Gaussian Mixture Model Based Soil Classification Using Multiple Cone Penetration Tests

Abstract

Abstract This study presents an application of the Gaussian mixture (GM) method for soil classification using multiple cone penetration tests (CPT). Compared to the hard clustering methods, the GM model classifies the CPT data by representing the probability density function of observed variables as a mixture of multivariate normal distributions. A GM model based expectation maximization (EM) algorithm with Bayesian information criterion (BIC) for selecting the optimal number of clusters is developed using six real CPT data performed at Dunkerque site in the north of France. The classification results are compared with the classical CPT based interpretation using the non normalized soil behavior type (SBT) index together with the Robertson chart. The results show that the GM model is able to identify accurately the soil layers. In addition, the combination of all CPTs, rather than considering them separately, may improve the soil layers identification because all the site information is considered.