基于集成学习建模的滑坡易发性评价  

Landslide Susceptibility Assessment Based on Ensemble Learning Modeling

作　　者：邬礼扬 曾韬睿 刘谢攀 郭子正 刘真意 殷坤龙[2] Wu Liyang;Zeng Taorui;Liu Xiepan;Guo Zizheng;Liu Zhenyi;Yin Kunlong(Hubei Center of Geological Disaster Control,Wuhan 430030,China;Faculty of Engineering,China University of Geosciences,Wuhan 430074,China;Institute of Geological Survey,China University of Geosciences,Wuhan 430074,China;School of Civil and Transportation Engineering,Hebei University of Technology,Tianjin 300401,China;China Railway Eryuan Engineering Group Co.,Ltd.,Chengdu 610031,China)

机构地区：[1]湖北省地质灾害防治中心,湖北武汉430030 [2]中国地质大学工程学院,湖北武汉430074 [3]中国地质大学地质调查研究院,湖北武汉430074 [4]河北工业大学土木与交通学院,天津300401 [5]中铁二院工程集团有限责任公司,四川成都610031

出　　处：《地球科学》2024年第10期3841-3854,共14页Earth Science

基　　金：国家重点研发计划项目(No.2018YFC0809402);国家自然科学基金项目(No.41877525)。

摘　　要：单一的机器学习模型往往难以满足滑坡易发性制图的需要,为提升滑坡易发性评价精度.提出了一种基于集成策略的机器学习模型组合优化的方法,以重庆市云阳县西部的12个乡镇为例进行滑坡易发性评价.首先,基于366处滑坡数据以及高程、坡度等9个指标因子构建易发性评价指标体系.然后以决策树模型(decision tree mode,DT)、逻辑回归模型(logistic regression,LR)和贝叶斯网络模型(bayesian network,BN)为基础模型,利用集成学习的三大类型,袋装法(bagging)、提升法(boosting)以及堆叠法(stacking)进行模型组合.并对各组合模型分别用粒子群算法(particle swarm optimization,PSO),贝叶斯算法(bayesian optimization,BO)进行超参数优化以及K最邻近算法(K-nearest neighbor,KNN)进行模型链接.最后采用ROC曲线与统计分析的方式来评估各集成学习模型精度.研究结果表明:与基础模型相比,三类集成学习模型精度均有提升,DT-LR-BN模型提升了3.5%~12.8%,RF模型提升了8%;BO-XGBoost模型提升了13.1%;KNN-stacking模型提升了7.4%~17%,BO-XGBoost模型的AUC值最高为0.811.集成学习能有效提升机器学习模型性能,提高滑坡易发性制图的精度,研究为机器学习模型之间的组合优化提供了新的思路与方法.A single machine learning model is often difficult to meet the needs of landslide vulnerability mapping,in order to improve the accuracy of landslide vulnerability assessment.In this paper,a method of machine learning model combination optimization based on integrated strategy is proposed,twelve townships in the west of Yunyang County,Chongqing were taken as an example.First,based on 366 landslide data and 9 index factors such as elevation and slope,the susceptibility evaluation index system was constructed.Then used the three algorithms of ensemble learning,bagging,boosting and stacking,to build combined models based on Decision Tree Mode(DT),Logic Regression(LR)and Bayesian Network(BN).The combined models used Particle Swarm Optimization(PSO),Bayesian Optimization(BO)for super parameter optimization and K-Nearest Neighbor(KNN)was used for model recombination.Finally,ROC curve and statistical analysis were used to calculate the accuracy of each integrated learning model.The research results show that compared with the basic classifier models,the accuracy of the three types of integrated learning models was improved.the DT-LR-BN model increased by 3.5%-12.8%,the RF model increased by 8%;the BO-XGBoost model increased by 13.1%;the KNN-stacking model increased by 7.4%-17%,and the AUC value of BOXGBoost model was the highest at 0.811.Integrated learning can effectively improve the performance of machine learning models,improve the accuracy of landslide susceptibility mapping,and provide a new idea and method for the combination optimization between machine learning models.

关 键 词：集成学习 袋装法 提升法 堆叠法 滑坡灾害 易发性评价 工程地质学 

分 类 号：P642.22[天文地球—工程地质学]