Scholarlink

<p>Although coupled models integrating statistical analysis and machine learning are widely used in landslide susceptibility mapping, coupling strategy selection and model interpretability require in-depth investigation. Herein, a fundamental dataset pertaining to Liangshan Yi Autonomous Prefecture was constructed based on 2,681 landslide points and 14 environmental factors. Following factor screening, four statistical models [frequency ratio, information value, weights of evidence (WoE), and certainty factor models] were employed to quantify the environmental factors. The results were used as input for five machine learning models [nonlinear generalized additive, gated recurrent unit network, random forest (RF), support vector machine, and extreme gradient boosting models] to establish 20 coupled evaluation models. Model performance was assessed using the area under the receiver operating characteristic curve, F1 score, and landslide ratio. The Shapley additive explanations (SHAP) algorithm was used to interpret factor contribution mechanisms. Among the factor linkage methods, the WoE model achieved the highest prediction accuracy (mean F1 score = 0.935) and stability (standard deviation = 0.005). From the perspective of prediction models, the RF model yielded the optimal accuracy (mean F1 score = 0.942) and notable stability (standard deviation = 0.002). The WoE–RF coupled model was identified as the optimal solution for the study area. The SHAP algorithm effectively revealed factor operating mechanisms at global and local scales, successfully linking prediction results with landslide causation. Thus, this work provides a theoretical basis for constructing coupled models and optimizing the factor system.</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

Statistical–machine learning coupled models predicting landslide susceptibility in Liangshan Yi Autonomous Prefecture: performance comparison and factor explanations

  • 期刊论文

  • 发表日期：

    2026年4月8日

查看全文
  • Zhijie Ning,
  • Yongbo Tie,
  • heyi Hou

摘要

Although coupled models integrating statistical analysis and machine learning are widely used in landslide susceptibility mapping, coupling strategy selection and model interpretability require in-depth investigation. Herein, a fundamental dataset pertaining to Liangshan Yi Autonomous Prefecture was constructed based on 2,681 landslide points and 14 environmental factors. Following factor screening, four statistical models [frequency ratio, information value, weights of evidence (WoE), and certainty factor models] were employed to quantify the environmental factors. The results were used as input for five machine learning models [nonlinear generalized additive, gated recurrent unit network, random forest (RF), support vector machine, and extreme gradient boosting models] to establish 20 coupled evaluation models. Model performance was assessed using the area under the receiver operating characteristic curve, F1 score, and landslide ratio. The Shapley additive explanations (SHAP) algorithm was used to interpret factor contribution mechanisms. Among the factor linkage methods, the WoE model achieved the highest prediction accuracy (mean F1 score = 0.935) and stability (standard deviation = 0.005). From the perspective of prediction models, the RF model yielded the optimal accuracy (mean F1 score = 0.942) and notable stability (standard deviation = 0.002). The WoE–RF coupled model was identified as the optimal solution for the study area. The SHAP algorithm effectively revealed factor operating mechanisms at global and local scales, successfully linking prediction results with landslide causation. Thus, this work provides a theoretical basis for constructing coupled models and optimizing the factor system.