Transfer Learning Aurora Image Classification and Magnetic Disturbance Evaluation

Abstract

Created by particles precipitating into the ionosphere, the aurora are a direct consequence of interactions between the solar wind and the magnetosphere. A connection between different auroral shapes and behaviors simultaneously across the polar region has been first shown in 1964, when analyzing auroral images taken across northern continental America and Antarctica (Akasofu, 1964). The established model of auroral substorms, describes the activity of the aurora from its beginning, when the aurora is calm, over active phases toward another calm phase. This model was later refined to include the three currently used phases "growth", "expansion" and "recovery" (McPherron et al., 1973). Interaction between the solar wind and the interplanetary magnetic field lead to storage of energy in the magnetosphere during the growth phase. In the expansion phase energy is released, before the magnetosphere returns to normal conditions in the recovery phase.Since the beginning of auroral research, images have therefore been an important tool to analyze and diagnose the complex processes in the ionosphere and magnetosphere, often supplemented by magnetometers measuring the local Earth's magnetic field or satellites performing similar measurements. All Sky Imagers, taking pictures of the night sky and capturing aurora in regular intervals, can for example, be found in Ny-Ålesund and Longyearbyen

Substorm Onset Prediction Using Machine Learning Classified Auroral Images

“…866357). The All Sky Image Classifier was made available by Sado et al (2022) on http://tid.uio.no/ TAME. We acknowledge NASA contract NAS5-02099 and V. Angelopoulos for use of data from the THEMIS Mission.…”

“…To remove artifacts like trees just above the horizon, a ring 20 px wide was removed. The images were then classified according to the method developed by Sado et al (2022).…”

“…Figure 1 shows an overview of how the data flows through the system. The all sky images are preprocessed and classified according to the classifier by Sado et al (2022). This process is detailed in Section 3.2.…”

“…Individual images are classified based on the classifier developed and demonstrated by Sado et al (2022); Clausen and Nickisch (2018). In this process, the images are analyzed by a pretrained neural network and the image features as defined by this network are extracted.…”

“…Their work shows how well solar wind data can be used to forecast onsets of substorms on a global level. Furthermore, Sado et al (2022) have shown that all sky images contain sufficient information that can be extracted by a neural network and be used to model the behavior of the Earth's local magnetic field in vicinity to the imager.…”