Assessing Diagnostic Error of Factors of Safety of Slopes Applying Bayesian Inference

Stability analyses are crucial for identifying soil slope failure susceptibility. Currently, several analytical methods are available to perform such analyses. However, the degree of agreement of the analytical factors of safety from the different analytical methods with the field condition is an ongoing discussion in the geotechnical field. This paper presents the use of Bayesian inference to assess the accuracy of slope failure diagnostics from analytical methods based on limit-equilibrium (LE) and finite-element (FE) analysis in two- and three-dimensional spaces. In addition, a comprehensive statistical assessment is conducted by performing a randomized complete block design (RCBD) that provides the relationship of the factor of safety (FS) among the studied methods. From the analyses, a threshold FS of 1.20 is needed to achieve accuracy and an area under the receiver operating characteristic curve (AUC-ROC) larger than 90% for all current methods when compared to field conditions. The proposed method was able to identify the most vulnerable analytical condition for accurate diagnosis of slope failures. All the analytical methods under this study have a relatively low accuracy when a threshold value FS of 1 is used to indicate stability. Considering the findings, the reliability of back-calculated parameters should be carefully studied.