融合空间物理参数的小样本极光涡旋自动分类研究  

Few shot classification of auroral vortices with space physics parameters

作　　者：韩冰[1] 吴墨豪 胡泽骏[2,3] 王平 张义生[4] HAN Bing;WU MoHao;HU ZeJun;WANG Ping;ZHANG YiSheng(School of Electronic Engineering,Xidian University,Xi'an 710071,China;MNR Key Laboratory for Polar Science,Polar Research Institute of China,Shanghai 200136,China;Ocean College,Zhejiang University,Zhoushan Zhejiang 316021,China;Institute of Applied Meteorology,Beijing 100029,China)

机构地区：[1]西安电子科技大学电子工程学院,西安710071 [2]中国极地研究中心自然资源部极地科学重点实验室,上海200136 [3]浙江大学海洋研究院,浙江舟山316021 [4]北京应用气象研究所,北京100029

出　　处：《地球物理学报》2023年第11期4478-4489,共12页Chinese Journal of Geophysics

基　　金：国家自然科学基金(62076190,41831072,41874195,42074199);国家重点研发计划(2021YFE0106400);陕西省重点研发计划项目(2022ZDLGY01-11);上海市科委项目(21DZ1206100);中国科学院空间科学先导专项(XDA15350202);空间环境地基综合监测网项目;应用气象研究所基础研究项目。

摘　　要：复杂的空间物理过程导致极光出现各类复杂的形态.极光形态的自动分类有助于分析极光发生机制以及了解空间物理过程.当前的极光分类主要采用传统的机器学习或深度学习的方法,需要大数据作为支撑.但在实际发生的极光事件中,一些事件相比其他极光事件出现频率小(如极光涡旋),这样的事件是小样本事件.本文提出一种基于注意力机制,结合空间物理参数信息的小样本极光事件分类方法.从北极黄河站2003—2017年全天空极光数据集中,构建了一个有85个极光涡旋序列的实验数据集.利用该方法,对极光涡旋这一类特殊小样本极光事件进行分类研究.研究结果显示,加入空间物理参数有助于极光涡旋的分类,分类准确率从56.37%提升到66.25%,同时也表明空间环境参数对极光涡旋的产生有着明显的调制作用.The complex morphological characteristics of auroras are closely related with the complex processes of space physics.The automatic classification of auroral morphology helps us to analyze the mechanism of auroral formation and understand the processes of space physics.Currently,traditional machine learning or deep learning methods are mainly adopted in the auroral classification,which need to be supported by large data.However,some specific aurora events appear less frequently,such as aurora vortices,and can be considered as small sample events compared with other aurora events.In this paper,it presents a few shot learning methods for aurora event classification based on attention mechanism and space physics parameter information,and a dataset with 85 auroral vortices events selected from 2003—2017 all-sky aurora dataset of the Chinese Yellow River Station in Arctic is constructed.The results show that space physics parameters is helpful for auroral vortices classification,the accuracy increases from 56.37%to 66.25%.In addition,the space environment parameters have obvious modulation effect on the occurrence of auroral vortices.

关 键 词：极光涡旋 空间参数 深度学习 小样本学习 多模态信息 

分 类 号：P353[天文地球—空间物理学]