随机森林机器学习方法用于冻雨现象自动识别的试验研究  

Experimental Study on Automatic Recognition of Freezing Rain by Random Forest Machine Learning Method

作　　者：王小兰[1,2] 仇建华 李翠翠 陶法 梁静舒 秦建峰 WANG Xiaolan;QIU Jianhua;LI Cuicui;TAO Fa;LIANG Jingshu;QIN Jianfeng(Meteorological Observation Center of CMA,Beijing 100081;Engineering Research Center of CMA,Beijing 100081;Nanjing University of Information Science and Technology,Nanjing 210044;Guizhou Meteorological Data Center,Guiyang 550002;Wuhan Meteorological Bureau,Wuhan 430074)

机构地区：[1]中国气象局气象探测中心,北京100081 [2]中国气象局工程技术研究中心,北京100081 [3]南京信息工程大学,南京210044 [4]贵州省气象数据中心,贵阳550002 [5]武汉市气象局,武汉430074

出　　处：《气象科技》2025年第2期191-200,共10页Meteorological Science and Technology

基　　金：中国气象局创新发展专项(CXFZ2024J057);中国气象局气象探测中心2023年观测试验计划(GCSY23-15)资助。

摘　　要：现行气象观测业务中尚缺乏对冻雨天气现象的自动监测,研究引入随机森林机器学习方法,利用Ka波段毫米波云雷达的观测数据,建立了冻雨现象的自动识别方法,为弥补观测业务中冻雨的自动识别和连续观测的缺乏提供一种可能。研究首先分析了2024年2月武汉地区的2次雨雪冰冻天气过程中不同降水现象(雨、冻雨、雪和雨夹雪)的Ka波段毫米波云雷达的回波强度、偏度值的分布特征,发现在数值范围和垂直高度分布上存在显著差异,由此确定回波强度、偏度及近地面气温作为识别变量。针对武汉地区、贵州地区多个雨雪冰冻过程分别建立RF机器学习方法的冻雨现象自动识别模型,经训练和验证计算,测试识别准确率(A _(cc))均超90%、冻雨命中率均超80%,使用独立的冻雨实例进行检验,检验A _(cc)可达到80%。与现行观测业务中电线积冰人工观测比较,该方法可以自动连续地识别分钟级冻雨现象,具备业务应用可行性。由于冻雨发生时的毫米波云雷达回波强度和偏度的地域特征明显,需要使用不同地区的冻雨样本数据建立针对不同地区的识别模型,扩充样本和优化模型参数及指标,可以进一步提升该方法的识别准确率,降低虚警率。There is still a lack of automatic monitoring for freezing rain in the current meteorological observation field.The Random Forest(RF)machine learning method is introduced here,and an automatic recognition method for freezing rain is established using the data from Ka-band Millimetre-Wave Cloud Radar(MWCR).This provides a possibility to make up for the lack of automatic recognition and continuous observation of freezing rain.Firstly,the distribution characteristics of echo intensity and skewness values of Ka-band MWCR with different precipitation phenomena(rain,freezing rain,snow,and mixed rain and snow)during two freezing weather processes occurring in the Wuhan area in February 2024 are analysed.Significant differences in the value range and vertical height distribution are found.Then,echo intensity,skewness values,and near-surface temperature are determined as identification variables.Automatic recognition models for freezing rain using the RF machine learning method are established for several freezing processes in Wuhan and Guizhou respectively.After training and verification calculations,the test recognition accuracy rate(A _(cc))exceeds 90%,and the freezing rain hit rate(P _(od))exceeds 80%.Independent freezing rain examples are used for verification,and the verifying A _(cc) reaches 80%.Compared with wire icing observation,this method can automatically and continuously identify freezing rain phenomena in minutes,which is also feasible for business application.Since the echo intensity and skewness values of MWCR during the freezing rain process have obvious regional characteristics,freezing rain sample data from different regions should be collected to establish recognition models for different regions.The recognition accuracy can be improved by expanding samples and optimising model parameters and indicators,as well as reducing the False Alarm Rate(F _(ar)).

关 键 词：冻雨 自动识别 机器学习 随机森林 毫米波云雷达 

分 类 号：P412[天文地球—大气科学及气象学]