SoilGrids250m

SoilGrids250m provides high-resolution (250 m) global predictions of standard numeric soil properties at seven standard depths (0, 5, 15, 30, 60, 100, and 200 cm) for environmental, agricultural, land-management, and climate research.

Key Features:

• Spatial resolution: 250 m global raster predictions.
• Depth layers: Predictions at seven standard depths (0, 5, 15, 30, 60, 100, and 200 cm).
• Predicted properties: Organic carbon, bulk density, cation exchange capacity (CEC), pH, soil texture fractions, coarse fragments, depth to bedrock, and WRB and USDA soil class distributions.
• Raster outputs: Approximately 280 raster layers covering properties and depths.
• Training data: Approximately 150,000 soil profiles used for model training.
• Covariate stack: 158 remote-sensing and environmental covariates derived from MODIS land products, SRTM DEM derivatives, climatic images, and global landform and lithology maps.
• Machine learning ensemble: Ensemble modeling including random forest, gradient boosting, and multinomial logistic regression.
• Software: Modeling implemented with R packages ranger, xgboost, nnet, and caret.
• Model validation: Performance assessed using 10-fold cross-validation.
• Reported accuracy: Ensemble models explained 56% (coarse fragments) to 83% (pH) of variance, with an overall average explained variance of 61%.
• Methodological improvements: Adoption of machine learning over linear regression, use of finer-resolution covariates, and inclusion of additional soil profiles.
• Ongoing development: Methods to incorporate input uncertainties and derive per-pixel posterior probability distributions, automation for spatial modeling, and multiscale merging with local or national gridded products toward finer resolutions (potentially up to 50 m).

Scientific Applications:

• Environmental research: Provide soil property maps for ecosystem and biogeochemical modeling.
• Agriculture: Inform soil fertility assessment, carbon stock estimation, and land-use planning.
• Land management: Support soil assessment, restoration planning, and land suitability analyses.
• Climate studies: Supply soil inputs for climate models and carbon cycle assessments.

Methodology:

Models link ~150,000 soil profiles to a stack of 158 remote-sensing and environmental covariates (MODIS land products, SRTM DEM derivatives, climatic images, landform and lithology maps) and fit an ensemble of random forest, gradient boosting, and multinomial logistic regression using R packages ranger, xgboost, nnet, and caret, with performance assessed by 10-fold cross-validation to generate ~280 raster layers at seven depths at 250 m resolution.