Scholarlink

<p>The uncertainty of positive-to-negative sample ratios (P/N ratio) and the limitations of negative sample selection methods in landslide susceptibility assessment (LSA) affect the accuracy of the results. This study proposes a method for selecting negative samples that combines Bayesian optimization of the P/N ratio with geographical similarity stratified sampling, using Fusong County in Jilin Province as a case study. The method employs a Bayesian probabilistic model to optimize the P/N ratio according to the characteristics of the sample data. Geographic features are then weighted to calculate geographical similarity, facilitating stratified selection of non-landslide samples from different intervals. The proposed approach is validated using Random Forest (RF) and Convolutional Neural Network (CNN) models, and SHapley Additive exPlanations (SHAP) analysis is conducted to investigate the contributions of various factors. The results show that the new sampling method significantly improves the performance of the RF and CNN models, which achieves maximum ROC-AUC scores of 0.9611 and 0.9523, and PR-AUC scores of 0.9187 and 0.9001. This represents an approximate 3–5% improvement over traditional sampling methods. Distance to road is identified as a significant contributor to landslide occurrences. The landslide susceptibility map (LSM) generated using this approach more accurately identifies regional risks, providing a reliable sampling technique for high precision landslide assessment and prevention.</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

Landslide susceptibility assessment based on Bayesian optimization and geographical similarity

  • 期刊论文

  • 发表日期：

    2026年3月18日

查看全文
  • Jiayu Yan,
  • Shengwu Qin,
  • Wendi Rao,
  • Chaobiao Zhang,
  • Jiangfeng Lv,
  • Zhijun Lin,
  • Feng Wan,
  • Jiasheng Cao

摘要

The uncertainty of positive-to-negative sample ratios (P/N ratio) and the limitations of negative sample selection methods in landslide susceptibility assessment (LSA) affect the accuracy of the results. This study proposes a method for selecting negative samples that combines Bayesian optimization of the P/N ratio with geographical similarity stratified sampling, using Fusong County in Jilin Province as a case study. The method employs a Bayesian probabilistic model to optimize the P/N ratio according to the characteristics of the sample data. Geographic features are then weighted to calculate geographical similarity, facilitating stratified selection of non-landslide samples from different intervals. The proposed approach is validated using Random Forest (RF) and Convolutional Neural Network (CNN) models, and SHapley Additive exPlanations (SHAP) analysis is conducted to investigate the contributions of various factors. The results show that the new sampling method significantly improves the performance of the RF and CNN models, which achieves maximum ROC-AUC scores of 0.9611 and 0.9523, and PR-AUC scores of 0.9187 and 0.9001. This represents an approximate 3–5% improvement over traditional sampling methods. Distance to road is identified as a significant contributor to landslide occurrences. The landslide susceptibility map (LSM) generated using this approach more accurately identifies regional risks, providing a reliable sampling technique for high precision landslide assessment and prevention.