Application of Deep Learning and Explainable AI to NIR Auroral Detection in Satellite Time‐Series Images

Abstract In this study, we employ a deep learning approach to detect auroras from near‐infrared (NIR) sequential images captured by the satellite‐based imager, the Enhanced Polar Outflow Probe (e‐POP)/Fast Auroral Imager (FAI). We also apply the Eigen‐Class Activation Map (Eigen‐CAM), an explainable AI technique in a broad sense, as a post‐hoc interpretability method to indicate the location of auroral emissions. Accurately identifying the location of auroras is crucial for understanding space weather dynamics. However, it is challenging to distinguish auroral emissions from static features such as clouds, mountains, and city lights in single‐channel NIR images. To overcome this problem, we employ a CNN‐based ResNeXt‐50 deep learning model to automatically detect auroras and discriminate them from non‐auroral features in e‐POP/FAI images. Through comparative experiments on input configurations, we find that the best‐performing model utilizes a sequence of three frames with temporal gaps (at a two‐second cadence). This temporal context allows the model to learn dynamic characteristics and effectively filter out static non‐auroral features. Eigen‐CAM is then applied to visualize the image regions that contribute most to the model's decision, illustrating its potential to highlight auroral structures without supervision. For our study, we use images from 2015 to 2017, with January and July as the test set. Our best model demonstrates high performance, achieving an accuracy of 0.84 and an F1‐score of 0.84. Ultimately, our approach facilitates large‐scale statistical studies of auroral dynamics by replacing manual classification, which also leads to better predictions about changes in Earth's upper atmosphere.