Attention driven reinforcement guided U Net for reliable SAR image segmentation with uncertainty estimation

Synthetic Aperture Radar (SAR) imagery is widely used for flood monitoring, oil spill detection, and coastal mapping. However, automatic SAR image segmentation remains challenging due to speckle noise, low signal-to-clutter ratio, and ambiguous object boundaries. To address these challenges, we propose R-CLAU-Net, a hybrid segmentation framework that integrates an attention-enhanced U-Net backbone with a reinforcement-guided Cellular Learning Automata (CLA) module. The attention-enhanced backbone improves discriminative feature extraction by emphasizing salient regions and suppressing irrelevant speckle-induced responses, while the CLA module performs output-level refinement to enhance boundary continuity and spatial consistency. Monte Carlo Dropout is further employed for pixel-wise uncertainty estimation, enabling confidence-aware predictions under noisy and ambiguous conditions. Experiments on a SAR dataset covering diverse scenes show that R-CLAU-Net achieves 83.5 ± 0.4% IoU, 90.0 ± 0.3% Dice, and 76.1 ± 0.5% Boundary IoU (BIoU). Ablation results demonstrate that removing attention gates, the CLA module, or uncertainty modeling degrades performance, confirming the complementary contribution of each component to region accuracy and boundary delineation. Robustness evaluations under seasonal shifts, noise perturbations, and unseen geographic regions, together with real-world case studies, further support the practical applicability of the proposed method for reliable SAR image segmentation.