3:30 PM - 3:45 PM
[PEM15-17] Binary Classification of SDO/AIA coronal hole Video Images for the relativistic electron enhancements
Keywords:Space weather, Coronal hole, radiation belt, Binary classification, Embedding technique
This study aims to develop a shape-preserving embedding technique for the binary classification of solar surface video images, particularly focusing on coronal hole structures. By assuming that these structures exhibit high sparsity and strong temporal evolution, we propose a method that transforms two-dimensional extreme ultraviolet (EUV) images into embedded vectors while preserving essential spatial features. This approach enables dimensionality reduction from three-dimensional video data to a two-dimensional time-sequential vector array, improving computational efficiency in classification tasks.
To validate our approach, we use NASA’s SDO/AIA 211Å images to extract coronal hole regions and the GOES-15 relativistic electron flux dataset as classification labels. We apply two embedding methods—presence/absence and occurrence ratio—to generate coronal hole vectors. These vectors serve as input for a deep learning-based binary classification model designed to predict relativistic electron enhancements in Earth's radiation belts. Our experimental results indicate that the occurrence ratio method significantly improves classification performance, demonstrating the importance of embedding quality in deep learning models.
Our findings highlight that a five-day sequence of coronal hole vectors achieves the highest classification accuracy, emphasizing the temporal dependency of coronal hole dynamics. However, further improvements are necessary to enhance prediction robustness and reduce model complexity. Future work will focus on refining embedding techniques and optimizing network architectures to improve prediction reliability in space weather forecasting.
To validate our approach, we use NASA’s SDO/AIA 211Å images to extract coronal hole regions and the GOES-15 relativistic electron flux dataset as classification labels. We apply two embedding methods—presence/absence and occurrence ratio—to generate coronal hole vectors. These vectors serve as input for a deep learning-based binary classification model designed to predict relativistic electron enhancements in Earth's radiation belts. Our experimental results indicate that the occurrence ratio method significantly improves classification performance, demonstrating the importance of embedding quality in deep learning models.
Our findings highlight that a five-day sequence of coronal hole vectors achieves the highest classification accuracy, emphasizing the temporal dependency of coronal hole dynamics. However, further improvements are necessary to enhance prediction robustness and reduce model complexity. Future work will focus on refining embedding techniques and optimizing network architectures to improve prediction reliability in space weather forecasting.