Substorm Event Retrieval Model in Ultraviolet Aurora Images Based on Contextual CNN Features

Abstract

Auroral substorms are one of the disturbance phenomena caused by high-energy charged particles from the solar wind precipitating into the Earth's magnetosphere and colliding with charged particles within the magnetosphere. Understanding the occurrence and evolution of substorms can help elucidate the physical processes governing the interaction between the solar wind and Earth's magnetosphere. Currently, ground-based and satellite-based imaging equipment have captured a vast amount of aurora images, and identifying auroral substorm events from these images is crucial for studying solar-terrestrial relations. The westward traveling surge (WTS) is a typical structure during substorm occurrences and is commonly used for auroral substorm identification. In this paper, we propose a method based on convolutional neural networks (CNNs) that uses a polar region partitioning strategy to locate image keypoints and determine the position and size of regional blocks. Multi-scale contextual CNN features are then generated to retrieve substorm events from ultraviolet aurora images. The results show that the multi-scale features extracted from convolutional and fully connected layers can effectively capture the characteristics of the WTS structure. The method achieves a mean average precision of 75.77% and a Recall@10 of 95.19%, demonstrating its effectiveness in retrieving auroral substorm events.