Development of a new 30-m soil properties map over the CONUS using pruned hierarchical Random Forests and iterative bias-correction

Digital soil maps provide important information into the spatial distribution of soil properties, supporting land management decisions and Earth system modeling. This study introduces a novel methodology for creating 30-m digital soil maps across the conterminous United States (CONUS). This approach reduces uncertainties in estimating soil property distribution. The predicted soil properties include soil texture, bulk density, soil hydraulic properties (pedotransfer function-derived), pH, and organic matter - through six standard depth intervals from surface to 2-m depth. Comparative analysis demonstrates improved performance over existing soil products over CONUS.

Our methodology uses a two-step process. First, we have developed a pruned hierarchical Random Forest (pHRF) method to generate prior distributions of each soil property. Key highlights of the pHRF method include: (1) efficient selection of soil covariates, such as Sentinel 1 and 2 satellites and GOES land surface temperature; (2) implementation of a 'moving polygon' algorithm that preserves natural landscape boundaries; (3) incorporation of point-based soil measurements; and (4) development of a pruned hierarchical Random Forest algorithm that reduces uncertainties in estimating soil properties and addressing the inherent imbalance in soil survey data (uneven distribution of soil observation and underrepresented soil classes).

To further enhance the performance of the soil dataset, we also incorporate bias correction as a post-processing process. This process incorporates additional soil profile data to iteratively correct histograms of soil properties at each location. This process continues until the residual variations of soil properties between iterations fall below a predetermined threshold, indicating convergence. Our method leverages the most probable predicted soil property values to correct their distributions while accounting for spatial correlations between different soil property layers. This data-driven approach makes no assumptions about the underlying distribution of soil properties and relies on non-parametric statistical models. The resulting posterior distributions show reduced prediction uncertainties. It also demonstrates reproducibility in the final soil maps. These soil maps provide valuable input for various land management and modeling applications.