Predicting soil bulk density for incomplete databases

- Proposed PTFs accurately predicted missing ρb values in databases.
- Information from other horizons in the soil profile was crucial for PTFs.
- Bulk density of other horizons in the soil profile was the most important variable.
- Threshold depth of 50 cm gave the best upward and downward predicting models.

Soil bulk density (ρb) is important because of its direct effect on soil properties (e.g., porosity, soil moisture availability) and crop yield. Additionally, ρb measurements are needed to express soil organic carbon (SOC) and other nutrient stocks on an area basis (kg ha− 1). However, ρb measurements are commonly missing from databases for reasons that include omission due to sampling constraints and laboratory mishandling. The objective of this study was to investigate the performance of novel pedotransfer functions (PTFs) in predicting ρb as a function of textural class and basic pedon description information extracted from the horizon of interest (the horizon for which ρb is being predicted), and ρb, textural class, and basic pedon description information extracted from horizons above or below and directly adjacent or not adjacent to the horizon of interest. A total of 2,680 pedons (20,045 horizons) were gathered from the USDA-NRCS National Soil Survey Center characterization database. Twelve ρb PTFs were developed by combining PTF types, database configurations, and horizon limiting depths. Different PTF types were created considering the direction of prediction in the soil profile: upward and downward prediction models. Multiple database configurations were used to mimic different scenarios of horizons missing ρb values: random missing (e.g., ρb sample lost in transit) and patterned or systematic missing (e.g., no ρb samples collected for horizons > 30 cm depth). For each database configuration scenario, upward and downward models were developed separately. Three limiting depths (20, 30, and 50 cm) were tested to identify any threshold depth between upward and downward models. For both PTF types, validation results indicated that models derived from the database configuration mimicking random horizons missing ρb performed better than those derived from the configuration mimicking clear patterns of missing ρb measurements. All 12 PTFs performed well (RMSPE: 0.10-0.15 g cm− 3). The threshold depth of 50 cm most successfully split the database between upward and downward models. For all PTFs, the ρb of other horizons in the soil profile was the most important variable in predicting ρb. The proposed PTFs provide reasonably accurate ρb predictions, and have the potential to help researchers and other users to fill gaps in their database without complicated data acquisition.