A Comparative Analysis of Decision Tree Models in Identifying Landslide Susceptibility and Type Classification

Abstract: Landslides pose a significant risk to human life and infrastructure, especially in Italy, which has a high frequency of landslide occurrences. To mitigate these hazards, Landslide Susceptibility Mapping (LSM) is crucial for identifying risk areas and developing appropriate mitigation strategies. Various methodologies have been adapted to perform LSM, with machine learning models seeing a rise in popularity due their predictive capabilities. The aim of this study is to compare two ensemble tree models, Random Forest (RF) and Extreme Gradient Boosting (XGB), in their predictive performances for landslide susceptibility by type. The typical methodology for assessing landslide type is performing susceptibility assessments individually for every class and aggregating the results. But with the RF and XGB models this process can be simplified by performing one multiclass analysis. The study found that the RF model significantly outperformed the XGB model in multiclass classification, with an overall accuracy of 95.83% compared to the 74.71% of the XGB model. No significant difference was found in the binary classification, with both models having an overall accuracy over 92%. The variables considered most important by both models were found to differ from heuristic models, suggesting a potential bias or incompleteness of the landslide inventory which should be considered in future studies. In conclusion, the RF model demonstrated its proficiency at making maps and high accuracy predictions for each landslide type.