90% prediction interval for each depth. ORCDRC = soil organic carbon content in permilles; PHIHOX = soil pH in water suspension. See also Table 1.
Figure 9.
Spatial predictions of WRB soil groups for SoilGrids1km (left) and HWSD data set representing conventional soil maps (right).
A zoom in on North of Italy. White pixels indicate missing values.
Figure 10.
Predicted global distribution of the soil organic carbon stock in tonnes per ha for 0–200 centimetres.
Total soil organic carbon stock (here displayed on a log-scale) was estimated as a sum of soil organic carbon stocks for six standard depths and adjusted for the depth to bedrock. Projected in the Sinusoidal equal area projection to give a realistic presentation of areas. Vast deserts (e.g. Sahara or Gobi) can be assumed to contain close to zero organic carbon stock. See also Figure 11.
Figure 11.
Lower and upper confidence limits (90% probability) of estimated soil organic carbon stock (tonnes per ha) for standard depths 0–30 and 30–60 centimeters for the same area as shown in Figure 9.
Derived using the procedure explained in Figure 6.
Figure 12.
Accessing SoilGrids1km from the SoilInfo app for mobile devices.
SoilInfo app is available for download via http://soilinfo.isric.org.
Figure 13.
Projected evolution of SoilGrids in the years to come.
We anticipate that the main drivers of success of SoilGrids will be use of machine learning methods for model fitting, development of spatio-temporal geostatistical models, use of new sources of field and remote sensing data and use of faster and more powerful computing capacities. Amount of variation explained by these models will eventually reach a ‘natural limit’ (short-range variation that cannot be explained using spatial prediction models), until there is a technological jump in soil remote sensing technology e.g. ground penetrating scanners.