Group-Wise Learning for Aurora Image Classification With Multiple Representations.

Area of Science:

• Geophysics and Space Physics
• Computer Science
• Artificial Intelligence

Background:

• Aurora image classification traditionally uses single feature representations, limiting the capture of complex morphologies.
• Existing multi-feature methods often overlook the inherent correlations between different representations.

Purpose of the Study:

• To develop an automatic aurora image classification method that effectively leverages multiple feature representations.
• To address the limitations of single-feature methods and the neglect of correlations in multi-feature approaches.

Main Methods:

• Proposed a group-wise learning (GWL) method for automatic aurora image classification using multiple representations.
• Constructed graphs in multiple feature spaces and used clustering to group correlated representations.
• Developed a GWL model for data-driven estimation of class labels and optimal feature weights, followed by a label fusion approach.

Main Results:

• The GWL method demonstrated improved performance on a dataset of 12,682 aurora images.
• Achieved approximately a 6% increase in classification accuracy compared to single-feature representation methods.

Conclusions:

• The proposed GWL method effectively utilizes diverse properties of multiple feature representations by grouping correlated ones.
• This approach offers a significant improvement in aurora image classification accuracy, outperforming conventional methods.