Multi-Label Auroral Image Classification Based on CNN and Transformer.

Area of Science:

• Space Physics
• Geophysics
• Atmospheric Science

Background:

• Auroral image classification is crucial in auroral physics.
• Current methods often fail to account for multiple aurora types coexisting or transitioning within a single image.
• This limitation hinders a comprehensive understanding of auroral dynamics.

Purpose of the Study:

• To develop an advanced multi-label auroral classification method (MLAC).
• To improve the analysis of complex auroral phenomena by recognizing multiple simultaneous aurora types.
• To enhance the extraction of physical information from auroral imagery.

Main Methods:

• Integration of Convolutional Neural Network (CNN) and Transformer architectures.
• Implementation of a multi-scale feature fusion framework for comprehensive feature representation.
• Utilization of a lightweight multi-head self-attention mechanism for capturing long-range dependencies.
• Development of a residual focused multilayer perceptron module with large kernel depth-wise convolution for enhanced spatial understanding.

Main Results:

• Achieved a mean average precision (mAP) of 88.20% on Yellow River Station data (2003-2008).
• Significantly outperformed state-of-the-art multi-label classification models in accuracy and computational efficiency.
• Demonstrated superior performance on public datasets (WIDER-Attribute, VOC2007), validating its generalizability.

Conclusions:

• The proposed MLAC method effectively leverages CNNs for local features and Transformers for global context.
• MLAC provides a more comprehensive and accurate approach to auroral image classification.
• This advancement facilitates deeper insights into the complex dynamics of auroral displays.