SoilGrids250m: Global gridded soil information based on machine learning

This paper describes the technical development and accuracy assessment of the June 2016 update of SoilGrids at 250m resolution. The system provides global predictions of standard numeric soil properties, including organic carbon, bulk density, cation exchange capacity, pH, soil texture fractions, and coarse fragments, at seven standard depths from 0 to 200 cm, along with depth to bedrock and soil class distributions under the WRB and USDA systems, roughly 280 raster layers in total. Predictions were based on about 150,000 soil profiles for training and a stack of 158 remote-sensing-based covariates, primarily derived from MODIS land products, SRTM DEM derivatives, climatic images, and global landform and lithology maps, fitted with an ensemble of machine learning methods (random forest, gradient boosting, and multinomial logistic regression) implemented in the R packages ranger, xgboost, nnet, and caret.

Ten-fold cross-validation shows the ensemble models explain between 56% (coarse fragments) and 83% (pH) of variation, averaging 61% overall. Relative accuracy improvements over the previous 1km-resolution version range from 60% to 230%, attributed to using machine learning instead of linear regression, substantial investment in finer-resolution covariate layers, and inclusion of additional soil profiles. The maps are released under the Open Data Base License, and the authors outline future directions such as incorporating input uncertainties, deriving per-pixel posterior distributions, and multiscale merging with local products.