Scholarlink

<p>Rainfall-induced clustered shallow landslides are difficult to predict because of their highly aggregated spatial patterns and short-duration triggering conditions. This study proposed a heterogeneous Stacking ensemble framework for landslide susceptibility assessment using 1969 shallow landslides triggered by the May 18, 2025 rainfall event in Gaozhou, Guangdong, China. Random Forest, Gradient Boosting Decision Tree, and XGBoost were used as base learners, while Logistic Regression served as the meta-learner. The Stacking model achieved the highest AUC of 0.952, outperforming RF, GBDT, and XGBoost by 5.0%, 4.0%, and 1.8%, respectively. It also identified more true positives, 542, than RF, GBDT, and XGBoost, which identified 531, 530, and 537, respectively. Susceptibility zoning and uncertainty analysis showed that the Stacking model provided more differentiated and reliable spatial predictions, with low uncertainty in low-susceptibility areas and relatively stable agreement in very-high-susceptibility zones. Feature importance analysis indicated that rainfall was the dominant factor, followed by slope and lithology. These findings demonstrate the effectiveness of heterogeneous ensemble learning for clustered shallow landslide susceptibility assessment in subtropical granite regions.</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

Ensemble learning-based susceptibility mapping of clustered shallow landslides triggered by the May 18, 2025 rainfall in Guangdong, China

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月24日

查看全文
  • Weijie Liao,
  • Xiaoyu Yi,
  • Fei Li,
  • Bin Qi,
  • Zhongliang Yang,
  • Kangchao Ning,
  • Xiaofei Zhong,
  • Wenkai Feng

摘要

Rainfall-induced clustered shallow landslides are difficult to predict because of their highly aggregated spatial patterns and short-duration triggering conditions. This study proposed a heterogeneous Stacking ensemble framework for landslide susceptibility assessment using 1969 shallow landslides triggered by the May 18, 2025 rainfall event in Gaozhou, Guangdong, China. Random Forest, Gradient Boosting Decision Tree, and XGBoost were used as base learners, while Logistic Regression served as the meta-learner. The Stacking model achieved the highest AUC of 0.952, outperforming RF, GBDT, and XGBoost by 5.0%, 4.0%, and 1.8%, respectively. It also identified more true positives, 542, than RF, GBDT, and XGBoost, which identified 531, 530, and 537, respectively. Susceptibility zoning and uncertainty analysis showed that the Stacking model provided more differentiated and reliable spatial predictions, with low uncertainty in low-susceptibility areas and relatively stable agreement in very-high-susceptibility zones. Feature importance analysis indicated that rainfall was the dominant factor, followed by slope and lithology. These findings demonstrate the effectiveness of heterogeneous ensemble learning for clustered shallow landslide susceptibility assessment in subtropical granite regions.