CD-TrGNN: A Complex-Domain Transformer-Graph Neural Network for ISAR Space Target Attitude Estimation

Area of Science:

• Space Surveillance
• Artificial Intelligence
• Signal Processing

Background:

• Accurate space target attitude estimation is crucial for space situational awareness.
• Existing inverse synthetic aperture radar (ISAR) image processing methods have limitations, including discarding phase information, restricted receptive fields in convolutional neural networks, and failure to model physical satellite topology.

Purpose of the Study:

• To address the limitations of current ISAR-based attitude estimation techniques.
• To propose a novel framework that unifies global context modeling and adaptive topological reasoning for enhanced space target attitude estimation.

Main Methods:

• Development of a complex-domain Transformer-graph neural network (CD-TrGNN).
• Implementation of a complex-domain Transformer module (CD-Transformer) for capturing long-range dependencies while preserving amplitude and phase information.
• Integration of a complex-domain graph convolution module (CD-GC) for encoding structural relationships among satellite components using learnable adjacency matrices.

Main Results:

• CD-TrGNN achieved a three-axis mean absolute error of 1.70° on a self-built ISAR complex image dataset, significantly outperforming six baseline methods.
• Ablation studies validated the contributions of complex-domain processing, global attention, and topological reasoning.
• The model demonstrated robustness, with an error of 2.81° at a 5 dB signal-to-noise ratio and accuracy below 2° for different satellite structures.

Conclusions:

• The proposed CD-TrGNN effectively utilizes the complete information present in ISAR complex images.
• This framework enables highly accurate and robust space target attitude estimation, advancing space situational awareness.