陈丹璐, 孙德亮, 文海家, 辜庆渝. 基于不同因子筛选方法的LightGBM-SHAP滑坡易发性研究[J]. 北京师范大学学报(自然科学版), 2024, 60(1): 148-158. DOI: 10.12202/j.0476-0301.2023098
引用本文: 陈丹璐, 孙德亮, 文海家, 辜庆渝. 基于不同因子筛选方法的LightGBM-SHAP滑坡易发性研究[J]. 北京师范大学学报(自然科学版), 2024, 60(1): 148-158. DOI: 10.12202/j.0476-0301.2023098
CHEN Danlu, SUN Deliang, WEN Haijia, GU Qingyu. LightGBM-SHAP landslide susceptibility by different factor screening methods[J]. Journal of Beijing Normal University(Natural Science), 2024, 60(1): 148-158. DOI: 10.12202/j.0476-0301.2023098
Citation: CHEN Danlu, SUN Deliang, WEN Haijia, GU Qingyu. LightGBM-SHAP landslide susceptibility by different factor screening methods[J]. Journal of Beijing Normal University(Natural Science), 2024, 60(1): 148-158. DOI: 10.12202/j.0476-0301.2023098

基于不同因子筛选方法的LightGBM-SHAP滑坡易发性研究

LightGBM-SHAP landslide susceptibility by different factor screening methods

  • 摘要: 以重庆市黔江区为例,选取23个评价因子构建滑坡致灾因子数据库,利用地理探测器与皮尔逊-主成分分析2种因子筛选方法选择因子最优组合;基于Bayesian-LightGBM-SHAP混合模型进行滑坡易发性评价,并对模型精度进行验证,分析影响黔江区滑坡发生的主导因子.初始模型的AUC值为0.801,Pearson Correlation Coefficient-Bayesian-LightGBM模型AUC值为0.824,GeoDetector-Bayesian-LightGBM模型AUC为0.835;由因子重要性可知,多年平均降雨量、高程、POI核密度与距河流距离是滑坡发生的最主要因子,而输沙指数、水流动力指数与坡位对滑坡的发生影响较弱.因子筛选法-Bayesian-LightGBM相结合的混合模型能够提高模型的准确性,为构建合理因子数据库提供参考框架;通过与因子重要性的结合分析,验证了地理探测器能够准确探测各因子对滑坡发生的贡献值,突出各滑坡地理因子组合之间的相关性,从而探究各因子与滑坡之间的关系.

     

    Abstract: To study Qianjiang District of Chongqing Municipality, 23 evaluation factors are selected to construct a landslide disaster-causing factor database, optimal combination of factors is chosen utilizing two factor-screening methods: geodetector and Pearson-principal component analysis. Bayesian-LightGBM-SHAP hybrid model for landslide susceptibility evaluation was applied to verify model accuracy, and to analyze dominant factors affecting landslide occurrence in Qianjiang District. The initial model had an AUC value of 0.801, Pearson Correlation Coefficient-Bayesian-LightGBM model had an AUC value of 0.824, whereas GeoDetector-Bayesian-LightGBM model had an AUC of 0.835. Importance of factors was ranked. Average multi-year rainfall, elevation, POI kernel density and distance from rivers were found the most important factors for landslides to occur, while the sand transport index, hydrodynamic index and slope position had a weaker effect. The hybrid model combining factor screening method-Bayesian-LightGBM could improve accuracy of the model and provide a reference framework for constructing a rational factor database. Integrating analysis with factor significance verified that geo-probe could accurately detect contribution value of each factor to landslide occurrence. This highlights correlation between each combination of landslide conditioning factors, thus clarifies the relationship between each factor and landslides.

     

/

返回文章
返回