Scholarlink

<p>Landslide susceptibility mapping (LSM) plays a critical role in the effective prevention and management of landslide hazards. Although machine learning-based LSM approaches have been widely applied in loess regions, few studies have incorporated regional-scale soil parameters (SPs), and the interpretability of these models remains limited. In this study, the Loess Plateau of western Shanxi (LPWS) was chosen as the study area, where regional-scale SPs—including cohesion, internal friction angle, permeability coefficient, and collapsibility coefficient—were integrated with conventional environmental variables. Three LSM models were developed using Support Vector Machine (SVM), Random Forest (RF), and Convolutional Neural Network (CNN). Their performances were evaluated using multiple metrics including AUC, ACC, F1-score, Kappa, and MIOU. To enhance interpretability, we propose a novel Weighted Multi-Model SHAP (WMM-SHAP) method that synthesizes feature attributions across the three models. The results indicate that: (1) incorporating SPs improved both predictive accuracy and generalization performance, with RF achieving the highest overall performance and SVM showing the most significant improvement; (2) WMM-SHAP identifies slope, elevation, distance to roads, and SPI as the dominant influencing factors, while SPs, although secondary, enhance predictive accuracy and help elucidate key physical mechanisms; and (3) dependency analysis reveals that the dominant factors exhibit Gaussian and Double Logistic relationships with landslide susceptibility, whereas SPs exhibit Exponential or Polynomial trends. The methodology proposed in this study and the resulting findings provide a scientific basis for land-use planning, site selection for major infrastructure, and the prioritization of landslide risk management strategies in the LPWS. It is important to emphasize that the model developed here is specifically calibrated and validated for the LPWS. Its direct transferability to other regions with distinct geotechnical and hydrological characteristics has not yet been established and warrants further investigation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable susceptibility mapping of loess landslides integrating soil parameters using machine learning and SHAP: a case study from the loess plateau in western Shanxi, China

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月21日

查看全文
  • Yongheng Zhou,
  • Yaming Tang,
  • Wei Feng,
  • Fan Feng,
  • Zhiheng Liu,
  • Fumeng Zhao,
  • Hongtao Wang,
  • Bo Hong,
  • Zepeng Wang

摘要

Landslide susceptibility mapping (LSM) plays a critical role in the effective prevention and management of landslide hazards. Although machine learning-based LSM approaches have been widely applied in loess regions, few studies have incorporated regional-scale soil parameters (SPs), and the interpretability of these models remains limited. In this study, the Loess Plateau of western Shanxi (LPWS) was chosen as the study area, where regional-scale SPs—including cohesion, internal friction angle, permeability coefficient, and collapsibility coefficient—were integrated with conventional environmental variables. Three LSM models were developed using Support Vector Machine (SVM), Random Forest (RF), and Convolutional Neural Network (CNN). Their performances were evaluated using multiple metrics including AUC, ACC, F1-score, Kappa, and MIOU. To enhance interpretability, we propose a novel Weighted Multi-Model SHAP (WMM-SHAP) method that synthesizes feature attributions across the three models. The results indicate that: (1) incorporating SPs improved both predictive accuracy and generalization performance, with RF achieving the highest overall performance and SVM showing the most significant improvement; (2) WMM-SHAP identifies slope, elevation, distance to roads, and SPI as the dominant influencing factors, while SPs, although secondary, enhance predictive accuracy and help elucidate key physical mechanisms; and (3) dependency analysis reveals that the dominant factors exhibit Gaussian and Double Logistic relationships with landslide susceptibility, whereas SPs exhibit Exponential or Polynomial trends. The methodology proposed in this study and the resulting findings provide a scientific basis for land-use planning, site selection for major infrastructure, and the prioritization of landslide risk management strategies in the LPWS. It is important to emphasize that the model developed here is specifically calibrated and validated for the LPWS. Its direct transferability to other regions with distinct geotechnical and hydrological characteristics has not yet been established and warrants further investigation.