Validation and Accuracy Assessment in Landslide Susceptibility Mapping: A Machine Learning Model Comparison

Developing reliable landslide susceptibility maps (LSM) represents a fundamental requirement for effective hazard mitigation, particularly within rapidly developing terrains. While receiver operating characteristic–area under the curve (ROC–AUC) metrics serve as standard benchmarks for assessing model performance, research demonstrates these measures alone provide insufficient validation for map dependability. This investigation examines the comparative performance of three machine learning approaches—logistic regression (LR), random forest (RF), and support vector machine (SVM)—for landslide hazard prediction within the mountainous terrain located east of Cairo, Egypt. The research utilized a balanced dataset comprising ۱۸۳ landslide occurrences and ۱۸۳ stable locations, identified through comprehensive field investigations and high-resolution satellite analysis. Fourteen predictor variables spanning topographic, geological, hydrological, anthropogenic, and triggering factor categories served as input parameters for LSM development. While all three algorithms demonstrated robust ROC–AUC performance (RF: ۰.۹۵, SVM: ۰.۹۰, LR: ۰.۸۸), supplementary evaluation using accuracy (ACC), recall, precision, F۱ score metrics, and spatial rationality assessment revealed substantial variations in model reliability. The RF algorithm emerged as the most dependable approach, exhibiting superior performance across all evaluation criteria and minimal classification errors in high-risk zones. Conversely, both SVM and LR models displayed elevated misclassification frequencies for hazardous and stable areas alike. These results emphasize that elevated ROC–AUC scores do not necessarily guarantee practical model reliability for landslide susceptibility assessment.

کلیدواژه ها:

Machine learning ، Random forest ، Landslide ، SHAP value ، ROC–AUC ، Cairo