Pedotransfer functions for estimating the maximum bulk density of major Brazilian agricultural soils: Model comparison and validation

Abstract

The Proctor compaction test is the standard method for determining the maximum bulk density (Dbmax) of a soil, a key parameter for obtaining the degree of compaction. However, this method is labor‐intensive and time‐consuming, making it impractical for routine soil quality assessment. We developed pedotransfer functions to estimate Dbmax of agricultural soils in Brazil from texture, soil organic carbon (SOC), and bulk density (Db). A dataset consisting of 234 compaction curves from peer‐reviewed studies was used to compare linear regression (LR) and machine learning methods using cross‐validation and two independent validation datasets, including additional data from Brazil and the Americas. Random forest (RF) and support vector regression (SVR) reduced prediction variance for the validation dataset of Brazilian soils but overestimated Dbmax; however, including Db reduced this bias. For the Americas dataset, LR showed high variance and low bias, SVR resulted in low variance but higher bias, whereas RF presented the best bias–variance trade‐off. Although RF provided the best predictions (RMSE of 0.069 and 0.078 g cm ⁻³ with and without Db, respectively), Dbmax for soils within the range represented by our dataset can also be estimated by Dbmax = 2.355 + 0.6 × SAND ² − 0.4 × SILT ² − 0.44 × ln(SAND) − 0.52 × ln(CLAY) − 0.38 × (CLAY:SAND) − 0.81 × (SOC/CLAY + SILT) and Dbmax = 1.41 + 0.32 × SAND ² − 0.10 × SILT ² − 0.23 × SOC + 0.37 × Db (texture and SOC in g kg ⁻¹ and Db in g cm ⁻³ ), with RMSE values of 0.084 and 0.066 g cm ⁻³ , respectively. This facilitates the estimation of Dbmax and supports the wider use of the degree of compaction in soil quality assessment.