Electromagnetic Ion Cyclotron Waves Pattern Recognition Based on a Deep Learning Technique: Bag-of-Features Algorithm Applied to Spectrograms

Abstract

Several studies have shown the importance of electromagnetic ion cyclotron (EMIC) waves to the pitch angle scattering of energetic particles in the radiation belt, especially relativistic electrons, thus contributing to their net loss from the outer radiation belt to the upper atmosphere. The huge amount of data collected thus far provides us with the opportunity to use a deep learning technique referred to as the Bag-of-Features (BoF). When applied to images of magnetic field spectrograms in the frequency range of EMIC waves, the BoF allows us to distinguish, in a semi-automated way, several patterns in these spectrograms that can be relevant to describe physical aspects of EMIC waves. Each spectrogram image provided as an input to the BoF corresponds to the windowed Fourier transform of a ∼40 minutes to 1 hour interval of Van Allen Probes’ high time-resolution vector magnetic field observations. Our data set spans the 2012 September 8 to 2016 December 31 period and is at geocentric distances larger than 3 Earth radii. A total of 66,204 spectrogram images are acquired in this interval, and about 45% of them, i.e., 30,190 images, are visually inspected to validate the BoF technique. The BoF’s performance in identifying spectrograms with likely EMIC wave signatures is comparable to the visual inspection method, with the enormous advantage that the BoF technique greatly expedites the analysis by accomplishing the task in just a few minutes.