Evaluating spatiotemporal variations in landslide susceptibility in Northwest Vietnam

Abstract Torrential rainfall is considered a primary trigger for catastrophic landslide events, which threaten the infrastructure, people, and economy of Vietnam. This study analysed spatiotemporal changes in landslide susceptibility using multitemporal landslide inventories and dynamic predictors, including average annual rainfall, land cover, and NDVI over six years. Landslide susceptibility models for three years were built using extreme gradient boosting (XGB), random forest (RF) and logistic regression (LR) algorithms with multitemporal landslide inventories and 12 predictors. Model performance was evaluated using different spatial block sizes, and a block size of 3,000 m was identified as the optimal configuration, balancing the reduction of spatial autocorrelation with stable predictive performance. The XGB model exhibited the best performance across most validation metrics. SHAP and ALE analyses identified rainfall and elevation as the most important predictors, with landslide probability increasing with annual rainfall up to 1,725 mm and elevation up to 2,000 m. Area of applicability AOA analysis confirmed the XGB model is reliable for producing susceptibility maps across almost the entire study area. Using successive landslide inventories and dynamic predictors retained information from previous events while capturing temporal environmental changes. Although these additions did not result in a continuous improvement in discriminative performance, they contributed to better calibration of the predicted probabilities.