Automatic Recognition and Localization of Poleward Moving Auroral Forms (PMAFs) From All‐Sky Auroral Videos

Abstract

Abstract Poleward Moving Auroral Forms (PMAFs) are one of the most common dayside auroral phenomena and are important for the study of dayside auroras and their dynamical processes. Accurate recognition and localization of PMAFs from a large number of all‐sky imager observations is the first critical step in PMAFs study, but is very difficult and tedious. This paper proposes an integrated model, namely RL‐PMAFs, for automatically recognizing whether an all‐sky auroral video contains PMAFs and, for the first time temporally locating PMAFs. RL‐PMAFs consists of a recognition network and a localization network. Taking the all‐sky auroral videos as input, the recognition network characterizes the morphology and motion of the aurora to determine whether the input videos contain PMAFs. Then, the feature sequences of the videos containing PMAFs are fed to the localization network to obtain the starting and ending times of PMAFs. RL‐PMAFs is evaluated using auroral observations at Arctic Yellow River Station from 2005 to 2007. RL‐PMAFs not only yields higher recognition accuracy of 91.67% than previous methods, but also achieves a precision of 81.90% and a recall of 79.62% for locating PMAFs in auroral videos. The experimental results show that it is a valuable attempt of artificial intelligence for space physics.

Keywords

• aurora
• PMAFs
• deep learning
• automatic recognition
• temporal localization