Abstract
This study explores a hybrid approach that combines BPSO with ensemble machine learning techniques to improve predictive accuracy in assessments of slope stability. The methodology employs BPSO to optimize the selection of features that are critical to the prediction process. In addition, a grid search technique is utilized to fine-tune the hyperparameters of the ensemble models. The research evaluates the performance of three ensemble models: RF, XGBoost, and LightGBM. For the predictive analysis, six features identified as potentially influential were selected, including: slope height (H), pore water ratio (ru), unit weight (Υ), cohesion (c), slope angle (β), and angle of internal friction (Φ). The effectiveness of the models was assessed using various performance metrics, including the AUC, kappa and accuracy of the predictions. The findings indicate that the hybrid approach, particularly the LightGBM model, significantly outperformed the other models, achieving an AUC of 0.871, a kappa of 0.658 and an accuracy rate of 0.832. This underscores the potential of the proposed hybrid method as a valuable tool for accurately predicting slope stability and mitigating risks associated with slope failures in engineering applications.