OptiX-Based SAR Image Simulation with Residual-Learning Refinement

Synthetic aperture radar (SAR) imagery plays an important role in disaster monitoring, environmental monitoring, and related applications. However, the acquisition of measured SAR data is limited by high costs and the difficulty of covering diverse environmental conditions, and the volume of SAR imagery is far smaller than that of optical remote sensing imagery, which restricts learning-based SAR applications. This paper proposes a multi-view SAR image simulation method based on OptiX and residual learning. The proposed method uses the NVIDIA OptiX ray-tracing API to simulate electromagnetic wave propagation, uses Monte Carlo sampling to model material interactions and coherent summation to form SAR images, and generates multi-view SAR image sequences of buildings. The simulator uses a scalar multiple-bounce approximation, and simulation, refinement, and evaluation use only the HH channel. The OptiX-based simulator is coupled with a residual U-Net that predicts the residual between simulated and measured images and compensates for missing fine-scale details. The method supports efficient multi-view generation. Quantitative validation on the SARMV3D dataset shows that the refined simulated images achieve low pixel-level error and high structural and perceptual similarity, yielding average values of 23.88 dB for the peak signal-to-noise ratio (PSNR), 0.83 for the structural similarity index measure (SSIM), and 0.20 for learned perceptual image patch similarity (LPIPS). The method provides an effective means for generating multi-view SAR images of buildings, has potential for use in downstream tasks such as target recognition, is physically interpretable, and shows promising generalization capability across the evaluated building scenes.