Combining Physical Models and Machine Learning for Efficient Slope Stability Prediction Under Extreme Rainfall

Climate change has increased both the frequency and intensity of extreme rainfall events worldwide, significantly intensifying the risk of rainfall-induced shallow landslides. This study proposes an approach combining physical models and machine learning to predict slope stability. Using a specific region in Singapore as the study area, the GEOtop, a distributed hydrological model, was employed to simulate volumetric water content (VWC) under two scenarios: maximum daily rainfall and maximum 5-day antecedent rainfall. The simulated VWC was subsequently input into the Scoops3D model to calculate the Factor of Safety (FoS). A Random Forest (RF) model was trained using multiple input variables, including VWC from GEOtop, with FoS from Scoops3D under the maximum 5-day antecedent rainfall as the target variable. The trained RF model was then applied to predict FoS under an unseen scenario-the maximum daily rainfall-and compared with corresponding Scoops3D predictions. The results demonstrated that the RF model's performance was overall comparable with Scoops3D. The predictions of the percentage of different risk levels and their variation trends were similar to those of Scoops3D. For very high-risk zones (FoS d 1), specifically, the predictions were particularly consistent. Spatial mapping of FoS distribution further revealed a strong alignment between the two methods. Additionally, two specific points were selected for comparison with two two-dimensional (2D) prediction strategies. The results showed that Scoops3D generally predicted lower FoS values than the 2D physical models, while the RF model closely matched Scoops3D in terms of FoS variations. This study highlights the effectiveness and generalisationof the proposed method in achieving reliable prediction accuracy and improved computational efficiency, providing a viable solution for regional slope stability assessments under extreme rainfall conditions.

Keywords: Unsaturated soil, Slope stability, Machine learning, Physical models, Comparative analysis.