Machine Learning in Auroral Image Research: Aurora Image Classification using Machine Learning Techniques and Substorm Forecasting

Abstract

Images taken of aurora are of key interest for research regarding space physics.To get as much material as possible, many imagers are positioned in key locations around the world, taking many images per year.Over the span of the last two decades, hundreds of millions of images have been taken.Manual evaluation of these images is tedious and time intensive, but automated analysis has so far not been feasible on a large scale.Using machine learning methods we trained a lightweight, easily scalable and open source classifier that is able to classify images into several "aurora" and "no-aurora" or aggregated classes thereof with near human-like accuracy in a fraction of the time.A byproduct of this process are machine-processable numerical features which we have shown to be an accurate descriptor of the images when modelling physical behaviour correlated with the aurora.Based on our classifier and these features in combination with solar wind data we were able to show the feasibility of machine learning based space weather forecasts.This will allow researchers to more accurately monitor and research space weather phenomena in the future.