High-Confidence Clustering and Lightweight U-Net for SAR Change Detection

Synthetic aperture radar (SAR) image change detection (CD) involves identifying changes between images captured at different times over the same geographical region. SAR provides significant advantages for disaster-related remote sensing due to its all-weather capabilities and ability to penetrate clouds and darkness. However, Synthetic aperture radar change detection under severe speckle is challenging yet essential for reliable all-weather monitoring. We introduce a hybrid framework that fuses classical and deep learning techniques: after edge-preserving denoising, three standard difference maps are combined into a single change indicator, on which a two-cluster Fuzzy C-Means selects high-confidence “changed” and “unchanged” pixels as training seeds. A compact three-level Mini-U-Net, trained with a weighted Binary Cross-Entropy and Dice loss, then refines the remaining uncertain areas via sliding-window inference. Experiments on three diverse SAR datasets demonstrate that our method achieves higher overall accuracy, F1 score, and IoU than state-of-the-art approaches, delivering robust speckle suppression and precise change delineation.