Evaluating a semi-automatic landslide inventory for machine learning-based shallow landslide susceptibility assessment

The acquisition of landslide inventories is the first step in landslide susceptibility assessment. Inventories indicate the geographical coordinates and the morphological and geological characteristics of areas where landslides have occurred, providing essential information for susceptibility analysis. Traditionally, the construction of landslide inventories is performed manually, relying on expert experience and requiring considerable time. Remote sensing techniques offer an alternative for faster mapping through automatic and semi-automatic approaches. Thus, this study evaluates the applicability of a semi-automatic landslide inventory within three susceptibility models: Logistic Regression (LR), Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGBoost). The study area is the Palmital–Gurutuba watershed located in the municipalities of Itaóca and Apiaí, São Paulo State, Brazil. The inventory was constructed following a single extreme rainfall event that triggered landslides on January 15, 2014. The results indicate good applicability of the semi-automatic shallow landslide inventory across all three models. For LR, the AUC-Success and AUC-Prediction were 0.77 and 0.80; for SVM, 0.88 and 0.82; and for XGBoost, 0.94 and 0.85. The Cohen’s Kappa index (k) was employed to evaluate the level of agreement among the susceptibility maps. The results showed an overall mean k value of 0.5; this constitutes a moderate level of agreement. These findings reinforce the potential of semi-automatic landslide inventories as a reliable basis for susceptibility modelling, particularly in scenarios where rapid responses are required after extreme events. Although semi-automatic approaches may still present limitations related to classification errors or the need for expert validation, they substantially reduce the time and effort needed to produce consistent inventories. Their integration with machine learning models demonstrates that, when properly constructed and validated, semi-automatic inventories can effectively support susceptibility assessments and contribute to more efficient hazard mapping and risk management strategies.