Ultraviolet Auroral Substorm Cross‐Modal Parameters‐ Image Retrieval Based on Deep Learning Approach

Abstract An auroral substorm is a critical phenomenon demonstrating magnetospheric energy accumulation and explosive release. It is crucial to automatically detect auroral substorms within the vast aurora data to ensure the safety of solar‐terrestrial space. Usually, the Westward Traveling Surge (WTS) is one of the typical morphological aurora structures, which can indicate the occurrence of auroral substorms to a certain extent. The existing auroral retrieval models are based on some auroral images to search for similar aurora images and special auroral events, but the studies show that the auroral activities are closely related to the parameters of space physics. We construct the cross‐modal retrieval model with Multi‐Head Cross Attention (MHCA) (Deep Supervised with Cross‐Modal Attention, DS_CMA) based on the deep learning method to find the relationship between space physics parameters and auroral substorm events by the image data observed by Special Sensor Ultraviolet Spectrographic Imager (SSUSI) carried on DMSP satellites and space physical parameters from OMNI data sets by the NASA (National Aeronautics and Space Administration). Since space physics parameters and image data are derived from different observation equipment, we employ a cross‐modal retrieval model embedded with Multi‐Head Cross Attention, which can effectively process these different modality data. The experiment results show that our DS_CMA model achieves the best retrieval results, with the mean average precision (mAP) of 77.78%, and the Recall@10 of 78.19%, which demonstrates that the proposed model can effectively retrieve auroral substorm images by using space physics parameters.