Using machine learning to predict soil bulk density on the basis of visual parameters : Tools for in-field and post-field evaluation

Bondi, Giulia; Creamer, Rachel; Ferrari, Alessio; Fenton, Owen; Wall, David


Soil structure is a key factor that supports all soil functions. Extracting intact soil cores and horizon specific samples for determination of soil physical parameters (e.g. bulk density (Bd) or particle size distribution) is a common practice for assessing indicators of soil structure. However, these are often difficult to measure, since they require expensive and time consuming laboratory analyses. Our aim was to provide tools, through the use of machine learning techniques, to estimate the value of Bd based solely on soil visual assessment, observed by operators directly in the field. The first tool was a decision tree model, derived through a decision tree learning algorithm, which allows discrimination among three Bd ranges. The second tool was a linear equation model, derived through a linear regression algorithm, which predicts the numerical value of soil Bd. These tools were validated on a dataset of 471 soil horizons, belonging to 201 soil profile pits surveyed in Ireland. Overall, the decision tree model showed an accuracy of ~ 60%, while the linear equation model has a correlation coefficient of about 0.65 compared to the measured Bd values. For both models, the most relevant property affecting soil structural quality appears to be the humic characteristics of the soil, followed by soil porosity and pedogenic formation. The two tools are parsimonious and can be used by soil surveyors and analysts who need to have an approximate in-situ estimate of the structural quality for various soil functional applications.