摘要
【目的】太空台风作为日地相互作用引发的典型现象,其在中高层大气中巨大的能量气旋往往伴随着极光现象的产生。通过对极光图像的识别可以帮助科学家寻找典型的太空台风事件,然而目前寻找事件主要依赖专家对极光图像的人工鉴别,较为低效。为解决上述问题,本文探索了基于深度学习的多波段图像联合目标检测方法,实现了对太空台风的事件识别与精准定位。【方法】本文利用DMSP/SSUSI的121.6 nm、135.6 nm、LBHS、LBHL四个波段极光图像识别太空台风事件,基于YOLOv8算法框架,引入了目标级融合与特征级融合策略,同时建立了单波段以及多波段融合的太空台风识别模型。【结果】在事件识别任务中,通过单波段基准模型与多波段融合模型的实验结果进行对比。结果显示特征级融合中的1216_LBHL组合表现最优,F1值达0.941;目标检测任务中,目标级融合中的1216_LBHL组合AP值最高,为0.917。【结论】特征级融合在太空台风事件识别中更具优势,目标级融合则更适用于目标检测任务,说明多波段互补性与融合策略的组合优化是提升检测性能的关键。
[Objective]As a typical phenomenon triggered by solar-terrestrial interactions,space hurricane often generates a huge energy cyclone in the middle and upper atmosphere,which is accompa-nied by the occurrence of auroral phenomena.Identifying auroral images can assist scientists in finding typical space hurricane events.However,currently,the search for such events mainly relies on experts'manual identification of auroral images,which is rather inefficient.To solve the above problems,this study explores a deep learning-based joint object detection method for multi-band images,achieving event recognition and precise localization of space hurricanes.[Methods]In this study,four-band auroral images(121.6nm,135.6nm,LBHS,LBHL)from DMSP/SSUSI are used to identify space hurricane events.Based on the YOLOv8 algorithm framework,target-level fusion and feature-level fusion strategies are introduced.Meanwhile,single-band and multi-band fusion models for space hurricane recognition are established.[Results]In the event recognition task,by comparing the experimental results of single-band baseline models and multi-band fusion models,it is shown that the 1216_LBHL combination in feature-level fusion performs best,with an F1 score of 0.941.In the object detection task,the 1216_LBHL combination in target-level fusion achieves the highest AP value of 0.917.[Conclusions]Featurelevel fusion demonstrates greater advantages in space hurricane event recognition,while target-level fusion is more suitable for object detection tasks.This indicates that the combined optimization of multi-band complementarity and fusion strategies is the key to enhancing detection performance.
作者
石珂
陆阳
陆盛
王勇
邹自明
SHI Ke;LU Yang;LU Sheng;WANG Yong;ZOU Ziming(University of Chinese Academy of Sciences,Beijing 100049,China;National Space Science Center,Chinese Academy of Sciences,Beijing 100190,China;National Space Science Data Center,Beijing 100190,China;Shandong University at Weihai,Weihai,Shandong 264209,China)
基金
国家重点研发计划“基础科研条件与重大科学仪器设备研发”重点专项(2022YFF0711400)
中国科学院“十四五”网络安全和信息化专项(CAS-WX2022SDC-XK15)
中国科学院网信专项(CAS-WX2022SF-0103)。
