Optimizing data grouping strategies for landslide susceptibility assessment using frequency ratio-machine learning coupled models

This study systematically evaluates 48 combinations of FR–ML models and grouping strategies, including four coupled models (FR–Convolutional Neural Network (CNN), FR–Random Forest (RF), FR–Extreme Gradient Boosting (XGBoost), FR–Logistic Regression (LR)), three grouping methods (K-means clustering, natural breaks, quantiles), and four group numbers (5, 8, 12, 20). Results show that all FR–ML models achieve AUC values above 0.76, outperforming the individual ML models and confirming the enhanced predictive capability of the FR framework. Increasing the number of groups generally improves the performance of FR–ML model, while the optimal grouping method varies by model type. The DeLong test further indicates significant differences in AUC with increasing group numbers, highlighting the strong influence of grouping strategy. Landslide susceptibility maps suggest optimal group numbers of at least 5 for FR–CNN, more than 8 for FR–RF and FR–XGBoost, and at least 8 for FR-LR. Moreover, grouping strategy substantially affects the importance ranking of conditioning factors.