Unsupervised Anomaly Detection With Variational Autoencoders Applied to Full-Disk Solar Images

Deep learning is successful in many fields due to its ability to learn strong feature representations without the need for hand-crafted features, resulting in models with high representational power. However, many of these models are based on supervised learning and therefore depend on the availability of large annotated data sets. These are often difficult to obtain because they require human input. A common challenge for researchers in space weather is the sparsity of annotations in many of the available data sets, which are either unlabeled or have ambiguous labels. To alleviate the data bottleneck of loosely annotated data sets, unsupervised deep learning has become an important strategy, with anomaly detection being one of the most prominent applications. Unsupervised models have been successfully applied in various domains, such as medical imaging or video surveillance, to distinguish normal from abnormal data. In this work, we investigate how a purely unsupervised approach can be used to detect and extract solar phenomena in extreme ultraviolet images from the NASA SDO spacecraft. We show how a model based on variational autoencoders can be used to detect out-of-distribution samples and to localize regions of interest for solar activity. By using an unsupervised approach, we hope to contribute to space weather monitoring tools and further improve the understanding of space weather drivers.