<![CDATA[Landslides constitute the predominant disaster type in Wonosobo Regency, Central Java, with 190 recorded events (62% of all disaster incidents) in 2024. Prior studies have reported inconsistent performance among machine learning algorithms due to variations in geographic context, dataset characteristics, feature selection, and evaluation frameworks. A review of the literature reveals a lack of comparative evaluations of Random Forest (RF), Support Vector Machine with Radial Basis Function kernel (SVM-RBF), and XGBoost that integrate Google Earth Engine (GEE), SHAP interpretability analysis, and three-class susceptibility classification specifically for Wonosobo Regency. This study aims to evaluate and compare the performance of these three algorithms in mapping landslide susceptibility using geospatial data. A total of 642 sample points (321 landslide and 321 non-landslide locations) were analyzed using eight conditioning factors extracted via GEE, seven of which were retained following Variance Inflation Factor (VIF) testing. Model performance was evaluated using accuracy, Cohen’s Kappa, Matthews Correlation Coefficient (MCC), and AUC-ROC with 10-fold stratified cross-validation. XGBoost achieved the highest performance on the study dataset, with an accuracy of 93.02%, an AUC-ROC of 0.974, and a cross-validation accuracy of 91.0% ± 3.1%. SHAP analysis indicated that rainfall made the largest contribution to the model predictions, highlighting its importance in the model’s decision-making process. Classification of prediction probabilities at the sample points using the Jenks Natural Breaks method yielded a susceptibility distribution of 49.2% Low, 6.4% Moderate, and 44.4% High.]]>
