Landslide inventories

Reasonable landslide susceptibility mapping (LSM) is helpful to avoid damage caused by landslides. At present, it is not clear which screening method can yield the optimal combination of causative factors. Moreover, the "black box" nature of machine learning models limits understanding of their internal mechanisms and the driving roles of causative factors in the landslide development process. In this work, GeoDetector, information gain (IG), and Gini index (Gini) were used to select the optimal combination of causative factors to explore the most effective method. The Random Forest (RF_GeoDetector, RF_IG, and RF_Gini), Support Vector Machine (SVM_GeoDetector, SVM_IG, and SVM_Gini), and eXtreme Gradient Boosting (XGBoost_GeoDetector, XGBoost_IG, and XGBoost_Gini) were constructed for LSM.Furthermore, the SHapley Additive exPlanation (SHAP) algorithm was applied to quantify the marginal contribution of each causative factor to the landslide susceptibility. The conclusion showed that (1) Giniis superior to GeoDetector and IG in screening the optimal combination of causative factors. (2) The SHAP revealed that proximity to roads was the most significant contributor to landslide occurrence, followed by lithology and annual average precipitation of March-July. It highlights the coupling effect of natural and anthropogenic factors on landslide susceptibility. (3) For LSM, the XGBoost_Ginimodel achieves the optimum performance. A reasonable LSM provides better technical support and more accurate disaster prevention.