基于深度学习的极光亚暴时-空自动检测  

Spatio-temporal detection of auroral substorm based on deep learning

作　　者：杨秋菊[1] 任杰 向晗 YANG QiuJu;REN Jie;XIANG Han(School of Physics and Information Technology,Shaanxi Normal University,Xi′an 710061,China)

机构地区：[1]陕西师范大学物理与信息技术学院,西安710061

出　　处：《地球物理学报》2022年第3期898-907,共10页Chinese Journal of Geophysics

基　　金：陕西省自然科学基础研究计划(2020JM-272);中央高校基本科研业务费专项基金项目(GK202103020)资助。

摘　　要：准确快速地检测极光亚暴具有重要的意义.现有利用机器学习技术自动检测亚暴起始时刻的方法无法同时兼顾检测精度和效率.本文基于深度学习技术提出了一个端到端的亚暴起始检测模型,该模型利用双流卷积网络提取亚暴的时-空特征,并用三个一维时序卷积层获得亚暴起始的概率序列.该模型在Polar卫星1996—1998年极光观测上获得了87.5%的准确率和393帧/s的检测速度,且定位的亚暴发生位置与现有物理结论高度一致,可用于大规模亚暴事件时-空自动检测.It is important to detect auroral substorms accurately and quickly.Existing methods for automatically detecting substorm onset using machine learning techniques cannot take into account the detection accuracy and computational efficiency at the same time.In this paper,we propose a novel end-to-end substorm onset detection network(SODN)based on deep learning techniques.SODN extracts the spatial and temporal characteristics of auroral substorm using two-stream convolutional networks and obtains the probability sequence of substorm onset with three one-dimensional temporal convolutional layers.SODN achieves an accuracy of 87.5%and a detection speed of 393 frames per second on the auroral observations of Polar satellite from 1996 to 1998,and the estimated substorm onset locations are highly consistent with existing physical conclusions.Experimental results show that SODN can be used for large-scale spatio-temporal detection of auroral substorm events.

关 键 词：极光亚暴 双流卷积网络 时序卷积网络 时-空检测 

分 类 号：P352[天文地球—空间物理学]