基于人工智能的张衡卫星闪电哨声波物理参数自动提取  

Automatic extraction of physical parameters of lightning whistler recorded by search coil magnetometer onboard Zhangheng satellite

作　　者：冯小康 袁静[1] 王桥 刘海军[1] 韩莹[1] 赵晨旭 丰继林[1] 申旭辉 泽仁志玛 黄建平 王亚丽 FENG XiaoKang;YUAN Jing;WANG Qiao;LIU HaiJun;HAN Ying;ZHAO ChenXu;FENG JiLin;SHEN XuHui;ZEREN ZhiMa;HUANG JianPing;WANG YaLi(Institute of Disaster Prevention,Langfang 065201,China;National Institute of Natural Hazards,Ministry of Emergency Management of China,Beijing 100085,China)

机构地区：[1]防灾科技学院,廊坊065201 [2]应急管理部国家自然灾害防治研究院,北京100085

出　　处：《地球物理学进展》2023年第6期2373-2391,共19页Progress in Geophysics

基　　金：河北省高等教育教学改革研究与实践项目(2021GJJG486);国家自然基金(41874174);中央高校基本科研基金项目(ZY20215155);防灾科技学院教研教改项目(JY2021A06);国际合作项目(the APSCO Earthquake Research Project Phase II and ISSI-BJ project)联合资助。

摘　　要：张衡卫星感应磁力仪(Search Coil Magnetometer)探测到了大量的低频波动事件,本文探索从中自动识别闪电哨声波(Lightning Whistler,LW)及其物理参数的算法,相关结果对进一步研究空间天气闪电事件的时空变化规律以及估计空间电子密度分布具有重要研究意义.该算法主要包含四部分内容:(1)预处理:以2 s的时间窗提取SCM原始波形数据,对其做重叠短时傅里叶变换得到时频图;(2)闪电哨声波轨迹识别和分割:整理时频图数据,对时频图中的LW轨迹区域进行逐像素点的语义标注;然后,搭建Unet神经网络,将标注数据输入该神经网络进行监督学习,得到LW轨迹分割模型;再应用该模型获得待检测时频图中所有的LW轨迹;(3)逐个闪电哨声波轨迹提取:采用闭运算增强LW轨迹的连通性同时断开不同LW轨迹之间的微弱连接;再采用种子填充法提取图像中的连通区域;然后计算连通区域的面积,删除小面积的连通区域,同时保留频率范围[500 Hz,5500 Hz]的区域;最后采用种子填充法提取不同的连通区域即可实现从所有的LW轨迹中逐个将LW轨迹一一提取出来;(4)闪电哨声波物理参数提取:根据每一条LW轨迹的时频位置获取原时频信号数据,并进行Eckersley公式拟合,拟合结果中二次项系数是LW的散射系数D0,其常数项为LW的初始触发时间t0.在600条闪电哨声波中开展实验(其中320条闪电哨声波用来训练Unet模型,180条LW做测试),得到实验结果是:本文算法能够自动提取180条闪电哨声波的物理参数,与人工提取的结果相当.Search Coil Magnetometer(SCM)onboard Zhangheng satellite has captured a large number of low-frequency wave events.This paper explores the algorithm for automatically identifying Lightning Whistler(LW)and its physical parameters.The results are of great significance for further studying the temporal and spatial variation of space weather lightning events and estimating the spatial electron density distribution.The algorithm mainly includes four parts:(1)Preprocess:SCM raw waveform data is extracted in a 2-second time window,and overlapping short time Fourier transform it to get the time-frequency diagram;(2)Segmentate the lightning whistler trace:collate the time-frequency map data,and annotate the LW trace area in the time-frequency map pixel by pixel;Then,build the U-net neural network,input the labeled data into the neural network for supervised learning,and obtain the LW trajectory segmentation model;Then use this model to detect and obtain all LW trajectories in the time-frequency diagram;(3)Extract the lightning whistler trace:use closed operation to enhance the connectivity of LW tracks and disconnect the weak connections between different LW tracks;use seed filling method to extract connected regions in images;calculate the area of the connected area,delete the small connected area,and keep the area of the frequency range[500 Hz,5500 Hz];Finally,use seed filling method to extract different connected regions,which can extract LW trace from all LW trace one by one;(4)Extract physical parameters of lightning whistler:According to the time-frequency position of each LW trace,obtain the original time-frequency signal data,and fitting it with Eckersley formula.In the fitting result,the coefficient of the quadratic term is the scattering coefficient D0 of LW,and its constant term is the initial trigger time t0 of LW(320 lightning whistlers were used to train the Unet model and 180 LWS were tested).The experimental results show that the algorithm can automatically extract the physical parameters of 180 lightning w

关 键 词：张衡卫星 Unet神经网络 闪电哨声波 散射系数 初始触发时间 

分 类 号：P352[天文地球—空间物理学]