An automated auroral detection system using deep learning: real-time operation in Tromsø, Norway

Abstract

The activity of citizen scientists who capture images of aurora borealis using digital cameras has recently been contributing to research regarding space physics by professional scientists. Auroral images captured using digital cameras not only fascinate us, but may also provide information about the energy of precipitating auroral electrons from space; this ability makes the use of digital cameras more meaningful. To support the application of digital cameras, we have developed artificial intelligence that monitors the auroral appearance in Tromsø, Norway, instead of relying on the human eye, and implemented a web application, “Tromsø AI”, which notifies the scientists of the appearance of auroras in real-time. This “AI” has a double meaning: artificial intelligence and eyes (instead of human eyes). Utilizing the Tromsø AI, we also classified large-scale optical data to derive annual, monthly, and UT variations of the auroral occurrence rate for the first time. The derived occurrence characteristics are fairly consistent with the results obtained using the naked eye, and the evaluation using the validation data also showed a high F1 score of over 93%, indicating that the classifier has a performance comparable to that of the human eye classifying observed images.

Characteristics of auroral electron precipitation at 1 geomagnetic latitude 67° over Tromsø

“…Since the 2015 result of Nanjo et al (2022) contain data from year 2016, it seems that the year of maximum occurrence rate in our study match with the optical aurora occurrence rate.…”

“…The occurrence rate increases monotonically from post-noon to morning hours. A similar MLT trend in the occurrence rate of optical aurora was reported byNanjo et al (2022) who found that occurrence rate of the diffuse aurora increases monotonically from 18 MLT until 04 MLT, and decreases afterwards. An explanation for the observed MLT dependence of auroral electron precipitation occurrence rate will be given in Section 4.2.…”

“…As shown in the figure, the auroral occurrence frequency has maxima during equinoctial months, in September and March, and a minimum in December to January. Nanjo et al (2022) found maximum optical auroral occurrence frequency in October and February, and minimum in December. The semi-annual variation in geomagnetic activity has long been recognized (Cortie, 1912;Chapman & Bartels, 1940;Russell & McPherron, 1973) and it gives peaks in geomagnetic activity at times close to equinoxes.…”

“…Morning sector electron precipitation is known to be associated with diffuse aurora, including pulsating aurora, (Newell et al, 2009;Nanjo et al, 2022). Medium-energy ( 1-30 keV) electrons are injected into the inner magnetosphere during periods of enhanced convection or substorms.…”

The Morphology of Poleward Moving Auroral Forms

“…The automation of PMAF detection has been found to be very difficult. While there have been recent advancements in the automation of detection, and to some extent, classification of auroral forms (Nanjo et al, 2022;Guo et al, 2022), there is currently no automated PMAF detection algorithm as far as we can tell. In the future our findings might enable the development of an automated PMAF detection algorithm based on the arciness index.…”