Accurate landslide susceptibility assessment remains challenging in mountainous regions with complex terrain, heterogeneous geology, and clustered landslide inventories. This study develops a slope–unit–based landslide susceptibility assessment framework for Tongren County, Qinghai Province, China, using a landslide inventory of 217 events, multi–source environmental data, Certainty Factor (CF)–based conditioning–factor analysis, and machine learning models. Eighteen conditioning factors derived from remote sensing, geological survey, and meteorological datasets were extracted at the slope–unit scale, and their collinearity was evaluated using Pearson’s correlation and the Variance Inflation Factor (VIF). Eight models—Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), AdaBoost, Decision Tree (DT), XGBoost, K–Nearest Neighbors (KNN), and Convolutional Neural Network (CNN)—were evaluated under a 70:30 train/test split. The results show clear performance differences among the tested models: SVM achieved the best overall balance between discrimination and landslide detection (AUC = 0.9489; recall = 0.879). The tested CNN baseline showed relatively weak performance under the current slope–unit–based tabular–data setting. Susceptibility zoning results showed that high– and very–high–susceptibility zones were mainly concentrated along the Longwu River and its tributaries, where middle–elevation dissected terrain, weak lithological materials, river–valley erosion, and human engineering activities spatially coincide. These results provide a practical basis for slope monitoring and land–use planning in Tongren County.

Landslide susceptibility assessment in cold, high-altitude regions is complicated by interactions among topographic, lithological, hydrological, and anthropogenic factors. Machine learning methods are widely used, but model selection, spatially independent evaluation, and interpretation of covariate-adjusted associations remain challenging. This study developed an integrated framework for landslide susceptibility assessment in Tongren City, on the northeastern margin of the Qinghai–Tibet Plateau. The framework combined H₂O AutoML, nested spatial cross-validation, Shapley additive explanations (SHAP), partial dependence plots (PDPs), and double machine learning (DML). Each of the 20 sampling realizations included 217 landslide points and 217 non-landslide points, with 13 conditioning factors retained after collinearity screening. DML used partially linear regression with spatial cross-fitting across 200 repetitions to estimate covariate-adjusted associations. Stacked Ensemble achieved the highest mean area under the receiver operating characteristic curve (AUC) of 0.953 ± 0.009 across the 20 realizations. Spatial out-of-fold evaluation showed that the 20% of the study area with the highest susceptibility contained 93.27% ± 1.51% of held-out landslides (mean ± standard deviation). Elevation, distance to roads, lithology, distance to faults, and land use/land cover together accounted for 76.69% of the total mean absolute SHAP values. PDPs revealed nonlinear relationships between conditioning factors and model-predicted susceptibility. Elevation, interpreted as a proxy for valley position, and distance to roads were the only factors with statistically significant negative covariate-adjusted associations after multiple-testing correction. The random common cause (RCC) test indicated that DML estimates remained stable under the random-covariate perturbations examined. The framework distinguishes predictive contributions from covariate-adjusted associations and supports landslide monitoring and spatial planning in cold, high-altitude plateau transition zones.