基于复合机器学习模型的无线微波降雨反演研究  

Research on Wireless Microwave Rainfall Inversion Based on Composite Machine Learning Models

作　　者：张驰 吴兵[4] 洪岱[5] 宋莹 郑鑫 陈渝青[1,2,3] 尧俊辉 谢彪[7] 王冲 杨涛 ZHANG Chi;WU Bing;HONG Dai;SONG Ying;ZHENG Xin;CHEN Yuqing;YAO Junhui;XIE Biao;WANG Chong;YANG Tao(State Key Laboratory of Hydrology-Water Resources and Hydraulic Engineering,Hohai University,Nanjin2g10098,China;Yangtze Institute for Conservation and Development,Nanjingg210098,China;College of Hydrology and Water Resources,Hohai University,Nanjing 210098,China;Huai'an Water Conservancy Survey and Design Institute Co.,Ltd,Huai'an 223005,China;China Information Consulting&Designing Institute Co.,Ltd,Nanjing 210098,China;Poyang Lake Hydrology and Water Resources Monitoring Center,Nanchang 330002,China;Jiangxi Hydrological Monitoring Center,Nanchang 330002,China;Water Resources Affairs Development Center of Jining Municipality,Jining 272100,China)

机构地区：[1]河海大学水文水资源与水利工程科学国家重点实验室,江苏南京210098 [2]长江保护与绿色发展研究院,河海大学,江苏南京210098 [3]河海大学水文水资源学院,江苏南京210098 [4]淮安市水利勘测设计研究院有限公司,江苏淮安223005 [5]中通服咨询设计研究院有限公司,江苏南京210019 [6]鄱阳湖水文水资源监测中心,江西南昌330002 [7]江西省水文监测中心,江西南昌330002 [8]济宁市水利事业发展中心,山东济宁272100

出　　处：《水文》2023年第6期33-39,共7页Journal of China Hydrology

基　　金：国家自然科学基金资助项目(51879068);国家重点研发计划资助项目(2018YFC0407906)。

摘　　要：为实现精细化的无线微波水文气象密集监测,提出基于无线微波链路的降雨增强反演方案。利用极值梯度提升(XGBoost)算法进行干湿判别,并结合支持向量回归机(SVR)和高斯过程回归(GPR)两种机器学习算法,构建XG_SVR与XG_GPR复合反演模型,与传统ITU-R模型反演结果比较分析。实验结果表明,XGBoost算法的干湿判别结果优于滑动标准差法,平均分类准确率为88%,1 h以下高时间分辨率的准确率可达90%以上;两种复合模型整体反演效果良好,XG_SVR与ITU-R模型均适用于高时间分辨率,相关性系数均值皆高于0.80;而XG_GPR模型在反演1 h及以上低时间分辨率降雨时有显著优势,相关性系数为0.95左右,且均方误差远小于其他两种模型。利用复合机器学习模型改进传统无线微波测雨的方案具有可行性及发展潜力。In order to realize the refined wireless microwave hydrometeorological intensive monitoring,a rainfall enhancement inversion scheme based on wireless microwave links was proposed.The eXtreme Gradient Boosting(XGBoost)algorithm was used to discriminate between dry and wet periods.It combined two machine learning algorithms,namely support vector regression(SVR)and Gaussian process regression(CPR),to construct a composite inversion model of XG_SVR and XG_GPR,then compared the inversion results with those of traditional ITU-R model.The results show that the discrimination of the XGBoost algorithm is better than those of the sliding standard deviation method,and the average classi-fication accuracy of XGBoost algorithm is 88%,the accuracy of high time resolution less than 1 h can reach more than 90%.The overall inversion results of the two composite models are good,both XG_SVR and ITU-R models are suitable for high time resolution,and the average correlation coefficient are higher than O.80.The XG_GPR model has a significant advantage in retrieving rainfalls over of 1 h and above low time resolution rainfall,the correlation coefficient is about 0.95,and the mean square errors are much smaller than the other two models.Using composite ma-chine learning model to improve the traditional wireless microwave rainfall measurement scheme has feasibility and development potential.

关 键 词：无线微波 降雨 雨衰关系 机器学习 干湿期判别 

分 类 号：P412.13[天文地球—大气科学及气象学] TV11[水利工程—水文学及水资源] P332