SARDNet: structure-anchored reference diffusion network for sim-to-real generation of SAR vehicle targets

Synthetic aperture radar (SAR) is widely used in Earth observation and automatic target recognition (ATR) because of its all-time and all-weather imaging capability. However, reliable SAR interpretation remains challenging because SAR images are governed by coherent imaging mechanisms and are affected by speckle fluctuations, target-dependent scattering responses, and complex background clutter. Simulation-based SAR image generation provides abundant labeled samples under controllable imaging conditions, but simulated images often exhibit substantial visual and statistical discrepancies from real SAR data, limiting their direct applicability to real-domain ATR. To address this problem within the scope of vehicle-target imagery, this paper proposes a structure-anchored reference diffusion network (SARDNet) for sim-to-real generation of SAR vehicle targets. SARDNet formulates SAR style transfer as a physically constrained conditional diffusion problem, in which simulated images serve as structural anchors to preserve target geometry, shadow layout, and label consistency, while real SAR images provide reference speckle, contrast, and clutter statistics. To mitigate structure-style interference, SARDNet introduces physics-enhanced dynamic region decomposition and implicit spatial routing, enabling foreground structures to be anchored by simulated inputs and background styles to be guided by real references. Residual feature-wise linear modulation, selective cross-attention, and dataset-level background statistical priors are further incorporated to improve real-domain realism while preserving discriminative scattering structures. Experiments on the SAMPLE and RaySAR-MSTAR datasets show that SARDNet achieves the best results among the compared methods on SSIM, HistBD, GLCM Dist, and TCR, while GAN-based baselines obtain lower KID values. Downstream SAR-ATR experiments further indicate that the generated samples can improve real-domain recognition accuracy, supporting the effectiveness of SARDNet for simulation-assisted SAR vehicle-target recognition.

Authors

Institutions

Publication Details

Journal
Geo-spatial Information Science
Published
2026-11-17
DOI
https://doi.org/10.1080/10095020.2026.2759561
Primary Topic
Advanced SAR Imaging Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

SARDNet: structure-anchored reference diffusion network for sim-to-real generation of SAR vehicle targets

Timo Balz, Yifei Li
Geo-spatial Information Science
Advanced SAR Imaging Techniques
article

SARDNet: structure-anchored reference diffusion network for sim-to-real generation of SAR vehicle targets

Timo Balz, Yifei Li
article en

Abstract

Synthetic aperture radar (SAR) is widely used in Earth observation and automatic target recognition (ATR) because of its all-time and all-weather imaging capability. However, reliable SAR interpretation remains challenging because SAR images are governed by coherent imaging mechanisms and are affected by speckle fluctuations, target-dependent scattering responses, and complex background clutter. Simulation-based SAR image generation provides abundant labeled samples under controllable imaging conditions, but simulated images often exhibit substantial visual and statistical discrepancies from real SAR data, limiting their direct applicability to real-domain ATR. To address this problem within the scope of vehicle-target imagery, this paper proposes a structure-anchored reference diffusion network (SARDNet) for sim-to-real generation of SAR vehicle targets. SARDNet formulates SAR style transfer as a physically constrained conditional diffusion problem, in which simulated images serve as structural anchors to preserve target geometry, shadow layout, and label consistency, while real SAR images provide reference speckle, contrast, and clutter statistics. To mitigate structure-style interference, SARDNet introduces physics-enhanced dynamic region decomposition and implicit spatial routing, enabling foreground structures to be anchored by simulated inputs and background styles to be guided by real references. Residual feature-wise linear modulation, selective cross-attention, and dataset-level background statistical priors are further incorporated to improve real-domain realism while preserving discriminative scattering structures. Experiments on the SAMPLE and RaySAR-MSTAR datasets show that SARDNet achieves the best results among the compared methods on SSIM, HistBD, GLCM Dist, and TCR, while GAN-based baselines obtain lower KID values. Downstream SAR-ATR experiments further indicate that the generated samples can improve real-domain recognition accuracy, supporting the effectiveness of SARDNet for simulation-assisted SAR vehicle-target recognition.

Geo-spatial Information Science
Wuhan University (CN)
Advanced SAR Imaging Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon