Data from: High-resolution soil total phosphorus mapping for the conterminous USA using machine learning

Accurate estimates of soil total phosphorus (TP) concentrations are
essential for sustainable nutrient management, food security, and water
quality protection. This study predicts and maps the spatial distribution
of TP in the top 5 cm and C horizon of soils across the conterminous USA
(CONUS) using data from the Geochemical and Mineralogical Data for Soils
of the Conterminous United States. We compare the performances of random
forest (RF) and inverse distance weighting (IDW) to model and generate
soil TP predictions. The RF incorporates 19 predictor variables, including
spatial coordinates, climate, soil properties, and topography, while IDW
relies solely on coordinates and interpolates between soil TP
observations. Models are evaluated using five-fold cross-validation. The
RF models outperform the IDW models and explain 52 % (RMSE =
0.22 log10 mg kg -1) and 56 % (RMSE = 0.26 log10 mg kg -1) of the
variance in soil TP for the top 5 cm and C horizon, respectively. As
expected, both model types identify higher TP concentrations in the top 5
cm than in the C horizon, particularly in agricultural regions, reflecting
anthropogenic influences. Furthermore, the RF-generated maps show more
realistic spatial patterns that capture the heterogeneity of the CONUS and
avoid the bullseye patterns often characteristic of IDW-generated maps.
Additional insights from the RF models show that coordinates, soil
texture, pH, and climate are top predictors of soil TP. Increased
availability of variables, such as iron and aluminum, that can bind with
phosphorus in soils, could improve RF model performance.