Influence of Sampling Methods on the Accuracy of Machine Learning Predictions Used for Strain-Dependent Slope Stability

Abstract

Supervised machine learning (ML) techniques have been widely used in various geotechnical applications. While much attention is given to the ML techniques and the specific geotechnical problem being addressed, the influence of sampling methods on ML performance has received relatively less scrutiny. This study applies supervised ML to the strain-dependent slope stability (SDSS) method for the prediction of the factor of safety (FoS) using hypoplasticity. It delves into different sampling strategies for training the ML model, emphasizing predictions of soil behavior in lower stress ranges. A novel sampling method is introduced to ensure a more representative distribution of samples in these ranges, which is challenging to achieve through traditional sampling approaches. The ML models were trained using traditional and modified sampling methods. Subsequently, slope stability analyses using SDSS were conducted with ML models trained from six different sampling methods. The results illustrate the impact of sampling methods on the FoS. Besides a noticeable improvement in predictions of shear stresses within the lower stress ranges, a decisive effect on the overall FoS was observed as well.