Automatic Classification of Auroral Images From the Oslo Auroral THEMIS (OATH) Data Set Using Machine Learning

Abstract

Based on their salient features we manually label 5,824 images from various Time History of Events and Macroscale Interactions during Substorms (THEMIS) all‐sky imagers; the labels we use are clear/no aurora, cloudy, moon, arc, diffuse, and discrete. We then use a pretrained deep neural network to automatically extract a 1,001‐dimensional feature vector from these images. Together, the labels and feature vectors are used to train a ridge classifier that is then able to correctly predict the category of unseen auroral images based on extracted features with 82% accuracy. If we only distinguish between a binary classification aurora and no aurora, the true positive rate increases to 96%. While this study paves the way for easy automatic classification of all auroral images from the THEMIS all‐sky imager chain, we believe that the methodology shown here is readily applied to all images from any other auroral imager as long as the data are available in digital form. Both the neural network and the ridge classifier are free, off‐the‐shelf computer codes; the simplicity of our approach is demonstrated by the fact that our entire analysis comprises about 50 lines of Python code. Automatically attaching labels to all available all‐sky imager data would enable statistical studies of unprecedented scope.