Predicting the dry bulk density of soils across Denmark: Comparison of single-parameter, multi-parameter, and vis–NIR based models

Dry bulk density (BD) is a key soil physical property, characterizing the level of soil compaction and controlling flow and transport of fluids and solutes. In this study, predictive BD models were developed from 2462 BD measurements of mineral soils representing both top- and subsoil horizons across Denmark. Three types of BD models were compared: (i) single-parameter based on soil organic carbon (SOC), (ii) multi-parameters based on SOC, soil texture and depth, and using multiple linear regression (MLR), and (iii) visible–near infrared spectroscopy (vis–NIRS) based, using the spectral information and partial least squares regression (PLSR). Also, three machine learning techniques (random forest (RF), regression rules (RR), and artificial neural networks (ANN)) were applied for multi-parameter and vis–NIRS based models to potentially improve predictions. BD models were calibrated on 70% of dataset and validated on the remaining 30%. Single-parameter models with SOC had the lowest predictive potential and were not able to predict BD above 1.62 g cm⁻³. Machine learning using RF, RR and ANN on soil properties resulted in significantly higher prediction accuracies (RMSE ≤ 0.11 g cm⁻³ and R² ≥ 0.60) than MLR (RMSE = 0.12 g cm⁻³ and R² = 0.41) and SOC-based models. The relative importance of variables was obtained from the multi-parameter models with SOC and clay as the most important variables. The prediction accuracies MLR and vis–NIRS did not differ but were better than SOC-based models, suggesting that vis–NIRS could be a cost- and time-efficient alternative to pedotransfer functions that generally require determination of one or more soil properties for developing the calibration models. Decomposing the spectral data into principal components and using selected components as inputs for RF, RR and ANN improved the model performances only for RF models. Instead and in perspective, combining techniques by supplementing spectral data with readily available soil information and using machine learning techniques may further improve BD predictions.