Unsupervised and Simultaneous Stratigraphic Interpretation of CPT Soundings at Site Scale

This paper presents a novel Bayesian machine-learning approach for unsupervised and simultaneous soil stratigraphic interpretation of cone penetration test (CPT) soundings at the site scale. The proposed approach interprets numerous CPT soundings in a joint manner, and it leverages the statistical similarity of the measured sounding data in feature space (i.e., the Robertson chart) and the spatial constraints induced from spatial correlations of the sounding data both vertically along a single CPT sounding and horizontally across multiple soundings in physical space. The mathematical core of the proposed approach consists of the following two parts: (1) a quasi-3D (or 3D axial-symmetric) hidden Markov random field (HMRF) model describing both the statistical and spatial patterns of the CPT soundings; and (2) a model inference process, in which the statistical and spatial patterns are extracted from the dataset using a Bayesian unsupervised learning algorithm. The joint interpretation strategy of the proposed approach facilitates the use of rich statistical information and spatial constraints contained in an ensemble of CPT soundings to enhance the accuracy and consistency of a stratigraphic interpretation at the site scale. The proposed approach has been tested in a real-world case consisting of 44 CPT soundings collected from a geotechnical investigation site. The interpretation results show that the proposed approach can extract the soil spatial and statistical patterns from multiple CPT soundings and significantly increase the accuracy and consistency of the soil stratigraphic interpretation results.