Scholarlink

<p>Machine learning has become a pivotal tool for geohazard susceptibility assessment; however, its performance is often constrained by the inherent scarcity and uncertain reliability of geohazard samples, a critical challenge in geoscience that can lead to model inaccuracies. To address this issue, this study introduces a Sample Enhancement Algorithm (SEA) to systematically improve training data quality. Using Guangzhou City as a case study, the SEA framework constructs a high-quality training dataset by quantifying sample reliability through a weighted environmental similarity approach based on prototype theory. The XGBoost model, trained on this enhanced dataset, demonstrates significant performance gains. For instance, in collapse scenarios, applying SEA with reliability thresholds of <i>R</i><sub><i>P</i></sub> ≥ 0.85 and <i>R</i><sub><i>N</i></sub> ≥ 0.5 improved model accuracy from 0.714 to 0.954 and AUC from 0.897 to 0.99. The study also reveals a critical trade-off, where overly strict reliability thresholds can reduce sample size and adversely affect model performance. Furthermore, validation against historical data highlights the impact of spatiotemporal heterogeneity on model generalization. Overall, these findings offer a quantitative, practical foundation for regional hazard risk management by underscoring the importance of balancing sample reliability with quantity and diversity to develop timely and generalizable models.</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

Enhancing geohazard susceptibility mapping with a sample reliability algorithm: a case study in Guangzhou, China

  • 期刊论文

  • 发表日期：

    2026年1月24日

查看全文
  • Yuchao Jiang,
  • Yan Gao,
  • Wenlong Li,
  • Rui Chen,
  • Zhiliang Zhang,
  • Quan Yuan,
  • Daixun He

摘要

Machine learning has become a pivotal tool for geohazard susceptibility assessment; however, its performance is often constrained by the inherent scarcity and uncertain reliability of geohazard samples, a critical challenge in geoscience that can lead to model inaccuracies. To address this issue, this study introduces a Sample Enhancement Algorithm (SEA) to systematically improve training data quality. Using Guangzhou City as a case study, the SEA framework constructs a high-quality training dataset by quantifying sample reliability through a weighted environmental similarity approach based on prototype theory. The XGBoost model, trained on this enhanced dataset, demonstrates significant performance gains. For instance, in collapse scenarios, applying SEA with reliability thresholds of RP ≥ 0.85 and RN ≥ 0.5 improved model accuracy from 0.714 to 0.954 and AUC from 0.897 to 0.99. The study also reveals a critical trade-off, where overly strict reliability thresholds can reduce sample size and adversely affect model performance. Furthermore, validation against historical data highlights the impact of spatiotemporal heterogeneity on model generalization. Overall, these findings offer a quantitative, practical foundation for regional hazard risk management by underscoring the importance of balancing sample reliability with quantity and diversity to develop timely and generalizable models.