GeoDiff-SAR: A Geometric Prior Guided Diffusion Model for SAR Image Generation

Abstract

Synthetic Aperture Radar (SAR) imaging results are highly sensitive to observation geometries and the geometric parameters of targets. However, existing generative methods primarily operate within the image domain, neglecting explicit geometric information. This limitation often leads to unsatisfactory generation quality and the inability to precisely control critical parameters such as azimuth angles. To address these challenges, we propose GeoDiff-SAR, a geometric prior guided diffusion model for high-fidelity SAR image generation. Specifically, GeoDiff-SAR first efficiently simulates the geometric structures and scattering relationships inherent in real SAR imaging by calculating SAR point clouds at specific azimuths, which serves as a robust physical guidance. Secondly, to effectively fuse multi-modal information, we employ a feature fusion gating network based on Feature-wise Linear Modulation (FiLM) to dynamically regulate the weight distribution of 3D physical information, image control parameters, and textual description parameters. Thirdly, we utilize the Low-Rank Adaptation (LoRA) architecture to perform lightweight fine-tuning on the advanced Stable Diffusion 3.5 (SD3.5) model, enabling it to rapidly adapt to the distribution characteristics of the SAR domain. To validate the effectiveness of GeoDiff-SAR, extensive comparative experiments were conducted on real-world SAR datasets. The results demonstrate that data generated by GeoDiff-SAR exhibits high fidelity and effectively enhances the accuracy of downstream classification tasks. In particular, it significantly improves recognition performance across different azimuth angles, thereby underscoring the superiority of physics-guided generation.

Figures (17)

• Figure 1: Radar chart comparison of five mainstream SAR image generation paradigms. The axes represent key performance metrics: Generation Speed, Accurate Azimuth Control, Physical Interpretability, Simplicity of Required Parameters, and Geometric Structure Accuracy. Compared to Electromagnetic Simulation, GANs, Autoregressive Models, and standard Diffusion Models, our proposed GeoDiff-SAR (represented by the green area) achieves the most balanced performance, effectively combining the high physical fidelity of simulations with the efficient generative capability of deep learning.
• Figure 2: Illustration of the research motivation and paradigm comparison. (Left) Traditional electromagnetic simulations ensure physical correctness but suffer from slow generation speeds and require complex parameter configurations (e.g., antenna dimensions, PRF). (Middle) Generic deep learning-based text-to-image models offer fast generation but often lack physical constraints, leading to geometric hallucinations and structural instability. (Right) The proposed GeoDiff-SAR framework bridges this gap by integrating a 3D Processing Module to extract physical point clouds as geometric priors. This hybrid paradigm achieves a synergy of high generation efficiency, precise geometric structure, and azimuth controllability.
• Figure 3: Visual comparison between real SAR samples (top row) and images generated by GeoDiff-SAR (bottom row) across ten distinct azimuth angles. The generated images accurately reproduce key SAR characteristics, including view-dependent geometric distortions (e.g., layover) and the distribution of strong scattering centers. The high degree of visual alignment with ground truth samples demonstrates the model's robustness in capturing both the physical structure and the speckle texture of SAR targets under varying observation geometries.
• Figure 4: The overall architecture of the proposed GeoDiff-SAR framework. (a) Construction of Geometric Priors: 3D models are processed via a point cloud processing module to extract explicit geometric scattering characteristics, such as multi-bounce reflections, serving as a robust physical prior. (b) Multi-modal Fusion Network: These physical features are synthesized with textual and visual conditions through a multi-modal fusion network to condition the Stable Diffusion 3.5 backbone, which is efficiently adapted to the SAR domain using LoRA (trainable) while keeping the pre-trained weights frozen. (c) Controllable Generative Model: As depicted in the top-right, the generated high-fidelity SAR images are utilized to augment scarce real training data for downstream PyTorch Image Models Multi-Label Classification tasks, thereby validating the practical utility and effectiveness of the proposed data augmentation strategy.
• Figure 5: Schematic diagram of the object-centric radar observation geometry employed in the point cloud simulation. The coordinate system is defined with the $X$-axis pointing West ($0^{\circ}$) and the $Y$-axis pointing South ($90^{\circ}$). The spatial position of the radar sensor is determined by the azimuth angle $\phi$ and the depression angle $\psi$, establishing the Line-of-Sight (LOS) vector pointing towards the target center $O$. This geometric configuration serves as the basis for constructing the transformation matrix $\mathbf{T}_{radar}$ and executing ray casting to capture view-dependent scattering characteristics.