三种机器学习方法在广东雷暴大风自动识别的应用效果评估  

THE ASSESSMENT OF APPLICATION EFFECTIVENESS OF THREE MACHINE LEARNING METHODS IN AUTOMATIC IDENTIFICATION OF THUNDERSTORM GALE IN GUANGDONG

作　　者：兰宇 罗聪[1] 伍志方[1] 唐思瑜 吴林 程兴国[2] 陈蝶聪 LAN Yu;LUO Cong;WU Zhifang;TANG Siyu;WU Lin;CHENG Xingguo;CHEN Diecong(Guangdong Meteorological Observatory,Guangzhou 510641,China;South China University of Technology,Guangzhou 510641,China;Guangdong Ecological Meteorology Center,Guangzhou 510641,China)

机构地区：[1]广东省气象台(南海海洋气象预报中心),广东广州510641 [2]华南理工大学,广东广州510641 [3]广东省生态气象中心,广东广州510641

出　　处：《热带气象学报》2023年第2期256-266,共11页Journal of Tropical Meteorology

基　　金：国家重点研发计划项目(2019YFC1510400);广东省基础与应用基础研究基金项目(2021B1515310001);国家自然科学基金项目(U2142210);广东省气象局科学技术研究项目(GRMC2021M09);广东省自然科学基金(2022A1515011814);南海海洋气象预报预警关键技术研发(CXFZ2021J025)共同资助。

摘　　要：基于2012—2019年自动站雷暴大风观测实况和对应雷达回波,利用传统机器学习方法(决策树)和深度学习方法(CNN、YOLO)等三种机器学习方法分别建立雷暴大风自动识别模型。根据广东雷暴大风回波特征,选取50dBZ高度、反射率因子强度梯度等5个回波参量作为决策树的特征因子;将1~9km高度的雷达回波分为11层,作为YOLOv3的输入层,使其由原3个特征层扩展到11层,训练优化后的YOLOv3可更合理刻画雷暴大风的空间结构特征。经批量测试和业务试运行试验,检验结果表明:三种模型中基于决策树的模型虚警最高,基于CNN的模型漏报最多,基于YOLO的模型识别效果最好,其POD和CSI均最高。通过对广东2020年汛期5次系统性和5次局地性雷暴大风过程进行分类型自动识别效果评估,并选取任意天气下长达30天连续时段进行不间断识别检验,结果表明该算法对于不同类型的雷暴大风均有较好的识别能力,具备业务化应用前景。On the basis of radar data and thunderstorm gale observation in Guangdong from 2012 to 2019,three machine learning algorithms,including a traditional machine learning algorithm(Decision Tree)and two deep learning algorithms(CNN and YOLO),were applied to establish automatic identification models for thunderstorm gale respectively.In line with the radar echo characteristics of thunderstorm gale in Guangdong,six echo parameters,such as 50 dBZ height and reflectivity factor gradient,were selected as the characteristic factors of the decision tree.In addition,the heights of 1 to 9 km were divided into 11 radar echo input layers of YOLOv3,expanding from the original three characteristic layers,which described the spatial structure characteristics of thunderstorm gale more reasonably after optimization.A series of business run-in tests indicated that among the three identification models,the model using the decision tree presents the highest FAR and the one based on CNN misses the most identifications,while the identification model based on YOLO behaves the best,with the highest POD and CSI.By evaluating the automatic identification effects of five systematic and five local thunderstorm gale processes during the rainy season of 2020 in Guangdong,and selecting continuous periods of up to 30 days under any weather condition for ongoing identification tests,the results reveal that the YOLO algorithm possesses good recognition ability for different types of thunderstorm gale with considerable business application prospect.

关 键 词：雷暴大风 自动识别 机器学习 雷达回波 深度学习 

分 类 号：P425.47[天文地球—大气科学及气象学]