Unsupervised TomoSAR Image Reconstruction Through Virtual Multiple Measurement Explorations

Tomographic synthetic aperture radar (TomoSAR) reconstruction produces 3D imaging of scenes from measurements and has recently been combined with data-driven deep learning techniques. While most existing methods rely on supervised learning with simulation-based and paired training samples, we focus...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Proceedings of the IEEE National Radar Conference (1996) s. 321 - 326
Hlavní autoři: Liu, Liang, Zeng, Tianjiao, Wang, Mou, Shi, Jun, Wei, Shunjun, Zhang, Xiaoling, Zhan, Xu
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 04.10.2025
Témata:
data-driven
Deep learning
Image reconstruction
Loss measurement
neural network inductive bias
Noise
Noise measurement
Physics
Redundancy
Three-dimensional displays
tomoSAR
Training
Unsupervised learning
ISSN:2375-5318
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Tomographic synthetic aperture radar (TomoSAR) reconstruction produces 3D imaging of scenes from measurements and has recently been combined with data-driven deep learning techniques. While most existing methods rely on supervised learning with simulation-based and paired training samples, we focus on unsupervised learning that operates without such samples. This approach promises more robust, flexible, and generalized reconstruction capabilities while leveraging deep learning to enhance reconstruction quality. We present a preliminary exploration using virtual multiple measurements derived from raw measurements. We leverage the recent finding that imaging areas possess both non-local and local similarities, resulting in highly correlated measurements. This property enables us to obtain additional virtual measurements beyond the raw data. By treating these measurements as multiple independent noisy samples of the same underlying signal, we propose a self-supervised robust reconstruction loss that incorporates an integrated physics forward measurement model. Our approach feeds measurement pairs to both the input and output of the reconstruction neural network, enabling it to explore the underlying signal while suppressing noise. We also adopt an encoderdecoder network architecture that implicitly incorporates scene priors through its inductive bias. By combining the physics model prior and scene prior while remaining robust to noise, we achieve high-fidelity reconstruction results. Experiments on both simulated and measured data demonstrate our method's effectiveness, achieving reconstruction performance comparable to supervised approaches while showing better generalization under varying noise conditions.
ISSN:2375-5318
DOI:10.1109/RadarConf2559087.2025.11205130