Predicting bulk density of soils with varying degree of structural degradation using single and multi-parameter based pedotransfer functions

Abstract

Bulk density (BD) is a key soil property needed to for example calculate carbon stocks or predict soil hydraulic properties. However, measuring soil BD on a large scale is often labor-intensive, time-consuming and expensive. While pedotransfer functions (PTFs) have been developed to predict BD from readily-available or easy-to-determine soil properties, they are often inaccurate. One of the reasons might be that PTFs typically do not include a soil structure related predictor variable. The objective of this study was to explore whether adding an easy-to-determine variable that represents soil structure or soil strength does improve the prediction accuracy of PTFs. 252 horizons were sampled on 42 agricultural fields in Belgium that cover all major texture classes according to the Belgian texture triangle. At each field, two positions of contrasting field traffic and thus difference in soil structure and soil strength were sampled at three depths. Several quantitative soil physical properties were determined using laboratory and field-based methods, including soil organic carbon content (SOC), texture (sand, silt, and clay content), penetration resistance (PR) at field capacity and a soil structural quality score (Sq) taken from a semi-quantitative visual soil assessment (based on VESS). PTFs with both single and multiple parameters were developed using two classical regression methods (multiple linear regression MLR and generalized additive model GAM) and two machine learning methods (support vector machine SVM and k-nearest neighbours KNN). In addition to the above mentioned soil properties, soil layer (depth) was as well considered as predictor variable. The BD in the Train dataset ranged from 1.10 to 1.74 Mg m⁻³ with a mean value of 1.45 Mg m⁻³. Similarly, in the Test dataset, BD values ranged from 1.14 to 1.67 Mg m⁻³ with the same mean value of 1.45 Mg m⁻³. Overall, it was found that the KNN method generally showed a better predictive performance with the higher R² and lower RMSE. Notably, while SVM also performed well, particularly on the Test dataset, its R² values exceeded those of KNN in some cases. PR and Clay content were the two most important variables for predicting BD in all models, SOC the third most important, while also the VESS-based Sq score was found to be a good potential predictor for BD PTF models. Grouping the dataset based on soil structure (by using Sq) could substantially improve the predictive ability of most single parameter PTFs, in contrast with grouping based on texture that resulted in a slight improvement only. Grouping based on Sq or texture also demonstrated improvements in the performance of multi-parameter models. Among these, the KNN and SVM models showed relatively better overall performance. For example, the KNN model with f(SOC, Clay, PR) grouping by Sq achieved an overall R² of 0.537 and RMSE of 0.057 Mg m⁻³ on the recombined complete dataset.