LightGBM Delivers the Strongest Validation Performance in Wushi Landslide Mapping
The 2024 Wushi Ms 7.1 earthquake left a dense field of coseismic landslides across the Biedieli River basin, and the resulting susceptibility map now reads less like a theoretical exercise than a practical hazard ledger. Investigators compiled an inventory of 904 landslides and used nine conditioning factors spanning topography, geology, infrastructure proximity, and shaking intensity to compare four predictive models across training and validation datasets.
The clearest result emerged quickly: LightGBM outperformed the competing models on validation, with an AUC of 0.961, accuracy of 0.890, and F1-score of 0.895. Its training-to-validation AUC decay was only 2.6%, a small drop that points to stronger generalization than is often seen in data-driven landslide mapping, where models can memorize local patterns and then falter outside the training set.
Elevation, Shaking, and Road Access Shaped the Landslide Pattern
Three factors rose to the top of the ranking: elevation, peak ground acceleration, and distance to roads. Elevation carried the largest AM value at 0.6296, followed by PGA at 0.5810 and distance to roads at 0.4934. The pattern is consistent with a hillside failure process driven by terrain energy, earthquake forcing, and the geomorphic imprint of human access routes.
Steep, high terrain concentrates gravitational stress, while strong shaking raises the likelihood that already marginal slopes will fail. Roads often cut into slopes, disturb drainage, and create mechanically weakened margins that can fail under seismic loading. In a narrow mountain basin, those three ingredients can align along the same corridor, which helps explain why the high and very high susceptibility zones clustered along valley sides, steep slopes, and the vicinity of the Maidan fault zone.
The susceptibility hotspots occupied mid-elevation valley zones between 2036.5 and 2959 m, slopes steeper than 40°, and areas exposed to PGA greater than 0.48 g. In other words, the earthquake did not distribute instability evenly across the basin; it activated a fairly specific geomorphic and tectonic window where slope angle, seismic energy, and structural weakness overlapped.
Why Ensemble Learning Beat the Simpler Baselines
The comparison was not merely a contest between one model and another. It was a test of whether a boosting-based ensemble could better absorb nonlinear terrain-seismic relationships than a neural network, a decision tree, or a naive Bayes classifier working from the same inputs.
ANN followed closely behind LightGBM with an AUC of 0.957, accuracy of 0.876, and F1-score of 0.876. Decision tree performance was lower, with an AUC of 0.947, accuracy of 0.856, and F1-score of 0.860. Naive Bayes produced the lowest AUC at 0.919, though its accuracy and F1-score remained competitive at 0.862 and 0.869, respectively. The ranking matters because the AUC captures discrimination across thresholds, while accuracy and F1-score indicate how well the model classified susceptible and non-susceptible terrain in practical terms.
LightGBM’s advantage likely comes from the way boosting sequentially corrects errors, allowing the model to focus on residual structure that simpler learners can miss. Leaf-wise tree growth and gradient-based one-side sampling also make it more efficient in high-dimensional settings, which is useful when landslide controls are partly redundant and partly interacting. The region’s terrain is not linear, and neither is the failure mechanism; slopes, lithology, fault distance, drainage, roads, and shaking interact in combinations that are difficult for a rigid rule set to capture.
The Basin Geometry Explains the Concentration of Risk
The Biedieli River basin sits in a tectonically active part of the South Tianshan Mountains, an area shaped by intense structural deformation and steep relief. The earthquake struck close to the middle reaches of the river, and the landslides clustered on both sides of the valley, where incision has left slopes exposed and mechanically stressed.
A mid-elevation belt of fractured, steep terrain is especially vulnerable after strong shaking because slope material may already be close to failure. Once seismic acceleration exceeds local resistance, the rupture surface can mobilize quickly, particularly where roads, river undercutting, or fault proximity have already weakened the slope. The map therefore has value beyond hazard visualization; it points directly toward the terrain most likely to complicate post-earthquake emergency access, highway function, and slope stabilization.
The study area’s cold, arid mountain setting also matters. Sparse vegetation, limited precipitation, strong freeze-thaw cycling, and deeply dissected topography create slope conditions that differ from better-studied landslide belts in wetter, monsoon-influenced regions. Those environmental differences help explain why model transfer across mountain systems cannot be assumed.
Validation Results Offer a Stronger Basis for Emergency Planning
Across all four models, the high and very high susceptibility zones covered 10.51% to 19.43% of the basin. The range is wide enough to matter operationally: in a post-earthquake setting, a map that concentrates risk more tightly can support targeted inspection, road clearance, and slope monitoring, while a broader map may push emergency teams toward a more conservative deployment strategy.
Because the inventory contained 904 landslides and the data were split into 70% training and 30% validation sets, the evaluation was large enough to compare model behavior with some confidence, yet still small enough that overfitting remained a real concern. The small AUC drop from training to validation in LightGBM suggests that the model captured persistent terrain controls rather than noise tied only to the sampled landslides.
The practical value is immediate for the Biedieli Port Highway and other critical infrastructure running through the valley. Landslide susceptibility maps do not replace field inspection, but they narrow where inspection should begin, which slopes deserve priority, and which road segments may fail again after aftershocks or seasonal weathering.
Model Choice Matters in Mountain Earthquake Zones
This comparison shows that landslide susceptibility mapping in the Tianshan cannot rely on a single algorithmic habit imported from another mountain belt. A model that performs well in one geological setting may not hold up in another where climate, tectonics, and topography combine differently.
LightGBM offered the best overall balance of discrimination and generalization in this basin, while ANN remained close behind. Decision tree and naive Bayes provided useful benchmarks but did not match the ensemble’s validation strength. The study does not remove uncertainty from earthquake-triggered landslide prediction, and it does not substitute for detailed field reconnaissance, but it does give emergency planners a sharper picture of where the hillsides are most likely to fail next.
The remaining challenge is operational: translating this susceptibility pattern into timely mitigation for roads, settlements, and river corridors before the next strong shaking episode redraws the map again.