Table 1.

Summaries of each of the eight topsoil SOC concentration maps.

Map; reference	Grid resolution (m); spatial coverage	Soil depth layers (cm)	SOC units	Prediction methods
ISRIC-201748	250 (resampled by taking the area-weighted mean of all cells in a 1 km grid; see “Methods” section); global	0, 5 & 15 (converted to a single 0–15 layer using the trapezium rule)	dg kg−1 (divided by 10 to get to g kg−1)	Applied machine learning, including random forest & gradient boosting, to a harmonised global soil observation dataset (WoSIS), using 90% of observations for calibration; 10% for validation. Covariates used for model prediction include (but are not limited to): EVI, night & day-time land surface temperature, land cover, monthly precipitation, lithologic units, and multiple topographic variables
ISRIC-202049	250 (resampled by taking the area-weighted mean of all cells in a 1 km grid; see “Methods” section); global	0–5 & 5–15 (converted to a single 0–15 layer by taking weighted means of the layers)	dg kg−1 (divided by 10 to get to g kg−1)	Same as above, but using: (1) A greater range of soil observations (updates to WoSIS soil database); (2) Improved model calibration & cross-validation; (3) Improved covariate selection & parameterisation; and (4) Prediction uncertainty quantified at the 90% prediction interval. Calibration on 90% of samples; validation on 10%
LUCAS47	500 (resampled by taking the area-weighted mean of all cells in a 1 km grid; see “Methods” section); Europe	0–20	g kg−1	Generalised additive model fitted to 85% of LUCAS 2009 survey points (15% used for validation), using slope, land cover, NPP, latitude & longitude as covariates for model prediction
OCTOP55	1000; Europe	0–30	% SOC (multiplied by 10 to convert to g kg−1)	Applied a pedo-transfer rule to all soil observations from the European Soil Database, with soil type, mean annual accumulated temperature, dominant surface textural class & land cover/use as covariates for model prediction. Validated using SOC data from Italy, England and Wales
CSGB-AIC53	1000; GB	0–15	g kg−1	Conventional upscaling/geo-matching to derive weighted-average SOC for different land units based on various combinations of land cover & parent material attributes. Inter-comparison of Akaike’s Information Criterion to judge model accuracies & to select the best map. Used all CS 2007 observations
CSGB-GAMM51	1000; GB	0–15	g kg−1	Applied a spatial GAMM modelling approach to all CS 2007 points. Covariates used included broad habitat class, soil group, CaCO3 rank, SO4, NH4 & NO3 deposition, 5-year means of seasonal temperature & precipitation. Validation by applying the model to LUCAS 2009 samples
CSGB-KRGS14	1000; GB	0–15	% LOI (multiplied by 5.5 as per28)	Interpolated a map of loss-on-ignition percentages from all CS 2007 sites (mean of 5 random points per square) using ordinary kriging. Sequential Gaussian simulation to estimate map uncertainty
CSGB-MLRF56	1000; GB	0–15	% LOI (multiplied by 5.5 as per28)	Applied a chain modelling approach of first using random forests to predict land cover from climate variables and then soil organic matter content from predicted land cover composition. Used CS 2007 in both modelling steps (80% of samples for model calibration; 20% for validation)