Spatial Heterogeneity of Dominant Controlling Factors for Seismic Landslide Susceptibility Zones: A Case Study of the Barkam Earthquake

ABSTRACT The 2022 Barkam earthquake triggered widespread landslide events, underscoring the urgent need for efficient and accurate susceptibility assessment and mechanism identification methods. This study focuses on the Ms 6.0 earthquake swarm zone in Barkam, constructing a susceptibility assessment framework based on 13 conditioning factors. Two ensemble learning models—Random forest (RF) and Extreme Gradient Boosting (XGBoost)—were compared, with XGBoost achieving superior performance (AUC = 0.893) over RF (AUC = 0.878). By establishing XGBoost sub‐models for zonal analysis and introducing SHapley Additive exPlanations (SHAP), the study quantified factor importance and response patterns across susceptible zones, identifying distinct spatial heterogeneity in dominant factors: high‐susceptibility zones are dominated by seismic activity (PGA > 0.3 g sharply increased susceptibility), precipitation (> 800 mm triggered strong effects), and slope (35°–45° as critical threshold), with vegetation (NDVI > 0.9) showing inhibitory effects; moderate susceptibility zones are controlled by topography‐vegetation coupling, where elevation (3300–4500 m) elevates susceptibility, and NDVI (0.3–0.8) exerted the strongest suppression, with precipitation > 770 mm also playing a role; low susceptibility zones are influenced by hydrological‐geomorphological factors, with elevation (> 4100 m) and SPI (> 0.5) promoting susceptibility, while gentle slopes showed suppressive effects. This study validates interpretable machine learning in susceptibility assessment and provides a scientific basis for zone‐specific disaster mitigation, enhancing the spatial precision of seismic landslide susceptibility management.