Diffusion models generate synthetic SAR data for new satellite sensors

A team led by Xuanting Wu from Nanjing University of Aeronautics and Astronautics has developed a novel framework to generate synthetic aperture radar (SAR) data for new satellite sensors using diffusion models and feature migration. Published in the Transactions of Nanjing University of Aeronautics and Astronautics (2025) and available on arXiv, the method addresses a critical challenge: different SAR sensors vary in resolution, polarization, and frequency bands, making existing models unusable for newly launched satellites that lack sufficient labeled data.

The framework integrates a stable diffusion model with low-rank adaptation (LoRA) modules inside a multimodal diffusion transformer (MM-DiT) architecture, enabling class-controllable SAR image generation from textual prompts. To ensure generated images statistically match the target sensor, the authors introduce an attention distillation mechanism that transfers sensor-specific features—such as spatial texture, speckle distribution, and structural patterns—from real target-domain data into the generative model. Extensive experiments on multi-class aircraft targets from two real spaceborne SAR systems demonstrate that the approach effectively alleviates data scarcity, supporting cross-sensor remote sensing applications without needing large volumes of new labeled data.