Bayesian estimation of probability distributions of undrained shear strength of soils with limited site data

To address the issue of poor reliability of the design parameters due to limited or incomplete geotechnical investigation data, a cohesive soil parameter database containing 1679 sets of data from 141 sites is established. The site-specific Bayesian method (SBM) and the hierarchical Bayesian method (HBM) are used to estimate the probability distribution of undrained shear strength of cohesive soils by utilizing the data from a specific site and multiple sites, respectively. The results show that compared with the SBM method, the HBM method can effectively reduce the uncertainty of parameter estimation when there is only limited measured data at the target site, and it is less affected by the number of measuring points at the target site. The leave-one-out cross-validation (LOO-CV) combined with the log pointwise predictive density (lppd) is used to compare the accuracy of the two methods. The results show that the lppd_loo-cv index of the HBM method is larger, indicating that the overall prediction accuracy of the HBM method is higher. Therefore, the HBM method is more suitable for the estimation of undrained shear strength parameters in the case of limited site data, and the posterior means obtained by the HBM method can be used for parameter estimation of new sites.