Scholarlink

<p>With the growing prevalence of landslides, risk mitigation has become critically significant. Although landslide susceptibility assessment is widely implemented globally, predicting the potential impact of landslides on human societies remains a challenge. The impact of landslides can be determined by the probability of sediment release and runout distance. This study aimed to develop a framework for evaluating the potential damage caused by landslides in an urbanized area. First, landslide occurrence susceptibility maps were produced by Artificial Neural Network (ANN), Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). XGBoost is a supervised learning algorithm that implements gradient-boosted decision trees and is optimized for efficiency, scalability, and high predictive performance on large datasets. Second, an impact probability map was produced by the constrained random walk. Finally, the concept of vulnerability was integrated with the impact probability map to assess landslide risk, providing a more comprehensive understanding of landslide hazards and their potential effects on human societies. This study was conducted in an area of King County, Washington, USA, with an inventory of 2331 landslides derived from lidar imagery. Each landslide area was divided into source and deposit subareas. Fourteen parameters spanning geomorphology, geology, hydrology, and human activity were selected as conditioning factors for landslide initiation and were preprocessed to address multicollinearity before model construction. The landslide inventory map was randomly split into training (75%) and validation (25%) datasets to train the ANN, SVM, RF, and XGBoost models. Model performance was evaluated using metrics, such as the area under the curve (<i>AUC</i>). The results show that the XGBoost model achieved the best performance in susceptibility analysis with an overall accuracy of 0.90 and an <i>AUC</i> value of 0.96, followed by the RF (accuracy = 0.87 and <i>AUC</i> = 0.94), ANN (accuracy = 0.85 and <i>AUC</i> = 0.92), and SVM (accuracy = 0.85 and <i>AUC</i> = 0.90) models. The impact probability from the random walk method achieves an overall accuracy of 0.82 and an <i>AUC</i> of 0.88, demonstrating the method's effectiveness.</p><p></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

GIS-based machine learning models for assessing landslide impact in King County, Washington, USA

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月22日

查看全文
  • Di Lu,
  • Takashi Oguchi

摘要

With the growing prevalence of landslides, risk mitigation has become critically significant. Although landslide susceptibility assessment is widely implemented globally, predicting the potential impact of landslides on human societies remains a challenge. The impact of landslides can be determined by the probability of sediment release and runout distance. This study aimed to develop a framework for evaluating the potential damage caused by landslides in an urbanized area. First, landslide occurrence susceptibility maps were produced by Artificial Neural Network (ANN), Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). XGBoost is a supervised learning algorithm that implements gradient-boosted decision trees and is optimized for efficiency, scalability, and high predictive performance on large datasets. Second, an impact probability map was produced by the constrained random walk. Finally, the concept of vulnerability was integrated with the impact probability map to assess landslide risk, providing a more comprehensive understanding of landslide hazards and their potential effects on human societies. This study was conducted in an area of King County, Washington, USA, with an inventory of 2331 landslides derived from lidar imagery. Each landslide area was divided into source and deposit subareas. Fourteen parameters spanning geomorphology, geology, hydrology, and human activity were selected as conditioning factors for landslide initiation and were preprocessed to address multicollinearity before model construction. The landslide inventory map was randomly split into training (75%) and validation (25%) datasets to train the ANN, SVM, RF, and XGBoost models. Model performance was evaluated using metrics, such as the area under the curve (AUC). The results show that the XGBoost model achieved the best performance in susceptibility analysis with an overall accuracy of 0.90 and an AUC value of 0.96, followed by the RF (accuracy = 0.87 and AUC = 0.94), ANN (accuracy = 0.85 and AUC = 0.92), and SVM (accuracy = 0.85 and AUC = 0.90) models. The impact probability from the random walk method achieves an overall accuracy of 0.82 and an AUC of 0.88, demonstrating the method's effectiveness.