Automated Classification of Auroral Images

Modern ground-based auroral All-Sky Imager (ASI) networks capture millions of images annually. While case studies still play an important role in auroral research, there are important reasons for developing automated analysis tools for use on large numbers of images, which is required for objective statistical studies. Future programmes such as NASA's THEMIS (Time History of Events and Macroscale Interactions during Substorms), which will produce about 84 million auroral images annually, can not fully utilise the data without automated tools. We have developed computer vision techniques for automatic classification of auroral images in terms of the type of aurora they contain. In this paper, we discuss our approach in formulating a numerical representation of the contents of each image, and the results of the application of these techniques to roughly 500,000 images from CANOPUS and MIRACLE. We also discuss possibilities for incorporating these techniques in multi-instrument statistical studies in which extensive ASI data sets will be combined with other obsevations, like solar wind or ground-based magnetometer data.