| Summary: | Soil shear strength is a critical parameter in slope stability. Shear strength properties may vary significantly over large areas. Thus, the spatial estimates of shear strength properties are necessary for deterministic slope susceptibility mapping over large areas. However, measurements of shear strength parameters are often limited as compared to other soil properties such as Atterberg limit, bulk density and grain size distribution. Multivariate methods have been shown to improve prediction accuracy, but these methods have rarely been used to predict shear strength. In this study, attempts were made to evaluate the effectiveness of using the aforementioned soil properties in predicting the spatial variation of shear strength properties: effective cohesion (c’) and effective friction angle (ϕ’). The performance of ordinary kriging (OK), Random Forest (RF) and regression kriging (RK) in predicting c’ and ϕ’ of residual soils in Singapore were compared and evaluated. In addition, the sensitivity of the three methods to the sample size was investigated. The results of RF analysis revealed that the northing coordinate and percentage of fines were the most important variables for predicting ϕ’. The spatial coordinates and ϕ’ were also important variables for predicting c’. The predicted c’ and ϕ’ using RF and RK resulted in higher spatial heterogeneity than OK. Overall, RF had the smallest error as compared to OK and RK in predicting c’ and ϕ’ at all sample sizes, except for the prediction of ϕ’ using the largest sample size. This study also showed that RF and RK were more sensitive to sample size than OK. These results highlight the benefits of using auxiliary variables when mapping shear strength properties.
|
|---|