U‐Net Based Machine Learning for Identification of Auroras in Space‐Based Images

Abstract

We use a Semantic Segmentation based U‐Net model to predict the appearance of various classifications of aurora present in near‐infrared images from the Fast Auroral Imager (FAI) onboard the CASSIOPE satellite. We manually labeled pixels from a set of 4091 FAI near‐infrared images into 7 different classes: no aurora, contaminated, diffuse, discrete, arc, ambiguous, and vignette. The U‐Net architecture was modified by replacing 3 layers and changing the input layer to be compatible with 280 × 256 pixel images. The challenges for auroral classification with machine learning when subclasses overlap with each other are explored. Tags for images were generated by analyzing predictions from the trained model over a series of successive images. Overall, the trained model was able to achieve 52.8% Mean Intersection over Union for unseen data. The use of this model and the resulting tags as a search filter enable fast and simple discovery of aurora in space‐based images.