Prediction of soil bulk density in agricultural soils using mid-infrared spectroscopy

Soil bulk density (BD) is a key physical parameter in soil quality control and in the calculation from soil organic carbon (SOC) mass (g/kg) content to area stock (kg/ha). However, BD laboratory analysis is time-consuming, labour intensive and expensive, especially for a national-scale soil assessment. Hence, how to fill the omissions of BD values for all or some records in soil databases is widely discussed. This study employed different chemometric and machine learning algorithms to estimate BD in Irish soil from 671 horizon-based samples from MIR spectral libraries by partial least square regression (PLSR), random forest, Cubist and support vector machine (SVM). The best performance was observed for the SVM model with a higher ratio of performance to interquartile distance (RPIQ = 3.61) and R2 (0.81) values and lower root mean square error of prediction (RMSEP = 0.132). Moreover, BD highly correlated wavenumber bands were determined by principal components analysis (PCA) and variable importance analysis. Soil organic matter (SOM) was identified as the primary factor in the spectral soil BD model. The generalisation error of predicting unknown samples using a spectral soil bulk density (BD) model was calculated by employing leave-one-out cross-validation (LOO-CV) on SVM. Estimation of BD by the spectral BD model was compared with published traditional pedo-transfer functions (PTFs), results were then compared for the overall models, different horizon types and specific depth categories. The spectral soil BD model is significantly better than traditional PTFs overall, with RMSEP equalling 0.132 g/cm3 and 0.196 g/cm3 respectively. The spectral soil BD model showed a similar accuracy on the A horizon, but considerable performance improvements were found on the other types of horizon. As for different depth categories, there is no significant accuracy difference between shallow (A-Samples: 5-20 cm) and deep (S-Samples: 35-50 cm) topsoil for the spectral soil BD model, which differs from traditional PTFs. The findings suggest that spectral modelling techniques, such as SVM, can provide high accuracy and homogenous performance across different depth layers, making them suitable for national soil surveys and large-scale carbon stock assessments. The best SVM model was then used to estimate BD values for a large archive of samples from the northern half of Ireland (Terra Soil project) and soil BD maps were generated at two different fixed-depth layers respectively. Besides that, all predicted soil BD values will be used for calculating soil carbon stock and assessing carbon deficit and sequestration potential in subsequent stages of the research.

Keywords: Soil; Bulk density; Mid-infrared; Spectroscopy; Chemometrics; Machine learning