Three national U.S. soil point datasets — NCSS Characterization Database, the National Soil Information System (NASIS), and the Rapid Carbon Assessment (RaCA) datasets — were combined with remote sensing images and detailed conventional soil polygon maps, and used to generate complete-coverage gridded predictions of soil properties (percent organic carbon, total nitrogen, bulk density, pH, and percent sand and clay) and classes (taxonomic great group and particle size in the control section) for the Conterminous U.S. Soil covariate layers included: DEM-based derivatives, long-term MODIS EVI seasonal images, MODIS cloud fractions, and temperature images per month, PRISM climatic datasets of precipitation, temperature, and vapor pressure deficit, and bioclimatic indicators, Landsat (cloud free) NIR, SWIR bands, gamma radiometric images, geological surface classes, land cover classes, globally produced predictions of soil properties (SoilGrids250m), and the SSURGO parent material and drainage maps. The soil property and class models were built within a high-performance computing system using parallelized random forest and gradient boosting algorithms. Predictions were generated at 100 meter spatial resolution for 7 standard soil depths (0, 5, 15, 30, 60, 100 and 200 cm) for soil properties and as probabilities per soil class.
Items in this Work