Advanced landslide susceptibility mapping and analysis of driving mechanisms using ensemble machine learning models

Abstract

This study focuses on advanced landslide susceptibility mapping in Bertioga, utilizing ensemble machine-learning models to identify and predict landslide-prone regions. Four algorithms—Adaptive Boosting (AdaBoost), Gradient-Boosting Decision Tree (GBDT), Multilayer Perceptron (MLP), and Random Forest (RF)—were employed to model landslide susceptibility at a 30-m spatial scale, using thirteen landslide conditioning factors (LCFs). These LCFs include topographical, geological, and environmental variables significantly influencing landslide occurrence. The performance of each model was evaluated based on precision (P), recall (R), F1-Score, and area under the curve (AUC), ensuring a robust assessment of prediction accuracy. Furthermore, the models demonstrated varying area coverage in terms of susceptibility classes. For instance, MLP identified 15% of the study area as very low susceptibility, 22% as low, 18% as moderate, 18% as high, and 27% as very high. RF predicted 40% of the region as very low susceptibility, whereas GBDT indicated a substantial 45% of the area as very high risk. AdaBoost, on the other hand, assigned the highest moderate risk coverage at 31%. These results provide a comprehensive understanding of the spatial distribution of landslide risks. Results show that MLP and GBDT achieved higher accuracies in the susceptibility mapping, with AUC values ranging from 0.85 to 0.96. For instance, in Parque CauiBura, MLP and GBDT performed exceptionally well with AUC scores of 0.98 and 0.97, respectively. The average prediction for all cities yielded a high accuracy of 0.95 in Parque CauiBura, followed by 0.92 in Centro. These findings highlight the importance of using ensemble machine learning techniques in regional landslide susceptibility mapping, offering valuable insights for risk management and mitigation strategies in landslide-prone areas such as Bertioga.