Scholarlink

<p>Recent studies have highlighted the potential of machine learning in landslide hazard analysis, leveraging their strong feature extraction capabilities for task execution. Nevertheless, existing approaches often regard landslide sample points as spatially uncorrelated isolated individuals, neglecting the correlation of their spatial positions. To address these gaps, this study introduced spatial autocorrelation(SA) information to characterize these features and investigated the enhancement effects on predictive performance for landslide occurrence. To be specific, we&#xa0;propose&#xa0;introducing&#xa0;SA information&#xa0;into sample’s environmental features to evaluate&#xa0;its improvement on landslide prediction model’s performance, using 9046 landslides points in Yunnan Province (China) as dataset, with data accumulated up to around 2019. We evaluated the improvement effect of SA information added as extra samples’ features on 4 traditional machine learning methods, with local Moran’s index characterizing this kind of information. Experimental results reveal that the convolutional neural network (CNN) model&#xa0;achieves&#xa0;a&#xa0;performance improvement after incorporating SA, evidenced by the increasing of binary classification (landslide or non-landslide) accuracy(from 65.20 to 96.74%), recall(from 67.20 to 95.49%), precision(from 63.19 to 97.72), F1(from 65.13 to 96.59%) and AUC(from 0.71 to 0.99). Moreover, a newly proposed two-dimensional visualization method demonstrates that the quantitative predicted values from the enhanced model exhibit closer alignment with the true labels of the sample data. Consistent improvements are also replicated in 3 popular machine learning models: logistic regression (LR), decision tree (DT) and support vector machine (SVM), validating&#xa0;the generalizability of this approach. This work substantiates that SA information can enhance the model's ability to accurately differentiate between landslide points and non-landslide points, offering a methodological reference for spatial modeling in other landslide-prone areas.</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

A hybrid framework for landslide prediction integrating machine learning and sample enhanced by spatial autocorrelation

  • 期刊论文

  • 发表日期：

    2025年9月8日

查看全文
  • Longyin Ding,
  • Wenliang Jiang,
  • Qiang Li,
  • Wenhao Shen

摘要

Recent studies have highlighted the potential of machine learning in landslide hazard analysis, leveraging their strong feature extraction capabilities for task execution. Nevertheless, existing approaches often regard landslide sample points as spatially uncorrelated isolated individuals, neglecting the correlation of their spatial positions. To address these gaps, this study introduced spatial autocorrelation(SA) information to characterize these features and investigated the enhancement effects on predictive performance for landslide occurrence. To be specific, we propose introducing SA information into sample’s environmental features to evaluate its improvement on landslide prediction model’s performance, using 9046 landslides points in Yunnan Province (China) as dataset, with data accumulated up to around 2019. We evaluated the improvement effect of SA information added as extra samples’ features on 4 traditional machine learning methods, with local Moran’s index characterizing this kind of information. Experimental results reveal that the convolutional neural network (CNN) model achieves a performance improvement after incorporating SA, evidenced by the increasing of binary classification (landslide or non-landslide) accuracy(from 65.20 to 96.74%), recall(from 67.20 to 95.49%), precision(from 63.19 to 97.72), F1(from 65.13 to 96.59%) and AUC(from 0.71 to 0.99). Moreover, a newly proposed two-dimensional visualization method demonstrates that the quantitative predicted values from the enhanced model exhibit closer alignment with the true labels of the sample data. Consistent improvements are also replicated in 3 popular machine learning models: logistic regression (LR), decision tree (DT) and support vector machine (SVM), validating the generalizability of this approach. This work substantiates that SA information can enhance the model's ability to accurately differentiate between landslide points and non-landslide points, offering a methodological reference for spatial modeling in other landslide-prone areas.