Multi-source data-driven prediction of cold-region slope failure using an SSA-PNN optimized stepwise reduction approach.

Abstract

Cold-region slopes are highly susceptible to instability due to freeze-thaw-creep coupling, which gradually degrades the mechanical properties of geomaterials and accelerates the formation of slip surfaces. To address the limitations of conventional strength reduction methods that fail to capture progressive failure mechanisms, this study proposes an integrated framework that combines the Stepwise Reduction Method (SRM) with a machine learning model based on a Sparrow Search Algorithm-optimized Probabilistic Neural Network (SSA-PNN). A multi-source dataset was established by incorporating 42 field monitoring segments, 78 numerical simulation samples, and laboratory tests of mechanical degradation under 0-60 freeze-thaw cycles, covering 15 environmental, material, structural, and response features. The model performance was evaluated using classification and regression tasks, with safety factor (FS) as the reference label. Results show that the SSA-PNN achieved an accuracy of 87.5%, macro-F1 of 0.869, weighted F1 of 0.877, macro-AUC of 0.979, and Brier score of 0.056 in classification, while in regression it obtained MAE = 0.041, RMSE = 0.053, and R² = 0.871, consistently outperforming benchmark models such as XGBoost, SVM, and Logistic Regression. Notably, in the critical stability interval (1.30 ≤ FS < 1.50), the SSA-PNN reduced misclassification rates by 12.4% compared with the conventional PNN, demonstrating a marked improvement in distinguishing borderline states. These findings confirm that the SRM-SSA-PNN framework effectively characterizes the spatiotemporal evolution of slope degradation under freeze-thaw effects, enhances the interpretability of instability mechanisms, and provides a reliable basis for risk assessment, intelligent monitoring, and early warning of geohazards in cold regions.