All-sky auroral image classification based on Low-rank representation and multi-level feature fusion

The morphology of the aurora is closely related to the solar-terrestrial electromagnetic activity. Morphological classification of auroras is an important tool to achieve attribution analysis of various auroral phenomena. This paper proposes an all-sky auroral image classification method based on low-rank representation and multi-level feature fusion. First, the local temporal stable part of the auroral sequence is extracted by sparse and low-rank decomposition. Local spatial features are extracted using ResNet50, and then multi-level local feature correlations are modeled at the global scale using a self-attention mechanism to highlight the local key structures. To enhance the discriminative ability of the features, the cross-level features are adaptively fused by a parameter learnable strategy. The proposed algorithm is proven to have good performance in all-shy auroral image classification by extensive experiments, and the effectiveness of each module is verified.