Automatic Electron Density Determination by Using a Convolutional Neural Network

Abstract

In this study, we developed a technique for automatically determining upper hybrid resonance (UHR) frequencies using a convolutional neural network (CNN) to derive the electron density along the orbit of the Arase satellite. We used three CNN models (AlexNet, VGG16 and ResNet) to determine the UHR frequencies without additional features based on an expert's knowledge. We also reproduced the multi-layer perceptron (MLP) model that had been used for the Van Allen probes mission, which requires observed electric field spectra and additional five features (i.e., decimal logarithm of electron cyclotron frequency (log10 fce), L-value, geomagnetic index (Kp), magnetic local time, and frequency bin with the highest power spectral density from the electric field spectra (binmax)). We confirmed that the proposed method using CNN more accurately determined the UHR frequencies than did the conventional method. The mean absolute error (MAE) of the VGG16 model was 3.478 bins when the input vector comprised both the observed electric field spectrum and the additional five features. In contrast, the MAE of the conventional method was 5.986 bins (72.1% worse). Moreover, we confirmed that the proposed method achieves a high accuracy regardless of the use of the additional five features (the MAE of the ResNet model was 3.664 bins when excluding the additional five features). This suggests that the feature map of the ResNet model acquired a representation ability beyond the five features.