Slope stability prediction using multi-stage machine learning with multi-source data integration strategy

Slope stability analysis is a crucial task in geotechnical engineering. Machine learning (ML) has been widely used for slope stability analysis based on inference relationships learned from training data. However, the performance of ML is constrained since available data from historical slopes is limited. The multi-source training dataset is generated in this study, including data from historical slopes, numerical simulations, and physical experiments. The multi-stage machine learning (MSML) model is proposed to utilize training data from different sources. We select stability status, factor of safety, and probability of slope stability as outputs, and design the hybrid loss function consisting of data-based and knowledge-based terms to further improve the performance of the MSML model based on the relationships among three outputs. The results of the proposed MSML model are then compared with the baseline ML methods using evaluation metrics. The comparisons demonstrate that the proposed MSML model outperforms all baseline ML methods with the highest accuracy (0.893), precision (0.845), recall (0.907), and F1-score (0.868). Even the proposed pre-trained model achieves performance similar to most baseline models. Finally, a further comparison using five landslide cases demonstrates that our proposed method provides accurate and reliable prediction results with excellent generalization.