融合多源异构数据的滑坡变形阶段智能判识方法  

An Intelligent Identification Method of Landslide Deformation Stage Based on Multi-source Heterogeneous Data

作　　者：蒲未来 刘敦龙 桑学佳 张少杰[3] 陈乔[4] PU Weilai;LIU Dunlong;SANG Xuejia;ZHANG Shaojie;CHEN Qiao(School of Software Engineering,Chengdu University of Information Technology,Chengdu 610225,China;Sichuan Information Application Support Software Engineering Technology Research Center,Chengdu 636499,China;Institute of Mountain Hazards and Environment,Chinese Academy of Sciences,Chengdu 610044,China;Chongqing Institute of Green and Intelligent Technology,Chinese Academy of Sciences,Chongqing 400722,China)

机构地区：[1]成都信息工程大学软件工程学院,四川成都610225 [2]四川省信息化应用支撑软件工程技术研究中心,四川成都636499 [3]中国科学院、水利部成都山地灾害与环境研究所,四川成都610044 [4]中国科学院重庆绿色智能技术研究院,重庆400722

出　　处：《灾害学》2023年第4期179-186,共8页Journal of Catastrophology

基　　金：国家自然科学基金青年项目(42001100);四川省自然科学基金(2023NSFSC0751);四川省信息化应用支撑软件工程技术研究中心开放课题(760115027)。

摘　　要：针对滑坡体不同变形阶段的监测数据样本不均衡,样本扩充量的限定研究较少以及判识模型准确率较低等现实问题,该文提出了一种少数类样本全局扩充量测算方法以及将分类结果混淆矩阵与GSA相结合的基于遗传的多分类样本合成方法MCGSA,可避免产生大量的合成样本,且有效解决了样本不均衡问题;其次借助堆栈泛化思想以及具有较强知识挖掘能力的机器学习模型,结合滑坡体的多源异构监测数据,构建了基于stacking的滑坡变形阶段智能判识模型;最后将该模型应用在多个滑坡隐患点上进行现场实验测试,并进行了对比实验分析,分析结果显示该判识模型的准确率可达89%,F1宏平均值达到了74%。模型的判识结果可为区域内滑坡隐患点的预警信息发布提供辅助决策。In view of the problems of uneven monitoring data samples at different deformation stages of the landslide,few studies on the limitation of sample expansion and low accuracy of the identification model,we propose a global expansion measurement method for minority samples and a genetic based multi-classification sample synthesis method MCGSA,which combines classification result confusion matrix with GSA.The method firstly can avoid producing a large number of synthetic samples and effectively solve the problem of sample imbalance.Secondly,with stack generalization and machine learning model with strong knowledge mining ability,combined with multi-source heterogeneous monitoring data of landslides,the intelligent recognition model of landslide deformation phase based on stacking is constructed.Finally,the model is applied to several hidden landslide points for field test,and comparative experimental analysis is carried out.The analysis results show that the accuracy of the model is up to 89%,and the average F1 macro is up to 74%.The identification results of the model can provide an auxiliary decision for the warning information release of landslide hidden points in the region.

关 键 词：滑坡变形阶段 多源异构 全局扩充量测算 MCGSA样本合成 混淆矩阵 

分 类 号：X43[环境科学与工程—灾害防治] X915.5[天文地球—地质学] P694