Abstract: Accurate determination of random field parameters and correlation function is the key for probabilistic characterization of inherent spatial variability (ISV) of soil properties. Based on indirect cone penetration test (CPT) data, the conventional statistical methods and Bayesian approaches are compared for their validity of estimating random field parameters and correlation function of sand effective friction angle, φ'. The reason for the difference between these two methods is presented. Effects of sampling size of indirect CPT data on the accuracy of these two methods are also investigated. The results indicate that the conventional statistical methods don't consider the model uncertainty associated with the transformation model between φ' and CPT data, leading to a low rate of correct identification of true correlation function of φ', overestimating standard deviation and underestimating scale of fluctuation. However, the Bayesian approaches can take the model uncertainty into proper consideration and reasonably determine the random field parameters and correlation function of φ'. Model uncertainty should be carefully considered when indirect test data is used to characterize ISV of soil properties. In addition, the rate of correct identification of true correlation function of φ' estimated from Bayesian approaches improves and the uncertainties of random field parameters of φ' gradually decrease with increasing sampling size of indirect CPT data. It is suggested that large numbers of site observation data should be collected for improving the accuracy of probabilistically characterizing ISV of soil properties.

Key words: inherent spatial variability, sand effective friction angle, cone penetration test, conventional statistical methods, Bayesian approaches