基于特征优选和逐步回归的黄土滑坡监测数据融合改进方法  

Improved Method for Fusion of Loess Landslide Monitoring Data Based on Feature Selection and Stepwise Regression

作　　者：王利[1,2,3] 张懿恺 舒宝 许豪 魏拓 雷体俊 WANG Li;ZHANG Yi-kai;SHU Bao;XU Hao;WEI Tuo;LEI Ti-jun(School of Geological Engineering and Geomatics,Chang'an University,Xi'an 710054,Shaanxi,China;State Key Laboratory of Geo-information Engineering,Xi'an 710054,Shaanxi,China;Key Laboratory of Western China's Mineral Resources and Geological Engineering of Ministry of Education,Chang'an University,Xi'an 710054,Shaanxi,China)

机构地区：[1]长安大学地质工程与测绘学院,陕西西安710054 [2]地理信息工程国家重点实验室,陕西西安710054 [3]长安大学西部矿产资源与地质工程教育部重点实验室,陕西西安710054

出　　处：《地球科学与环境学报》2023年第3期511-521,共11页Journal of Earth Sciences and Environment

基　　金：国家自然科学基金项目(41877289);国家重点研发计划项目(2021YFC3000503,2021YFC3000501)。

摘　　要：针对滑坡监测多源异构数据融合处理中存在的影响因子筛选难、结果差异大、数据处理复杂程度高等问题,提出一种基于最大互信息系数(MIC)、灰色关联分析(GRA)和逐步回归的黄土滑坡多源多点位异构监测数据融合方法。该方法首先将最大互信息系数和灰色关联分析结合起来,采用基于加权关联度的特征优选方法综合筛选滑坡变形影响因子,提取具有代表性的影响因子并剔除关联性差的影响因子;然后,通过逐步回归方法赋予各监测点位移和优选后的影响因子对应的重要性权重系数,获取多源异构数据融合序列;最后,采用甘肃黑方台党川滑坡监测设备所获取的全球卫星导航系统(GNSS)监测数据、裂缝位移计数据及气象数据进行实验验证。结果表明:在滑坡变形影响因子筛选性能方面,基于加权关联度的特征优选方法优于传统的Pearson相关系数法;基于特征优选和逐步回归的多源多点位异构数据融合模型的预测精度较传统的BP神经网络有所提升,其中均方根误差(RMSE)降低了51.8%,平均绝对百分比误差(MAPE)降低了2.26%,拟合优度达到了0.964。Aiming at the problems of difficult screening of influencing factors,large difference of results and high complexity of data processing in multi-source heterogeneous data fusion processing of landslide monitoring,a multi-source and multi-point heterogeneous monitoring data fusion method of loess landslide based on maximal information coefficient(MIC),grey relational analysis(GRA)and stepwise regression was proposed.Firstly,the maximum mutual information coefficient and grey correlation analysis are combined,and the feature optimization method based on weighted correlation degree is used to comprehensively screen the influencing factors of landslide deformation,extract representative features and eliminate factors with poor correlation.Secondly,the importance weight coefficient corresponding to the displacement of each monitoring point and the optimized influence factor is given by the stepwise regression method,and the multi-source heterogeneous data fusion sequence is obtained.Finally,the global navigation satellite system(GNSS)observation data,crack displacement meter data and meteorological data obtained by Dangchuan landslide monitoring equipment in Heifangtai of Gansu are used for experimental verification.The results show that the feature selection method based on weighted correlation degree is superior to the traditional Pearson correlation coefficient method in the screening performance of landslide deformation influencing factors;compared with the traditional BP neural network,the prediction accuracy of the multi-source and multi-point heterogeneous data fusion model based on feature selection and stepwise regression is improved;the root mean square error(RMSE)is reduced by 51.8%,the mean absolute percentage error(MAPE)is reduced by 2.26%,and the goodness of fit reaches 0.964.

关 键 词：黄土滑坡 变形监测 多源数据融合 多点位监测 加权关联度 特征优选 逐步回归 黑方台 

分 类 号：P228[天文地球—大地测量学与测量工程] P642.22[天文地球—测绘科学与技术]