Landslide Susceptibility Evaluation in Southern Ningxia Under Future Heavy Rainfall Scenarios Based on GCCM-KNN

Objectives: This study proposes an integrated framework combining causal factor screening, machine-learning susceptibility modeling, and R95P-driven dynamic evaluation to assess landslide susceptibility in southern Ningxia, China, under predicted heavy-rainfall conditions from 2025 to 2034. Methods: Based on 777 historical landslides and 17 conditioning factors, a Geographical Convergent Cross Mapping method coupled with k-Nearest Neighbors, termed GCCM-KNN, was developed to identify causally relevant factors through adaptive neighborhood selection and local weighted fitting. The selected factors were compared with those obtained using original GCCM, Principal Component Analysis, ReliefF, and the full-factor scheme. Support Vector Machine, Random Forest, and XGBoost models were constructed for static susceptibility assessment, with SHAP used for interpretation. For dynamic evaluation, annual R95P from 1960 to 2024 was predicted using an ensemble model integrating Support Vector Regression, XGBoost, and Graph Neural Networks with ridge-regression weighting. Predicted R95P layers for 2025–2034 were then incorporated into the optimal susceptibility model while other factors remained unchanged. Results: GCCM-KNN identified nine key factors: gully density, elevation, relief degree, lithology, distance to fault, precipitation, distance to river, population density, and distance to road. Compared with other factor-selection schemes, GCCM-KNN improved the balance between model simplicity, causal interpretability, and spatial discrimination. The GCCM-KNN-XGBoost model achieved the best static performance, with an Accuracy of 0.881, F1-Score of 0.878, and Precision of 0.900. The R95P prediction model performed well during 2015–2024, with an average RMSE of 24.77 mm, MAE of 20.37 mm, and R² of 0.88. From 2025 to 2034, predicted R95P showed moderate interannual fluctuation, and dynamic susceptibility remained generally stable with local upgrading in transition zones. By 2034, medium and high susceptibility areas increased to 10.7% and 13.6%, respectively. In addition, 48% of newly identified 2025 landslides fell within dynamically predicted high or very high susceptibility zones. Conclusions: GCCM-KNN improves landslide factor selection by integrating causal constraints and adaptive spatial neighborhoods. The GCCM-KNN-XGBoost model provides accurate and interpretable susceptibility assessment for southern Ningxia. Coupling predicted annual R95P with the optimal model enables dynamic evaluation of rainfall-driven susceptibility changes. Future heavy rainfall may not reshape the overall susceptibility pattern, but it can promote local upgrading in valley margins, piedmont zones, and transition belts. However, annual R95P cannot fully represent short-duration rainstorms, rainfall peak intensity, or fine-scale precipitation heterogeneity, which should be further addressed in future studies.