A hidden Markov random field model based approach for probabilistic site characterization using multiple cone penetration test data

This paper presents a new probabilistic site characterization approach for both soil classification and property estimation using sounding data from multiple cone penetration tests (CPTs) at a project site. A hidden Markov random field (HMRF) model based Bayesian clustering approach is developed, which can describe not only the heterogeneity of properties in statistically homogeneous soil layers, but also the correlation between spatial distributions of different soil layers. The latter has not been well considered in the existing CPT interpretation methods. A Monte Carlo Markov chain based expectation maximization (MCMC-EM) algorithm is adopted to calibrate the established HMRF model, so that both the subsurface soil/rock stratification and the pertinent soil properties can be estimated in a probabilistic manner. The proposed CPT interpretation approach is validated and demonstrated using a series of numerical examples, including using real CPT data. It is shown that the proposed method is able to accurately identify soil layers, pinpoint their boundaries, and provide reasonable estimates of the associated soil properties. In addition, comparative studies show that combining analysis of CPT data from multiple soundings, rather than interpreting them separately, can significantly enhance the accuracy of interpretation and simplify the subsequent task of interpreting stratigraphic profiles.