Scholarlink

<p>Although stacking models have been widely applied in geological hazard susceptibility mapping, their increased predictive complexity introduces greater uncertainty compared to single models. In this study, we conducted a susceptibility assessment of land subsidence in Hangzhou using a carefully selected set of eight evaluation factors covering geological, hydrological, and human engineering aspects. By employing the SHapley Additive exPlanations (SHAP) tool, we achieved a detailed interpretation of feature contributions to subsidence occurrence and further analyzed the role of individual base models within the XGBoost-LightGBM-CatBoost ensemble model, which demonstrated the best performance in our study. The results indicate that engineering activities and soft soil distribution are the most influential factors in predicting land subsidence in Hangzhou. However, their contributions vary significantly between positive and negative samples, highlighting the necessity of integrating real-world considerations. Within the ensemble model, XGBoost and LightGBM act as conservative contributors focused on risk control, whereas CatBoost serves as the primary decision-making component. By using disaster prediction as an example, our study provides a replicable framework for both in-depth interpretation of decision-making processes in stacking ensemble models and disaster reduction in urban planning.</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

Explainable machine learning-based land subsidence susceptibility mapping: from feature importance to individual model contributions in ensembled system

  • 期刊论文

  • 发表日期：

    2025年5月20日

查看全文
  • Bofan Yu,
  • Huaixue Xing,
  • Weiya Ge,
  • Jiaxing Yan,
  • Yun-an Li

摘要

Although stacking models have been widely applied in geological hazard susceptibility mapping, their increased predictive complexity introduces greater uncertainty compared to single models. In this study, we conducted a susceptibility assessment of land subsidence in Hangzhou using a carefully selected set of eight evaluation factors covering geological, hydrological, and human engineering aspects. By employing the SHapley Additive exPlanations (SHAP) tool, we achieved a detailed interpretation of feature contributions to subsidence occurrence and further analyzed the role of individual base models within the XGBoost-LightGBM-CatBoost ensemble model, which demonstrated the best performance in our study. The results indicate that engineering activities and soft soil distribution are the most influential factors in predicting land subsidence in Hangzhou. However, their contributions vary significantly between positive and negative samples, highlighting the necessity of integrating real-world considerations. Within the ensemble model, XGBoost and LightGBM act as conservative contributors focused on risk control, whereas CatBoost serves as the primary decision-making component. By using disaster prediction as an example, our study provides a replicable framework for both in-depth interpretation of decision-making processes in stacking ensemble models and disaster reduction in urban planning.