Maximum bulk density (BDmax) and critical water content (θc) (i.e., Proctor parameters) are valuable parameters to evaluate soil compactness and optimum moisture of workability for tillage. There are two novelties in the present study: First, no study has been conducted so far to estimate the Proctor parameters from CaCO3, saturated and field water contents in agricultural lands using state-of-the-art methods. Second, no study has been done to compare the estimation accuracy of linear (LR) and nonlinear (NLR) regression, support vector machine (SVM), and artificial neural networks (ANNs) methods in estimating Proctor parameters in agricultural soils. In total, 105 soil samples were taken from agricultural lands of Ardabil plain, northwest of Iran. Pedotransfer functions (PTFs) were constructed using SVM, ANNs, LR and NLR methods to estimate BDmax and θc from readily available soil properties including organic carbon (OC), CaCO3, particle size distribution (PSD), bulk (BD), and particle (Dp) density, total porosity (n), penetration resistance (PR), and saturated (θs) and field (θf) water contents. The results of the LR, NLR, ANNs, and SVM estimations showed that θs, θf, OC, and Dp were the most suitable estimators in estimating BDmax and θc. The values of root mean square error (RMSE) criterion in the best LR, NLR, ANNs, and SVM PTFs were obtained 2.3, 3.29, 2.19 and 3.09 g g−1 for θc and 0.05, 0.07, 0.05 and 0.07 g cm−3 for BDmax in the testing data set, respectively. Overall, Proctor parameters of agricultural soils could be accurately estimated by the ANNs compared with the LR, NLR and SVM.
