融合雷达回波时空特征的下击暴流智能识别方法  

Downburst intelligent recognition method based on fusion of spatiotemporal characteristics of radar echo

作　　者：王兴[1] 吕晶晶[1] 周可 詹少伟 WANG Xing;LV Jingjing;ZHOU Ke;ZHAN Shaowei(National Demonstration Center for Experimental Atmospheric Science and Environmental Meteorology Education,Nanjing University of Information Science&Technology,Nanjing 210044,China;Nanjing Xinda Institute of Meteorological Science and Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学大气科学与环境气象国家级实验教学示范中心,江苏南京210044 [2]南京信大气象科学技术研究院,江苏南京210044

出　　处：《现代电子技术》2022年第4期23-29,共7页Modern Electronics Technique

基　　金：国家自然科学基金青年科学基金项目(41805033);国家自然科学基金面上项目(41675136);南京信息工程大学2020年度地球科学虚拟仿真实验教学课程建设项目(XNFZ2020A02)。

摘　　要：当前利用雷达等气象资料对下击暴流进行识别预警的准确性很低,因此文中提出一种以深度神经网络为基础的下击暴流智能识别方法。该方法以雷达回波时序图像和雷达径向速度场时序图像作为输入,将两种图像的时空四维特征融合到深度神经网络中,通过深度学习寻求“雷达回波图像和径向速度场图像”与“是否发生下击暴流”之间的函数映射关系。运用数据增强和损失函数优化技术,改善因样本数据不均衡所导致的识别结果偏向于样本中大概率事件的问题。再结合K折交叉验证,避免模型训练过程陷入局部极值。实验结果表明,文中方法对下击暴流识别的成功率达到95%,可实现对其识别预警的自动化,增强预报的时效性,同时该方法也适用于小尺度天气系统中对因下沉气流辐散所形成的大风的识别。As the accuracy of identifying and early warning downburst using radar and other meteorological data is very low,a downburst intelligent identification method based on depth neural network is proposed.In this method,the radar echo time⁃serial images and the radar radial velocity field time⁃serial images are taken as input,the spatiotemporal four⁃dimensional features of the two kinds of images are fused into the deep neural network,and the function mapping relationship between"radar echo image and radial velocity field image"and"whether downburst occurs"is sought by means of the deep learning.The data enhancement and loss function optimization techniques are used to improve the problem that the identification results tend to high probability events in the sample due to the imbalance of sample data.By combing with K⁃fold cross⁃validation,the problem of falling into local extremum in the process of model training is avoided.The experimental results show the success rate of downburst identification in the method proposed in this paper is 95%,it can realize the automation of identification and early warning,enhance the timeliness of prediction,and is also suitable for the small⁃scale weather systems to identify the strong wind caused by the divergence of downdraft.

关 键 词：智能识别 雷达回波 时空特征 下击暴流识别 特征融合 数据预处理 统计分析 交叉验证 

分 类 号：TN915-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]