Predicting bulk density in Brazilian soils for carbon stocks calculation: a comparative study of multiple linear regression and Random Forest models using continuous and categorical variables

Soil bulk density (ρb) is vital for assessing soil organic carbon (SOC) and nutrient stocks, as well as for modeling soil processes. Although ρb can be measured using traditional methods, these are labor-intensive and time-consuming. Consequently, there is a growing interest in developing pedotransfer functions (PTFs) to predict ρb based on more easily measured soil properties. The vast Brazilian territory, with different climates and biomes, presents a challenge for development of national-scale PTFs, particularly for ρb. In this study, a comprehensive dataset was compiled from various sources to develop ρb PTFs across Brazil. The dataset includes particle size distribution (PSD), SOC, ρb, sampling depth, soil order, land use, and geographic coordinates, allowing for the incorporation of additional numerical and categorical variables. Rigorous data preprocessing ensured quality and reliability. PTFs were developed using multiple linear regression (MLR) and Random Forest (RF) models. Model accuracy was evaluated using mean absolute error, bias, root mean square error (RMSE), and coefficient of determination. Both MLR and RF models accurately predicted ρb, with log-transformed PSD and SOC emerging as key predictors. The RF model slightly outperformed the MLR model (RMSE = 0.12 vs. 0.13 g cm−3) on the test dataset, underscoring the importance of environmental and categorical variables in predicting ρb. The developed PTFs, along with other PTFs for Brazil, were applied to estimate SOC stocks across different biomes and land uses. Best estimations were obtained with the RF model, with an R2 of 0.97, emphasizing the value of categorical variables in improving SOC stock estimations.