A data-driven evaluation of landslide susceptibility mapping in Jiaokou County, an ecologically sensitive loess coal-mining region in Shanxi Province, China, has assessed a hybrid negative sampling framework designed to tackle persistent uncertainty in training data labels. Reliable identification of non-landslide reference samples remains a structural challenge in spatial hazard prediction, particularly in settings where comprehensive field inventories are incomplete and unrecorded ground failures cannot be ruled out. The newly investigated approach couples empirical frequency ratio pre-screening with a positive–unlabeled learning algorithm known as SPY, incorporating candidate non-landslide units through differentiated training weights rather than treating all unrecorded terrain as verified negative instances.
Supervised spatial modeling typically relies on mapped historical failures as positive training units, but defining negative training samples across regional terrain poses a classic methodological dilemma. In common practice, land areas without documented failures are drawn at random, offset by geographic buffers, or selected through empirical susceptibility indices, yet none of these techniques independently confirm that an unlabeled slope has never experienced a failure. When unrecorded failures contaminate negative training pools, supervised classifiers risk learning distorted boundaries between environmental triggers and stable ground. Conversely, discarding all but the most conservatively screened candidate negatives risks limiting the diversity of terrain conditions available for model training.
To navigate this trade-off, the study conceptualized landslide inventory units under a positive–unlabeled learning perspective across Jiaokou County, which encompasses approximately 1,259.92 square kilometers across elevations ranging from 809 to 2,046 meters. The landscape is characterized by forested valleys in the west transitioning to fragmented loess hilly-gully terrain in the east, where extensive coal extraction introduces human-induced disturbances alongside fragile geological structures. Rather than utilizing arbitrary computational grid cells that disregard geomorphic boundaries, the researchers partitioned the study area into 1,914 individual slope units using curvature-based watershed segmentation applied to digital elevation models within ArcGIS Pro, establishing terrain-bounded mapping entities for spatial assessment.
The historical baseline dataset obtained from the National Geological Hazard Point Spatial Distribution Dataset encompassed 157 documented geological hazard locations across Jiaokou County, comprising 74 landslides, 58 collapses, 12 rockfalls, and 13 debris flows. These records were mapped into the segmented terrain, defining 157 inventory-positive slope units, while the remaining terrain units were designated as unlabeled background rather than verified negative ground. The authors noted that individual occurrence dates were not available in the archive and positional accuracy could not be independently verified through renewed field inspections, positioning the investigation strictly as a spatial susceptibility evaluation rather than an event-specific temporal hazard forecast.
To characterize the conditioning environment, the modeling framework compiled an initial suite of 18 predisposing factors spanning topography, geology, hydrology, ecology, and human activity indicators. Topographic metrics computed from 30-meter elevation models included slope, elevation, and synthetic curvature, which integrated profile and plan curvature components to eliminate mathematical redundancy. Hydrological parameters incorporated long-term June-to-September mean monthly precipitation averaged over the 2001 to 2020 period as a static hydroclimatic background, alongside topographic wetness, sediment transport, and distance to river networks. Ecological factors encompassed land cover and fractional vegetation cover, while anthropogenic influences were represented by distance to roads, distance to mining sites, and multi-year nighttime light observations spanning 2017 to 2022. Multicollinearity screening utilizing Variance Inflation Factor and Spearman correlation thresholds removed stream power index and individual curvature components, retaining 15 predictors with variance inflation factors strictly below 10.
The hybrid negative sampling workflow established an initial fuzzy negative candidate pool by calculating frequency ratio indices across the 15 conditioning factors using only the positive records within each designated training partition. Summed landslide susceptibility index scores were normalized and grouped into five natural break classes, with unlabeled units in the two lowest classes forming the fuzzy candidate pool. Next, a semi-supervised SPY procedure withheld a randomly selected 15 percent subset of known positive units as hidden spies while fitting a random forest screening classifier on the remaining positives and the initial fuzzy negatives. Unlabeled units generating positive-class probability scores strictly below the 1st percentile of the spy sample distribution were designated as high-confidence negatives. When assembling final training sets, high-confidence negatives received a sample weight of 1.0, whereas fuzzy negatives were assigned a differentiated weight of 0.5 to retain environmental diversity while dampening the influence of label uncertainty.
Predictive evaluation was structured around a five-fold spatial holdout validation design using K-means clustering of slope unit centroids, ensuring that negative sample screening, frequency ratio calculations, and hyperparameter tuning through Optuna occurred strictly within training partitions without leakage into held-out folds. Across both eXtreme Gradient Boosting and radial basis function support vector machine classifiers evaluated at positive-to-negative ratios of 1:1 and 1:2, the hybrid strategy achieved the highest mean area under the receiver operating characteristic curve. For XGBoost, the hybrid method recorded mean ROC-AUC values of 0.8135 ± 0.0440 at a 1:1 ratio and 0.8121 ± 0.0406 at a 1:2 ratio, while corresponding support vector machine configurations yielded 0.7849 ± 0.0566 and 0.7974 ± 0.0553. However, paired two-sided exact sign-flip tests across the five spatial folds revealed that these modest numerical margins over baseline methods, such as buffer-based sampling, did not achieve statistical significance at the 0.05 level.
Beyond overall curve metrics, detailed classification statistics highlighted operational trade-offs rather than uniform superiority across evaluation dimensions. Under an XGBoost classifier with a 1:2 sample ratio, the hybrid negative sampling framework delivered a higher recall of 0.820 compared to 0.766 for buffer-based sampling, successfully capturing a greater proportion of reference positive units. This gain was counterbalanced by lower precision of 0.145 compared to 0.169, lower specificity of 0.579 compared to 0.676, and an F1-score of 0.243 versus 0.272, reflecting an increased rate of classifying unlabeled background units as positive. Sensitivity ablations adjusting fuzzy sample weights between 0.25 and 1.00 produced minimal ROC-AUC shifts between 0.8112 and 0.8136, confirming that differentiated sample weighting primarily adjusted operating thresholds between sensitivity and conservative specificity rather than transforming absolute discriminatory capacity.
Spatial susceptibility mapping across Jiaokou County using common natural break cutoffs revealed distinct regional distributions among the sampling strategies. Across all models, low-susceptibility territory dominated the forested western sector, whereas elevated susceptibility clustered along the loess hilly-gully landscapes of the eastern district and central corridors, closely tracking the spatial concentration of documented hazard records. Random sampling produced an overly conservative map allocating only 16.7 percent of the county to the high-susceptibility tier, whereas SPY sampling designated an expansive 45.1 percent as very high susceptibility. The hybrid negative sampling strategy generated an intermediate configuration, classifying 35.1 percent of the territory as very low, 11.6 percent as high, and 35.8 percent as very high susceptibility, balancing spatial hazard capture without excessive territorial over-prediction.
Model interpretation conducted using SHapley Additive exPlanations on the trained hybrid XGBoost model illuminated the structural influence of individual conditioning factors on spatial predictions. Mean absolute SHAP values established fractional vegetation cover as the leading contributor, followed by nighttime lights, lithology, topographic wetness index, and regional elevation. Dependence distributions demonstrated that fractional vegetation cover exhibited a pronounced transition around 0.5, where lower canopy density contributed positively to predicted susceptibility and higher coverage exerted a stabilizing negative contribution. Elevated nighttime light values generally correlated with increased susceptibility predictions, acting as a regional spatial proxy for infrastructure and mining activity rather than a physical triggering mechanism. Similarly, intermediate slope-unit mean angles around 13 to 16 degrees contributed positively to risk scores, reflecting terrain aggregation characteristics across broad slope units rather than direct geotechnical failure thresholds.
The authors framed their conclusions within clear technical boundaries, emphasizing that statistical associations derived from machine learning should not be conflated with direct physical causation. A significant limitation remains the temporal mismatch between multi-year static indicators, such as nighttime light composites, and historical hazard records lacking exact failure timestamps, precluding dynamic triggering analyses. Furthermore, grouping diverse movement mechanisms ranging from rockfalls to debris flows under a generalized susceptibility umbrella models broad slope-failure propensity rather than distinct mechanical processes. Because truly verified negative terrain remains absent in unmonitored mining regions, statistical screening provides a structured method for balancing sample confidence and diversity, though its adoption must be weighed against its added complexity relative to simpler baseline sampling strategies.
Subject of Research: Landslide susceptibility modeling and negative sampling strategies in mining areas
Article Title: A hybrid negative sampling strategy for landslide susceptibility assessment in mining area
Article References: Xu, H., Liu, Z., & Li, S. (2026). A hybrid negative sampling strategy for landslide susceptibility assessment in mining area. Discover Geoscience, 4(1), Article 443. https://doi.org/10.1007/s44288-026-00821-x
Image Credits: AI Generated
DOI: 10.1007/s44288-026-00821-x
Keywords: landslide susceptibility, machine learning, negative sampling, positive-unlabeled learning, XGBoost, geohazards, spatial modeling, data-driven geology, hybrid, negative, sampling, strategy
Cite Scienmag News
Violet Maxwell. (November 9, 2026). Hybrid Sampling Strategy Evaluated for Mining Landslide Susceptibility Mapping. Scienmag. https://scienmag.com/hybrid-sampling-strategy-evaluated-for-mining-landslide-susceptibility-mapping/
Violet Maxwell. "Hybrid Sampling Strategy Evaluated for Mining Landslide Susceptibility Mapping." Scienmag, 9 November 2026, https://scienmag.com/hybrid-sampling-strategy-evaluated-for-mining-landslide-susceptibility-mapping/. Accessed 17 November 2026.
Violet Maxwell. "Hybrid Sampling Strategy Evaluated for Mining Landslide Susceptibility Mapping." Scienmag. November 9, 2026. https://scienmag.com/hybrid-sampling-strategy-evaluated-for-mining-landslide-susceptibility-mapping/








