SoilGrids250m: Global gridded soil information based on machine learning

This paper documents the technical development and accuracy assessment of the June 2016 update of SoilGrids, a system that produces global gridded soil information at 250m spatial resolution. SoilGrids predicts standard numeric soil properties, including organic carbon, bulk density, cation exchange capacity, pH, soil texture fractions, and coarse fragments, at seven standard depths down to 200cm, and additionally predicts depth to bedrock and the distribution of soil classes under the WRB and USDA systems, amounting to roughly 280 raster layers. The predictions were fit from about 150,000 training soil profiles and a stack of 158 remote-sensing-based covariates derived largely from MODIS products, SRTM elevation derivatives, climate images, and landform and lithology maps, using an ensemble of machine learning methods: random forest, gradient boosting, and multinomial logistic regression.

Ten-fold cross-validation showed the ensemble models explain between 56% of the variation for coarse fragments and 83% for pH, averaging 61% overall. Relative accuracy improved by 60 to 230% compared with the previous 1km-resolution SoilGrids, gains attributed to using machine learning instead of linear regression, investing in finer-resolution covariate layers, and adding more soil profiles. Released under an open database license, SoilGrids250m became a widely used global soil dataset, with future directions including uncertainty quantification and multiscale merging with local products.