A comprehensive approach for Bayesian soil classification using Cone Penetration Test data

Classification is a critical soil characterisation task, meaning there is considerable value in methods which can accurately and reliably predict soil type from cost-effective Cone Penetration Test (CPT) data. Traditional methods involve plotting measurements against 2-dimensional charts. However, these have evolved only incrementally over past decades despite several key limitations: (i) inference in two dimensions is suboptimal, as more than two measurements are available; (ii) the quantitative specification of the charts is relatively untransparent; and (iii) the outputs are deterministic, which can be misleading as the estimation process is highly uncertain. This paper introduces a comprehensive CPT-based classification approach which is entirely data-driven and delivers probabilities for each possible soil type. Reliable training data is curated from a large database through automated pairing of laboratory and CPT data, with the latter subject to novel pre-processing techniques to remove uncharacteristic features. Bayesian inference is used to characterise a multinomial logistic regression model: given training data and an optional prior model density, the posterior density is computed. Class prediction for new data is then performed through Monte Carlo integration over the posterior density. The resulting model is found to deliver strong classification accuracy, and the probabilistic output is highly practical for interpretation; in particular, being well suited to automatic stratification. Additionally, an existing model can easily be adapted to a new set of classes and re-trained, enabling refinement or site-specific calibration, with robust continuity facilitated through use of the prior.