Weakly Supervised Vortex Detection for Studying Correlation Between Multiscale Auroral Events

Abstract

Aurora is the most visible manifestation of the sun’s effect on Earth. The ground-based all-sky imager (ASI) can observe a wealth of multiscale morphological features. Auroral image classification is an important tool for studying magnetospheric regimes and dynamic activities of aurora. Previous studies of automated auroral image classification focused more on auroral large-scale features across the entire image, ignoring small-scale auroral structures. In this letter, we introduce an object detection approach to investigate the small-scale features of auroral morphology. Since the small-scale auroral structures are nonrigid, morphologically diverse, and undefined boundaries, pixel-level labeling is labor-intensive and error-prone for human experts. Therefore, a weakly supervised object detection method for auroral vortexes is proposed in the absence of pixel-level annotations. We first perform global semantic identification and coarse localization using image-level labels as supervision. Considering the motion properties of vortexes, the global and local semantic information in a spatiotemporal volume is leveraged as semantic and location continuity constraints to generate high-confidence pseudo-labels. The experiments demonstrate that the proposed method can identify the vortexes more accurately. The method can retrieve small-scale auroral events in the aurora image dataset, allowing the study of the correlation of multiscale auroral events to be carried out.