Single Target SAR 3D Reconstruction Based on Deep Learning.

Area of Science:

• Geospatial analysis
• Remote sensing technology
• Artificial intelligence in imaging

Background:

• Synthetic aperture radar tomography (TomoSAR) is crucial for 3D mapping but demands numerous observation orbits.
• Constraints like funding and time-sensitive target changes limit the number of available orbits.
• Insufficient orbits degrade 3D reconstruction quality, reducing signal-to-noise ratio (SNR), peak-to-sidelobe ratio (PSR), and resolution.

Purpose of the Study:

• To address the limitations of insufficient orbits in TomoSAR.
• To enhance the resolution and SNR of 3D reconstructions using deep learning.
• To enable effective 3D mapping with significantly fewer observation orbits.

Main Methods:

• A deep learning approach utilizing a 3D super-resolution convolutional neural network (CNN) was developed.
• The CNN was trained on pairs of low-resolution (from ~3 orbits) and high-resolution (from all available orbits) 3D voxel-grid reconstructions.
• The model learns target prior distributions to improve reconstruction quality.

Main Results:

• The proposed deep learning algorithm significantly improved the quality and quantity of 3D reconstructions from sparse TomoSAR data.
• Experiments on the Civilian Vehicle Radar dataset demonstrated enhanced resolution and SNR.
• The model exhibited good generalization capabilities on unseen targets.

Conclusions:

• Deep learning effectively overcomes the challenge of limited observation orbits in TomoSAR.
• The developed method enhances the practical applicability of TomoSAR for 3D mapping.
• This approach offers a viable solution for improving 3D reconstruction quality in data-scarce scenarios.