Auroral Image Processing Techniques - Machine Learning Classification and Multi-Viewpoint Analysis

Every year, millions of scientific images are acquired in order to study the auroral phenomena. The accumulated data contain a vast amount of untapped information that can be used in auroral science. Yet, auroral research has traditionally been focused on case studies, where one or a few auroral events have been investigated and explained in detail. Consequently, theories have often been developed on the basis of limited data sets, which can possibly be biased in location, spatial resolution or temporal resolution. Advances in technology and data processing now allow for acquisition and analysis of large image data sets. These tools have made it feasible to perform statistical studies based on auroral data from numerous events, varying geophysical conditions and multiple locations in the Arctic and Antarctic. Such studies require reliable auroral image processing techniques to organize, extract and represent the auroral information in a scientifically rigorous manner, preferably with a minimal amount of user interaction. This dissertation focuses on two such branches of image processing techniques: machine learning classification and multi-viewpoint analysis. Machine learning classification: This thesis provides an in-depth description on the implementation of machine learning methods for auroral image classification; from raw images to labeled data. The main conclusion of this work is that convolutional neural networks stand out as a particularly suitable classifier for auroral image data, achieving up to 91 % average class-wise accuracy. A major challenge is that most auroral images have an ambiguous auroral form. These images can not be readily labeled without establishing an auroral morphology, where each class is clearly defined. Multi-viewpoint analysis: Three multi-viewpoint analysis techniques are evaluated and described in this work: triangulation, shell-projection and 3-D reconstruction. These techniques are used for estimating the volume distribution of artificially induced aurora and the height and horizontal distribution of a newly reported auroral feature: Lumikot aurora. The multi-viewpoint analysis techniques are compared and methods for obtaining uncertainty estimates are suggested. Overall, this dissertation evaluates and describes auroral image processing techniques that require little or no user input. The presented methods may therefore facilitate statistical studies such as: probability studies of auroral classes, investigations of the evolution and formation of auroral structures, and studies of the height and distribution of auroral displays. Furthermore, automatic classification and cataloging of large image data sets will support auroral scientists in finding the data of interest, reducing the needed time for manual inspection of auroral images