Machine Learning-Based Evaluation of Tower Foundation Frost Jacking in the Marginal Permafrost of Northeast China

Rapid localized permafrost degradation heightens geological risk and threatens the long-term stability of transmission towers in cold regions. Based on systematic field investigations of 336 towers in the marginal permafrost of Northeast China, thirteen conditioning factors were incorporated to evaluate five machine learning models (Random Forest, XGBoost, LightGBM, CatBoost, Logistic Regression) for foundation frost jacking susceptibility assessment, with SHapley Additive exPlanations (SHAP) and Geographical Detector (GeoDetector) used to quantify factor contributions and interactions. Results show that differential frost jacking poses the greatest structural hazard. Peat-rich wetland geohydrological settings constitute the highest-risk terrain for foundation uplift. Random Forest achieved the highest area under the curve of 0.987 and showed competitive performance across multiple evaluation metrics. Ground ice content was the most influential predictor, followed by land cover, soil type, and surface water. Ice content and land cover exhibited the strongest explanatory power (q = 0.78). These findings provide a quantitative and interpretable basis for frost-jacking susceptibility assessment and preliminary engineering screening within the investigated transmission corridor.