Examining the Influence of Different Inventories on Shallow Landslide Susceptibility Modeling: An Assessment Using Machine Learning and Statistical Approaches

Shallow landslides are one of the most common natural hazards in Brazil and worldwide. Susceptibility maps are powerful tools to analyze the spatial probability of shallow landslide occurrences. The outputs of susceptibility maps strongly depend on the type of landslide inventory used. The aim of this study is to examine the influence of different inventories on shallow landslide susceptibility modeling using the different methods LR, SVM, and XGBoost. Three different shallow landslide inventories were compiled following a single extreme rainfall event in the Ribeira Valley, São Paulo, Brazil. The results indicate that inventories generated through different landslide detection methods and imagery produce diverse susceptibility maps, as evidenced by the calculated Cohen’s Kappa coefficient values (0.33–0.79). The agreement among the models varied depending on the specific model: LR exhibited the highest agreement (0.79), whereas SVM (0.36) and XGBoost (0.33) showed lower numbers. Conversely, the accuracy numbers suggest that XGBoost achieved the highest success rate in terms of AUC (85–78%), followed by SVM (82–76%), and LR (80–71%). Inventories obtained through different detection methods, using distinct datasets, can directly influence the susceptibility assessment, leading to varying classifications of the same area. These findings demonstrate the importance of well-established landslide mapping criteria.