Classification of Space Objects by Using Deep Learning with Micro-Doppler Signature Images.

Area of Science:

• Radar signal processing
• Machine learning for defense applications

Background:

• Micro-doppler frequency analysis is crucial for radar target classification in missile defense.
• Current methods heavily rely on feature extraction, limiting classifier generalization.
• Existing deep learning approaches using CNNs and GANs have drawbacks, including single-feature reliance and data augmentation challenges.

Purpose of the Study:

• To improve the generalization performance and robustness of radar target classifiers.
• To address limitations of single-feature input and data augmentation in existing methods.
• To propose a novel transfer learning-based parallel network for enhanced classification.

Main Methods:

• Developed a parallel network architecture utilizing both spectrogram and cadence velocity diagram (CVD) as inputs.
• Generated an electromagnetic (EM) simulation-based dataset using the shooting and bouncing rays concept.
• Simulated radar-received signals based on target dynamics and relative aspect angles.

Main Results:

• The proposed parallel network achieved higher accuracy compared to pre-trained networks using single input features.
• Demonstrated an accuracy improvement of approximately 0.01% to 0.39%.
• The EM simulation dataset provided realistic target features for robust model evaluation.

Conclusions:

• The parallel network architecture effectively leverages multiple input features for improved radar target classification.
• The proposed method offers a more robust and generalizable solution for missile defense systems.
• EM simulation provides a viable alternative for generating high-fidelity datasets for radar target classification research.