Cleaning radar satellite images to improve object detection

What is it about?

Synthetic aperture radar, or SAR, is widely used because it can image the Earth at night and through clouds. A common issue in SAR images is speckle, a grainy pattern that makes images harder to interpret and can interfere with automated analysis. In this paper, we introduce an AI-based method to reduce speckle using a diffusion model that cleans images. We guide the training process by starting from easier cases with weaker speckle and gradually moving to harder cases with stronger speckle. We also test two additional strategies: cleaning the image in two stages, and providing a traditional SAR filtered image as an extra input to help preserve important structures. We evaluate the method on both simulated data and real SAR images, and we also test how the cleaned images affect object detection results.

Featured Image

Photo by NASA on Unsplash

Photo by NASA on Unsplash

Why is it important?

Our work shows that combining diffusion models with established SAR knowledge leads to more stable and trustworthy results, especially when models are applied to new sensors or regions that differ from the training data. What is unique about our approach is that it does not rely solely on large amounts of labeled real SAR data. Instead, it uses noise-level guidance, staged denoising, and classical SAR filters to constrain how the model removes speckle. This makes the method more robust to domain shift and easier to deploy in real-world settings where clean reference data are limited. By evaluating both image quality and object detection performance, the study highlights a practical path toward using modern generative models in operational SAR applications rather than only for visual improvement.