Classification of Ground‐Based Auroral Images by Learning Deep Tensor Feature Representation on Riemannian Manifold

• Published in: Journal of geophysical research. Machine learning and computation Vol. 1; no. 2
• Main Authors: Hu, Yangfan, Zhou, Zeming, Yang, Pinglv, Zhao, Xiaofeng, Zhang, Peng
• Format: Journal Article
• Language: English
• Published: 01.06.2024
• Subjects: auroral image classification; dimensionality reduction; Riemannian manifold learning; second‐order tensor feature
• DOI: 10.1029/2023JH000109
• ISSN: 2993-5210

Automatically classifying a huge amount of ground‐based auroral images is essential to facilitate aurora morphology statistical research and aid in comprehending the magnetospheric dynamics. However, facing the challenge of insufficient labeled images, deep learning methods perform suboptimally on small auroral image data sets, and traditional machine learning methods based on handcrafted features heavily rely on expert knowledge. In this paper, we propose a novel method that leverages the merits of both traditional machine learning and deep learning methods by extracting deep second‐order tensor features to train a support vector machine (SVM). To improve compactness and discriminative ability of the features, we comply the intrinsic data geometry on Riemannian manifold to employ dimensionality reduction and map the dimensionality‐reduced features from Riemannian space to Euclidean space for the SVM classifier. Experimental results on small aurora data sets conclusively demonstrate the effectiveness of our method, exhibiting competitive performance with recent aurora classification methods. Plain Language Summary Auroral images constitute the primary optical observation of aurora. However, the annotation for auroral images is time‐consuming and tedious, provoking the need for leveraging a small number of annotated samples to improve auroral image classification accuracy. Auroral image classification has been accomplished by traditional machine learning or deep learning methods. Combining the merits of both, we propose a novel method harnessing the robust feature extraction ability of a pretrained convolutional neural network (CNN) and the superior classification performance of a support vector machine (SVM) for small auroral image data sets. Our method represents deep second‐order tensor features of auroral images by capturing correlations between channels in CNN‐based features. Since deep second‐order tensor features reside on Riemannian manifold, we measure the feature similarity by geodesic distance and adopt dimensionality reduction to refine those features by imposing a constraint that contracts the data points' intra class and separates the data points inter classes. Moreover, dimensionality‐reduced features still locate in Riemannian space, in the manner of more compactness and higher discriminative ability; we map the features from Riemannian space to Euclidean space for SVM applicability. Experimental results on small data sets conclusively demonstrate the effectiveness of our method, which consistently outperforms recent auroral image classification approaches. Key Points We leverage the merits of traditional machine learning and deep learning to improve the classification accuracy of auroral images We apply Riemannian manifold learning to improve the discriminative ability of feature and preserve its non‐Euclidean geometry structure Experiments demonstrate the proposed method is superior to recent classification results on small ground‐based auroral image data sets