SoilGrids

SoilGrids provides global predictions of numeric soil properties at multiple depths for use in environmental, agricultural, land‑management, and climate analyses.

Key Features:

• Spatial resolution and depths: Global predictions at 250 m spatial resolution for seven standard depths: 0, 5, 15, 30, 60, 100, and 200 cm.
• Soil properties predicted: Numeric predictions include organic carbon content, bulk density, Cation Exchange Capacity (CEC), pH, soil texture fractions, coarse fragments, depth to bedrock, and soil class distributions according to WRB and USDA systems.
• Output layers: Generates approximately 280 raster layers representing predicted properties and classes.
• Training data: Models trained on about 150,000 soil profiles.
• Covariates: Uses a stack of 158 remote sensing‑based covariates primarily from MODIS land products, SRTM DEM derivatives, climatic images, and global maps of landform and lithology.
• Modeling algorithms: Applies ensemble machine learning methods including random forest, gradient boosting, and multinomial logistic regression implemented via R packages ranger, xgboost, nnet, and caret.
• Validation: Model assessment used 10‑fold cross‑validation with explained variance ranging from 56% (coarse fragments) to 83% (pH) and an overall average of 61%.
• Relative improvement: Reports relative accuracy improvements of 60%–230% compared to the previous 1 km resolution version, attributed to machine learning adoption, finer covariate layers, and additional soil profiles.
• Research and development directions: Active developments include incorporating input uncertainties to derive per‑pixel posterior probability distributions, automating spatial modeling for more variables, and multiscale merging with local/national gridded products potentially to 50 m resolution.

Scientific Applications:

• Environmental research: Support for global and regional studies of soil carbon, nutrient status, and soil‑driven ecosystem processes.
• Agriculture: Informing soil fertility assessments, management planning, and input targeting based on predicted soil properties.
• Land management: Enabling mapping of soil constraints such as coarse fragments and depth to bedrock for land use and infrastructure planning.
• Climate studies: Providing spatially explicit inputs for Earth‑system models and carbon budget assessments via organic carbon and bulk density maps.

Methodology:

Models were trained on ~150,000 soil profiles combined with 158 remote sensing and thematic covariates (MODIS, SRTM DEM derivatives, climatic images, landform and lithology maps) using ensemble machine learning (random forest, gradient boosting, multinomial logistic regression) implemented with R packages ranger, xgboost, nnet, and caret, and evaluated with 10‑fold cross‑validation.