Machine learning reveals lithology and soil as critical parameters in landslide susceptibility for Petrópolis (Rio de Janeiro State, Brazil)

Abstract

Petrópolis, located in the mountainous region of Rio de Janeiro, Brazil, is frequently impacted by severe landslides, exacerbated by intense rainfall, steep topography, and unregulated urban growth. This study employs machine learning to assess and predict landslide susceptibility, integrating geological, hydrological, and anthropogenic factors. Five models—Random Forest, CatBoost, Support Vector Machine, Artificial Artificial Neural Network (ANN), and XGBoost—were evaluated, with CatBoost emerging as the optimal model (F1-score: 0.82; AUC-ROC: 0.88). Variable importance analysis revealed soil type and erodibility as critical soil parameters influencing susceptibility, alongside lithology, underscoring the significance of geological over purely topographic factors. These findings emphasize the utility of machine learning for landslide modeling, providing scalable methodologies applicable to similar geospatial risk assessments worldwide. Beyond local applications, this work offers actionable insights for urban planning and disaster risk management in mountainous urban regions.