By Sota Nanjo, PhD student at University of Electro-Communications, Tokyo, Japan
As an activity of citizen science, images taken by amateur photographers using commercial digital cameras have been used for scientific research. Such digital images contain a lot of background noise in various wavelengths, unlike professional optical data captured with a narrow-band optical filter. However, digital cameras have three RGB channels and are able to capture what we see in full color. This capability often enables us to identify characteristic features that are difficult to be noticed in professional monochromatic data (MacDonald et al., 2018; Shiokawa et al., 2018). Furthermore, it has been suggested that colors in a digital image can be used to estimate the average energy of pulsating auroral electrons (Nanjo et al., 2021); thus, digital cameras have the potential to play an essential role in scientific research of auroras. However, unlike professional researchers, who automatically observe auroras overnight, photographers have to wait for an auroral appearance to take pictures, which is often a heavy burden on their observations. To reduce this constraint, it is important to provide a system that notifies them of the auroral appearance in real-time. One of the reasons why such a system has not been developed so far is that the aurora's complex and diverse morphology has prevented highly accurate automatic detection. However, with the improvement of deep learning techniques, that problem has almost been solved (Clausen & Nickisch, 2018; Kvammen et al., 2020). In this study, we developed an AI classifier that automatically detects auroras instead of the human eye using a deep neural network model. We then combined it with real-time observation using a digital camera in Tromsø, Norway (Nozawa et al., 2018) in order to build a website “Tromsø AI” that notifies users of auroral occurrences in real-time. In the presentation, we will explain how to use the website and discuss the solar activity and seasonal and local time dependence of auroral occurrence rates obtained by classifying images from the past 10-year observations in Tromsø.