Bohn PhD Defense 2022

Digital soil mapping (DSM) is widely promoted as a next-generation approach for producing soil maps with greater accuracy and spatial precision than traditional soil survey products. Advances in fine-resolution digital terrain analysis, remote sensing, geospatial workflows, and machine learning (ML) have greatly improved spatial prediction capabilities over the past two decades. However, the degree to which DSM provides clear improvements over conventional soil maps is not always evident, making evaluation essential. As global demands increase for agronomic efficiency while reducing nutrient and sediment losses, greenhouse gas emissions, erosion, and enhancing carbon sequestration, there is growing need for updated, subfield-scale soil property information to support agroecosystem research and management. This dissertation evaluated the effectiveness of DSM using ML through multiple approaches: regional-scale comparisons of map accuracy and resolution (Chapter 2), assessment of validation methods for unbiased accuracy estimation (Chapter 3), evaluation of the predictability of diverse soil properties and depths at the plot scale (Chapter 4), and testing the impact of DSM-derived soil inputs on agroecosystem model simulations (Chapter 5). In Chapter 2, DSM products were compared to SSURGO in a high-quality soil survey region. Target variables included key morphologic indicators of topsoil thickness and hydrologic regime—depth of mollic colors (mollic depth) and depth to reduced redoximorphic depletions (RRD depth)—as well as sand, silt, clay, and soil organic matter (OM) across six depth intervals to 2 m. DSM generally produced more accurate 10-m maps than SSURGO, improving mollic depth and RRD depth predictions by ~20 cm. DSM also outperformed SSURGO for most texture and OM predictions, with a few depth-specific exceptions. Chapter 3 evaluated how different validation strategies influence estimated map accuracy. Hypotheses tested whether (1) cross-validation underestimates error relative to independent validation, (2) independent validation underestimates error relative to geographically disjoint validation, and (3) spatial autocorrelation between training and validation points reduces estimated error. Only hypothesis (2) was supported. Chapter 4 assessed DSM-ML prediction performance for 18 agronomically relevant soil properties across seven depth intervals to 2 m at the plot scale. SOM fraction properties were easiest to predict, followed by carbonate properties and soil-water solution properties, while texture was most difficult. The 15–30 cm interval was most predictable, and accuracy generally declined with depth. Chapter 5 integrated the DSM products into agroecosystem process modeling to evaluate sensitivity to soil input data. Simulations using DSM-derived inputs produced substantially lower average maximum maize biomass and yield than simulations driven by SSURGO, demonstrating that soil input choice can strongly influence model outcomes.