Advanced machine learning techniques for enhanced landslide susceptibility mapping: Integrating geotechnical parameters in the case of Southwestern Cyprus

This study explores the potential enhancement of the performance of machine-learning-based landslide susceptibility analysis by the incorporation of key geotechnical parameters, namely Plasticity Index, Clay Fraction and Geological Strength Index (GSI), alongside geomorphological, geological, and hydrological factors. Utilizing geotechnical parameters, which are often overlooked in conventional probabilistic landslide susceptibility studies, can provide benefits, as they are directly related to the shear strength of the ground and the problem of slope stability. Herein, three methods, namely Logistic Regression, Random Forest and XGBoost are employed, to develop landslide susceptibility classifiers for the southwestern part of Cyprus, a region for which a detailed landslide inventory and geotechnical data are available. A dataset of 2500 landslide points and an equal number of non-landslide points were split into training (70%) and validation (30%) subsets. After processing the feature importance of 17 causal factors, lithology emerged as the most influential factor, followed by rainfall and land use, while GSI and plasticity index ranked sixth and seventh in the importance hierarchy. The capabilities of the three machine learning models were assessed and compared based on ROC curve analysis and 6 statistical metrics. Generally, the machine learning algorithms achieved high accuracy and predictive capability, succeeding in identifying more than 90% of the recorded landslides as areas of high to very high landslide susceptibility. The incorporation of geotechnical parameters resulted in modest but marked increase of statistical performance metrics.