Keypoint Extraction of Auroral Arc Using Curvature-Constrained PointNet++.

The study of auroral morphology and evolution process is an effective tool to study the influence of solar wind on the Earth's magnetosphere-ionosphere, which can be used for space weather prediction. Aurora is caused by charged particles originating from the solar wind that precipitate along magnetic field lines toward Earth and collide with neutral constituents of the upper atmosphere. In this paper, we propose a new perspective for the study of auroras, which is no longer limited to morphological information but based on point sets. Specifically, the U-Net network is used to efficiently and accurately segment the aurora arc region, and then the aurora arc region is represented by a set of points. Considering the diversity of auroral shapes, we first calculate the curvature of the aurora arc and combine it with the aurora arc curvature image to constrain the number of keypoints extracted by the PointNet++. Experiments show that the extracted keypoints can represent the objects' structural information and local details well. In addition, it can be further applied to other tasks.