UAV SAR drops the straight-and-level assumption: a diffusion prior plus motion compensation holds 21.3 dB at half physical resolution

Synthetic aperture radar imaging from a low-altitude ISAC UAV no longer depends on the uniform-linear-motion assumption, because imaging under an arbitrary known three-dimensional trajectory is recast as an exact linear inverse problem. Conventional range-Doppler and chirp scaling algorithms assume uniform linear motion and approximately decouple range from azimuth processing; when a UAV flies a flexible 3D path to balance communication coverage against sensing geometry, pixels within one range cell see different acceleration projections and higher-order range variations, so a shared range cell migration correction cannot compensate all of them and a common azimuth matched filter cannot focus them, producing range defocusing and azimuth smearing. Across an ideal uniform-linear-motion path, a smooth curved path and a Z-shaped path with the same path length and the same number of slow-time samples, the method varies by only 1.50 dB in PSNR and less than 0.024 in SSIM at 20 dB SNR; on the Z-shaped path it retains 36.39 dB PSNR and 0.969 SSIM, while the best baseline, PnP-BM3D, reaches only 16.62 dB and 0.681.

The imaging pipeline separates the physical observation model from the image prior: a pretrained diffusion model learns the prior from SAR reflectivity data alone, while the echoes enforce measurement consistency through likelihood guidance during reverse diffusion, so changing trajectory only requires updating the measurement matrix rather than retraining per scene. Earlier compressed-sensing methods used handcrafted sparsity priors that fail to capture the non-Gaussian statistics and spatial structure of real SAR scenes, and supervised echo-to-image networks bind their learned mapping to the sensing geometry seen in training, so a change of trajectory, MIMO configuration or OFDM sampling strategy forces retraining. The diffusion prior is trained on the FAIR-CSAR dataset, which provides 12,000 training and 2,700 testing single-look complex images of scenes such as airports and harbors; inference uses DDIM acceleration, and 20 sampling steps closely match the quality of the full 1,000-step sampler.

For unknown position errors caused by wind and rotor vibration, image reconstruction and motion compensation alternate inside the same reverse diffusion process: the current noisy state yields a clean reflectivity estimate, that estimate refines the position errors, and the corrected measurement operator then improves the data-consistency guidance for later steps. Traditional motion compensation and autofocus methods rely on isolated dominant scatterers and assume spatially invariant phase errors, which breaks down under the slow-time and space-variant coupling of MIMO-OFDM spotlight mode; optimizing the time-domain position-error matrix directly involves many variables and is highly sensitive to unreliable intermediate reflectivity estimates. Position errors are reparameterized in the discrete cosine transform domain so only low-order modes are optimized, with a coarse-to-fine spectral schedule; at an injected position-error root-mean-square of 0.5 cm and 20 dB SNR, the joint method reaches 23.65 dB PSNR and 0.91 SSIM with a residual position-error RMSE of 0.19 cm and a mean 3D Euclidean error of 0.27 cm, whereas PnP-BM3D shows an RMSE of 2.94 cm.

The results come from numerical simulation: a UAV acts as a dual-functional aerial base station serving ground communication users while performing spotlight SAR imaging from downlink OFDM communication echoes, over a 100 m by 100 m target scene discretized onto a 256 by 256 grid with 0.39 m physical pixels. The intended readers are researchers and engineers working on ISAC, UAV remote sensing and generative inverse-problem solvers. The simulated synthetic aperture is short, which limits the accumulation of navigation and motion errors; taking this pipeline to a real platform would next require validation on real flight data.

All conclusions come from simulation, with no validation on real UAV flight data. Position-error compensation keeps a low residual RMSE and only slowly declining PSNR as the injected error grows to 0.75 cm, and degrades further at 1 cm while the residual RMSE stays around 0.16 cm; a small residual remains even at zero injected error because motion compensation stays active, so noise and intermediate estimation errors produce nonzero updates and create an estimation floor. Under bandwidth truncation the default guidance parameter performs sub-optimally and must be lowered from 2 to 1 to raise the relative weight of the diffusion prior.