A Superposed Epoch Analysis of Auroral Oval Coverage During Substorms Using Deep Learning-Based Segmentation Models

As an effect of solar-terrestrial activity, the aurora has always been a focus of substorm research. The expansion of the aurora is one of the important characteristics during substorms, which could be directly reflected by the change of auroral oval coverage. However, the related studies were limited because of the absence of a reliable automated method to obtain auroral ovals from a large number of images. In this paper, we propose a new strategy to achieve this. Based on the Segment Anything Model, we design a process for annotating the auroral oval region that requires little manual work. A new aurora segmentation model, HrSeg, is then developed to obtain auroral ovals more efficiently and accurately. Through 5-fold cross-validation, it is determined that the average intersection over union, Dice coefficient, and pixel accuracy are all greater than 0.97. Furthermore, images of 590 substorms observed by the Polar satellite Ultraviolet Imager are segmented. We present superposed epoch analyses of the auroral oval coverage calculated from the segmentation results. Generally, the coverage decreases slightly before onset, then rapidly increases for tens of minutes after onset, and finally decreases gradually. Moreover, the auroras in different magnetic local time (MLT) sectors exhibit different evolutions in coverage. It is also revealed that the evolution pattern of auroral coverage depends on interplanetary magnetic field orientations and seasonal conditions. The results quantify the variation of auroral morphology in terms of coverage, which complete the evolution pattern of aurora during substorms and provide a more comprehensive understanding of substorms. Aurora has always been a focus of substorm research. When a substorm occurs, the size of aurora varies dramatically. In this paper, we use the auroral oval coverage to measure the size of the aurora, which could directly reflect the expansion or contraction of the aurora. We propose a deep learning-based segmentation model to extract the auroral oval. Based on the segmentation results, a superposed epoch analysis is presented to explore how auroral oval coverage evolves during the substorms and how interplanetary magnetic field orientations and seasonal conditions influence the evolution of auroral oval coverage. The analysis results quantitatively describe the variation of auroral morphology in terms of coverage. The findings complete the evolution pattern of aurora during substorms and help us comprehensively understand substorms. Based on the Segment Anything Model, a semi-automatic strategy for annotating the auroral oval region is proposed An HRNet-based neural network is developed to segment the auroral oval region from the images during substorms, which is more accurate The superposed epoch analyses of auroral oval coverage are presented to show the differences in MLT and the influences of IMF and season