摘要
准确开展崩塌滑坡灾害易发性评价是花岗岩分布区崩塌滑坡灾害防治的关键,可有效规避潜在风险。为探索崩塌滑坡灾害易发性评价模型的适用性及合理性,以河南罗山县花岗岩分布区为研究区,构建涵盖地形地貌、地层岩性、水文及土地利用四个层次14个因子的评价体系,采用逻辑回归(Logistic Regression,LR)模型、反向传播(Back Propagation,BP)神经网络、粒子群优化-反向传播(Particle Swarm Optimization-Back Propagation,PSO-BP)神经网络模型、类别提升(Categorical Boosting,CatBoost)、极限梯度提升(eXtreme Gradient Boosting,XGBoost)和轻量级梯度提升机(Light Gradient Boosting Machine,LightGBM)6种机器学习算法,开展崩塌滑坡灾害易发性评价研究。结果表明,BP模型(BP、PSO-BP)和Boosting模型(CatBoost、XGBoost、LightGBM)精度远大于LR模型,6种模型的AUC分别为0.95857、0.96055、0.98613、0.98606、0.98619和0.73905。LightGBM模型在AUC、频率比和野外验证中均表现最优,更适用于研究区崩塌滑坡灾害易发性建模,其极低、低、中、高和极高易发性区域面积占比分别为86.06%、2.76%、3.00%、3.50%和4.68%。6种算法的高及极高易发性分区结果基本一致,主要分布在人类活动强烈的山区。距道路距离、年平均降雨量、距断层距离、地貌和高程是诱发崩塌滑坡灾害的主控因子。研究旨在探究不同机器学习模型的精度和预测能力,优化建模思路,获取最优评价结果,为类似地区崩塌滑坡灾害的科学防治提供参考。
Accurate evaluation of landslide and collapse susceptibility is essential for disaster prevention in granite regions,helping to mitigate potential risks.This study examines the applicability and effectiveness of susceptibility evaluation models,focusing on the granite area of Luoshan County,Henan Province.An evaluation system was developed using remote sensing interpretation,field surveys,drone aerial photography,and literature analysis,incorporating 14 factors across four categories:topography,lithology,hydrology,and land use.Six machine learning algorithms—Logistic Regression(LR),Back Propagation(BP)neural network,Particle Swarm Optimization-Back Propagation(PSO-BP)neural network,Categorical Boosting(CatBoost),eXtreme Gradient Boosting(XGBoost),and Light Gradient Boosting Machine(LightGBM)—were employed to assess landslide and collapse hazard susceptibility.The study first defined classification criteria for the 14 factors,then generated susceptibility zoning results using these six models.Model performance was evaluated through the Area Under the ROC Curve(AUC),frequency ratio analysis,and field validation.The results indicate that BP-based models(BP,PSO-BP)and Boosting models(CatBoost,XGBoost,LightGBM)significantly outperformed the LR model,with AUC values of 0.95857,0.96055,0.98613,0.98606,0.98619,and 0.73905,respectively.Among them,the LightGBM model demonstrated the highest accuracy(AUC=0.98619).High-and extremely highsusceptibility zones identified by all models showed strong spatial consistency and were primarily concentrated in mountainous areas with intensive human activity.Key regions included southeastern Zhutang Township,central Lingshan Town,northern and central Tiepu Township,northern and southeastern Pengxin Town,central and southern Shandian Township,central Zhoudang Town,and northeastern/southwestern Dingyuan Township.Frequency ratios for high-susceptibility zones showed a positive correlation with AUC values,with LightGBM achieving the highest frequency ratio(18.43)for extremely high-susceptibility areas.Field validation confirmed that two newly identified landslide and collapse sites were located within the high-or extremely high-susceptibility zones predicted by the boosting models,with LightGBM demonstrating the best spatial alignment.The LightGBM model classified 86.06%,2.76%,3.00%,3.50%,and 4.68%of the study area into extremely low-,low-,moderate-,high-,and extremely high-susceptibility categories,respectively,which closely matched the actual hazard distributions.SHAP interpretability analysis identified the top five controlling factors for susceptibility as distance to roads,rainfall,distance to faults,geomorphology,and elevation,with road proximity being the most influential.This highlights the significant role of human engineering activities as the primary trigger.This study optimizes modeling approaches for susceptibility assessment and offers a scientific foundation for disaster prevention in similar granite regions.
作者
曹攀
陈婕
齐小帅
包峻帆
杨泽强
CAO Pan;CHEN Jie;QI Xiaoshuai;BAO Junfan;YANG Zeqiang(The Third Institute of Geology and Mineral Resources,Henan Bureau of Geology and Mineral Resources,Xinyang 464000,Henan,China;Henan Natural Resources Science and Technology Innovation Center(Information Perception Technology Application Research),Xinyang 464000,Henan,China;School of Geophysics and Spatial Information,China University of Geosciences(Wuhan),Wuhan 430000,China)
出处
《安全与环境学报》
北大核心
2025年第12期4751-4763,共13页
Journal of Safety and Environment
基金
河南省自然资源科研项目(2022-12,2023-8)。
关键词
公共安全
机器学习模型
花岗岩分布区
崩塌滑坡灾害
易发性
public safety
machine learning models
granite distribution areas
landslide and collapse hazards
susceptibility
