Detection of UHR Frequencies by a Convolutional Neural Network From Arase/PWE Data

Abstract

We have developed the automatic detection scheme for upper hybrid resonance (UHR) frequency using a convolutional neural network (CNN) from the electric field spectra obtained by the plasma wave experiment (PWE) aboard Arase. In this paper, we investigate the practical capability of this scheme in terms of actual scientific use case. We find that the average error rate is below 7.8% when the wave frequency is above 30 kHz and the wave spectral intensity is less than 10 −5 mV 2/m 2 /Hz. About 91% of the data obtained by the high-frequency analyzer (HFA) aboard the Arase satellite satisfies these conditions. To improve the accuracy of the determined UHR frequencies in a wide frequency range, we used another receiver, the onboard frequency analyzer (OFA), which enables us to detect low-frequency UHR emissions. We confirmed that the averaged error rate derived by the OFA spectra becomes better than that derived from the HFA spectra in a frequency range below 20 kHz. We report the performance of the UHR frequency determination under the different geomagnetic conditions. We find that the UHR frequency can be determined with good accuracy using the CNN from the frequency-time diagram both during geomagnetically quiet and disturbed conditions. We conclude that the CNN-based UHR frequency determination is a reliable method to derive the electron density along the satellite orbit through observations of UHR frequencies, and this method contributes to studies on dynamics of the plasmasphere. Plain Language Summary Determining the upper hybrid resonance (UHR) frequency is the most popular way to determine the quantitative electron density in space. The high-frequency analyzer (HFA) and onboard frequency analyzer (OFA) aboard the Arase satellite measure electric field spectra at high-and low-frequency ranges. The nominal time resolutions of the HFA and OFA spectra are 8 and 1 s, respectively. Determining the UHR frequency by conventional visual inspection requires huge resources for researchers; therefore, some automatic determination methods have been proposed in recent years. In this study, we evaluate the accuracy of UHR frequency determination by convolutional neural networks (CNN). We find that the averaged error rate of the automatically determined UHR frequency is less than 7.8% for the most events (91% of the data set), and we conclude that the CNN-based UHR frequency determination is the reliable method to estimate the electron density along the satellite orbit.

Modeling of the cold electron plasma density for radiation belt physics

“…NURD is applied to Van Allen Probes EMFISIS data in Allison et al (2021) to show that the plasma density has a controlling effect over acceleration of radiation belt electrons to ultra-relativistic energies. ML-based methods for automatically determining the UHR frequency have also been applied to the Arase satellite using convolutional neural network (Hasegawa et al, 2019;Matsuda et al, 2020) and the CLUSTER mission using several automated pipelines based on neural network methods (Gilet et al, 2021). Machine models of the electron density discussed in this article are listed and succinctly synthetized in Table 4.…”

Tracking the Upper Hybrid Resonance Emission in the Inner Magnetosphere using a Convolutional Neural Network

“…S. Matsuda (1)* , T. Hasegawa (2) , A. Kumamoto (3) , F. Tsuchiya (3) , Y. Kasahara (1) , Y. Miyoshi (4) , Y. Kasaba (3) , A. Matsuoka (5) , I. Shinohara (6) (1) Kanazawa University, Kanazawa, 920-1192, Japan e-mail: matsuda@staff.kanazawa-u.ac.jp;…”

Assessment of the Earth’s Cold Plasmatrough Modeling by Using Van Allen Probes/EMFISIS and Arase/PWE Electron Density Data

“…The local electron density data are derived from the frequency spectra determined by the High Frequency Analyzer (HFA; Kumamoto et al., 2018) subsystem, which is part of the PWE (Kasahara et al., 2018), and using the local magnetic field measured by the Magnetic Field Experiment (MGF; Kasahara et al., 2021; Matsuoka et al., 2018). A semiautomated procedure to derive the electron density has been also developed with an average error rate below 7.8% when the wave frequency is above 30 kHz and when the wave spectral intensity is less than 10 −5 mV 2 /m 2 /Hz (Matsuda et al., 2020). More sophisticated techniques from deep learning without additional features based on expert knowledge allow to determine the upper hybrid frequency and, thus, the electron density with high accuracy (Hasegawa et al., 2019).…”