A large data set of auroral poleward/equatorward boundary, which are automatically identified from more than 60,000 images acquired by Polar ultraviolet imager, is used to study the variation of auroral boundaries associated with the interplanetary/geomagnetic environments and dynamic parameters of the magnetosphere. A total number of 3,805,000/1,215,000 points for poleward/equatorward boundary were identified from the data set. It is found that the poleward/equatorward boundaries show an equatorward movement with the increase of Kan‐Lee electric field. The poleward/equatorward boundaries in the nightside sector (1800–0600 magnetic local time (MLT)) show an equatorward motion with the increase of the value of IMF B_y and the north‐south electric field. The equatorward boundary (1800–0300 MLT) shows a quasi‐linear equatorward displacement with the increase of solar wind dynamic pressure and speed. The poleward boundary (excluding midday sector) and equatorward boundary expand equatorward with the increase of AE index. The auroral oval boundary model with input parameters of the three components of IMF, solar wind speed and density, and AE is developed by using multivariate regression method. Evaluation of the model shows that the mean absolute deviation of the model in every MLT sector is 1.3–2.1° magnetic latitude (MLAT) for the poleward boundary and 1.3–2.5° MLAT for the equatorward boundary, respectively, when input parameters are within valid value coverage (three components of IMF are each from −8 to +8 nT, and the values of solar wind speed and density, and AE are not more than 550 km/s, 20/cm⁻³, and 520 nT, respectively).

Extracting auroral key local structures (KLS) containing both morphological information and spatial location from large amount of auroral images is the key for automatic auroral classification and event recognition and thus is very important for improving the efficiency of aurora study. However, it is labor‐intensive for human experts and challenging for computer execution because of the blurry boundary and inhomogeneous luminosity of aurora. In this work, we formulate the automatic KLS extraction to learn a mapping from an all‐sky auroral image to a KLS mask style with artificial intelligence technique. We explore the Cycle‐consistent Generative Adversarial Network to generate the KLS masks based on the unpaired training data set consisting of 2508 all‐sky auroral images observed between years 2003–2009 at the Arctic Yellow River Station, of which only 200 images were manually annotated as the ground truth of KLS masks. The extracted KLS are consistent with human visual perception, and the intersection‐over‐union metric is significantly improved by at least 17% than previous method by Niu et al. (2018) (https://doi.org/10.1109/TGRS.2018.2848725). Classification accuracy has been improved when combining the KLS masks further indicates the effectiveness and value of this work.

Auroral forms are correlated with certain physical processes in the magnetosphere and ionosphere. It is therefore desirable to automatically classify the vast amount of observed auroral images and make large statistical studies. The key problem in classification tasks is image representation. In this paper, using the adaptive feature learning ability of convolutional neural networks, an end-to-end auroral image classification network is proposed, which automatically classifies the auroral images observed at Chinese Yellow River Station into four classes: arc, drapery corona, radial corona, and hot-spot corona. Based on AlexNet, our method exploits the advanced spatial transformer network (STN) and large margin Softmax (L-Softmax) loss function to extract auroral features. STN is able to learn invariance to translation, scaling, and rotation, whereas L-Softmax increases the difficulty of auroral feature learning so that encourages the intra-class compactness and inter-class separability between learned features. The proposed method was validated on auroral image datasets by supervised classification, image retrieval, and statistical analysis of the temporal occurrence distributions of the four auroral categories. Experimental results showed that, after trained on the winter auroral observations in 2003, the proposed model achieves an average classification accuracy of 93.7% on the auroral data of the following five winters (2004—2009) while maintaining high efficiency, which is superior to previously reported papers.

Auroral image classification has long been a focus of research in auroral physics. However, current methods for automatic auroral classification typically assume that only one type of aurora is present in an auroral image. This oversight neglects the complex transition states and coexistence of multiple types during the auroral evolution process, thus limiting the exploration of the intricate semantics of auroral images. To fully exploit the physical information embedded in auroral images, this paper proposes a multi-label auroral classification method, termed MLAC, which integrates convolutional neural network (CNN) and Transformer architectures. Firstly, we introduce a multi-scale feature fusion framework that enables the model to capture both fine-grained features and high-level information in auroral images, resulting in a more comprehensive representation of auroral features. Secondly, we propose a lightweight multi-head self-attention mechanism that captures long-range dependencies between pixels during the multiscale feature fusion process, which is crucial for distinguishing subtle differences between auroral types. Furthermore, we design a residual focused multilayer perceptron module that integrates large kernel depth-wise convolution with an improved multilayer perceptron. This integration enhances the model’s ability to represent complex spatial structure, thus improving local feature extraction and global contextual understanding. The proposed method achieves a mean average precision (mAP) of 88.20% on the auroral observation data collected at the Yellow River Station from 2003 to 2008. This performance significantly surpasses that of the most advanced multi-label classification models while maintaining competitive computational efficiency. Moreover, our method also outperforms the state-of-the-art multi-label methods in both computational efficiency and classification accuracy on two publicly available multi-label image datasets: WIDER-Attribute and VOC2007. These results demonstrate that our method skillfully leverages the robust feature extraction capability of CNNs for local features and the superior global information processing capability of Transformer.