Application of Image Segmentation for Predicting Slope Failure Mechanism in Spatially Variable Soils

The representation of the spatial variability of soil properties in the form of random fields permits advanced and realistic probabilistic assessment of slope stability. As the intensity and the spatial distribution of soil properties vary in different random field realizations, the resulting failure mechanisms will also be different. Not only does the location of the failure surface vary, but the failure mode itself could also change. While the random field finite element method (RF-FEM), which adopts the shear strength reduction technique to naturally seek out the failure mechanism, is advantageous over other numerical schemes that need the failure mode to be defined a-priori, it suffers from a lack of computational efficiency. In this paper, the application of the image segmentation technique for predicting slope failure mechanism in spatially variable soils is investigated. The deep learning model based on the Resnet-18 architecture is applied to study the failure mechanism of a soil slope modified from a real case study. With sufficient training samples, the deep learning model can function as an image segmentation tool to delineate the sliding mass and the intact mass in the slope, based on which the mode and the location of the slip surface can be predicted without the need to perform the time-demanding random field finite element analysis. By using such an image segmentation tool, a large number of Monte-Carlo realizations of random fields can be efficiently analyzed. The spatial distribution of failure surfaces and the statistical distribution of sliding volume can then be calculated, leading to an improved understanding of slope reliability.

Keywords: Spatial variability, Slope stability, Failure mechanism, Image segmentation.