The role of conditioning factors in machine learning-based landslide spatial probability

Conditioning factors (CFs), such as topographic, hydrological, and environmental factors, significantly influence the accuracy of predicting landslide spatial probability (LSP). This study applied three machine learning models – random forest (RF), deep neural network (DNN), and extreme gradient boosting (XGBoost) – to predict LSP in Phuoc Son, Quang Nam, Vietnam. Historical landslide data were split into training and validation sets to create and assess LSP maps. Model performance was assessed using receiver operating characteristic curves, with success rate analyses showing area under the curve (AUC) values of 87.3% for RF, 92.4% for DNN, and 93.7% for XGBoost. Validation AUC values were similarly high, reaching 87.5% for RF, 92.5% for DNN, and 93.8% for XGBoost, confirming the models’ robustness. To clarify the contributions of CFs, their relative importance was evaluated using permutation, decision tree, and coefficient-based measures. In addition, SHapley Additive exPlanations and Local Interpretable Model-Agnostic Explanations were used to provide both global and local insights into how each CF affects LSP. This framework combines model-dependent and model-agnostic approaches, providing a comprehensive interpretation of landslide driving mechanisms, which have been rarely explored in previous studies. Among the CFs analysed, the normalised difference vegetation index and slope were identified as the most influential in predicting LSP.