Scale discrepancies in landslide susceptibility assessment – a comprehensive analysis in Southwestern China

Large-scale landslide susceptibility assessment and mapping (LSAM) is a vital tool for disaster mitigation and spatial planning. However, discrepancies often arise between the results of large-scale assessments and fine-scale evaluations within sub-regions, hindering scientific understanding and practical application of landslide risk management. In this study, an explainable XGBoost-SHAP model was developed to systematically explore how scale variation affects LSAM outcomes across Southwestern China and its four high landslide density regions. The results reveal that large-scale assessments can produce inaccurate predictions in certain areas, characterized by reduced model performance (e.g. lower ROC values) and overestimated landslide susceptibility index (LSI). SHAP-based interpretation further demonstrates that the contribution of conditioning indicators to landslide susceptibility varies significantly across regions. These findings indicate that indicator heterogeneity is a key driver of scale-related discrepancies in LSAM. This study highlights the importance of recognizing local uniqueness within large-scale assessments and provides methodological insight for improving the reliability and interpretability of regional LSA.