ARSRNet: accurate space object recognition using optical cross section curves.

Area of Science:

• Space Surveillance and Object Recognition
• Artificial Intelligence in Aerospace
• Computational Optics

Background:

• Ground-based optical observations yield limited optical cross section (OCS) data for space objects (SOs).
• Sparse OCS data and unclear labels hinder neural network performance for SO recognition.
• Retraining deep neural networks is necessary for identifying new SOs or categories, impacting efficiency.

Purpose of the Study:

• To develop a novel neural network model, ARSRNet, for accurate SO recognition and attitude identification.
• To improve the generalization and training convergence speed of SO recognition networks.
• To enable SO recognition with minimal, unlabeled OCS data and adapt to new recognition tasks.

Main Methods:

• Introduction of the ARSRNet neural network architecture.
• Development of the AdamRprop network optimization algorithm.
• Training and testing the model on a dataset of OCS data for SOs.

Main Results:

• ARSRNet achieves 90.60% recognition accuracy on the test OCS dataset.
• The network effectively identifies SOs and their attitude using limited, unlabeled OCS data.
• The AdamRprop optimizer accelerates ARSRNet training and convergence compared to other algorithms.

Conclusions:

• ARSRNet offers a robust solution for SO recognition challenges posed by limited and unlabeled OCS data.
• The proposed model demonstrates adaptability to new recognition tasks without extensive retraining.
• The AdamRprop optimizer enhances the efficiency and accuracy of ARSRNet, making it suitable for practical applications.