A Tailored Deep Learning Network with Embedded Space Physical Knowledge for Auroral Substorm Recognition: Validation Through Special Case Studies

The dynamic morphological characteristics of the auroral oval serve as critical diagnostic indicators for auroral substorm recognition, with each pixel in ultraviolet imager (UVI) data carrying different physical implications. Existing deep learning approaches often overlook the physical properties of auroral images by directly transplanting generic models into space physics applications without adaptation. In this study, we propose a visual–physical interactive deep learning model specifically designed and optimized for accurate auroral substorm recognition. The model leverages the significant variation in auroral morphology across different substorm phases to guide feature extraction. It integrates magnetospheric domain knowledge from space physics through magnetic local time (MLT) and magnetic latitude (MLAT) embeddings and incorporates cognitive features derived from expert eye-tracking data to enhance spatial attention. Experimental results on substorm sequences recognition demonstrate satisfactory performance, achieving an accuracy of 92.64%, precision of 90.29%, recall of 93%, and F1-score of 91.63%. Furthermore, several case studies are presented to illustrate how both visual and physical characteristics contribute to model performance, offering further insight into the spatiotemporal complexity of auroral substorm recognition.