A hybrid negative sampling strategy for landslide susceptibility assessment in mining area

Reliable delineation of non-landslide samples remains a challenge in data-driven landslide susceptibility assessment, particularly where landslide inventories are incomplete. This study proposes a hybrid negative sampling (HNS) strategy that integrates frequency ratio (FR)-based susceptibility pre-screening with SPY-based positive–unlabeled (PU) learning and differentiated instance weighting. Specifically, an initial pool of fuzzy negatives was derived from frequency ratio-based susceptibility indices, followed by the SPY algorithm to identify relatively high-confidence negative candidates. The integration of these two sources allows candidate negatives with different screening confidence levels to contribute to model training through distinct weights. Applied to Jiaokou County, a loess mining area in China (157 inventory-positive geological hazard records, 1,914 slope units), the methodology was evaluated using a common five-fold spatial holdout validation protocol with XGBoost and SVM, four baseline sampling strategies, and positive-to-negative ratios of 1:1 and 1:2. HNS achieved the highest mean ROC-AUC across all four classifier–ratio configurations, including 0.812 ± 0.041 for XGBoost and 0.797 ± 0.055 for SVM at 1:2, although its improvements over baseline strategies were modest and not statistically significant. Sensitivity and ablation analyses indicated that differentiated weighting primarily influenced the trade-offs among Recall, Precision, and Specificity, without consistently improving overall discrimination or area-based capture performance. Furthermore, SHAP analysis identified fractional vegetation cover (FVC), nighttime lights (NTL), and lithology as the leading contributors to model predictions rather than direct causal factors. Overall, HNS provides a structured approach to incorporating candidate negatives with different screening confidence levels, offering competitive predictive performance with application-dependent benefits and trade-offs.