Comparative Analysis of Ensemble Learning Approaches for Slope Stability Prediction

Due to the complex nature of slope engineering, accurately predicting slope stability using traditional techniques can be difficult. It is, therefore, crucial to identify the correct technique for slope stability prediction in order to prevent disasters caused by slope failures. This study provides a comprehensive analysis of three ensemble models: Random Forest (RF), CatBoost, and Stacking. The models were evaluated for a wide range of hyperparameters to find the optimal settings for each model, resulting in the best solution. Six potentially relevant features, including height (H), pore water ratio (ru), unit weight (Ƴ), cohesion (c), slope angle (β) and angle of internal friction (ɸ), were selected as prediction indicators. The generalization ability of classification models is enhanced by using a 5-fold CV. Evaluation indicators such as AUC and accuracy were analyzed, and Stacking was found to outperform the other ensemble models with the highest AUC of 0.898 and accuracy of 0.854. The analysis of engineering examples shows that Stacking is a highly effective tool for predicting slope stability due to its ability to enhance capacity and efficiency in deformation prediction models. This makes it the most accurate tool available for forecasting slope stability. In addition, a comprehensive analysis of parameter sensitivity was conducted to determine the most significant characteristics for predicting slope stability.