Bright shadow detection in cross-polarized SAR analysis for moving ships discrimination

Accurate estimation of vessel motion in synthetic aperture radar (SAR) imagery is essential for maritime surveillance and navigation safety. In this study, we introduce a novel deep learning-based framework for the automated segmentation of ship-induced “bright shadows” in cross-polarized SAR images, harnessing these delineated features as a proxy for motion estimation. We leverage the newly compiled XShadowBright dataset, which encompasses a wide variety of vessel types, radar incidence angles, and sea-state conditions, to train a U-Net architecture equipped with a ResNet-18 encoder for robust feature extraction. The network was optimized using a pixel-wise cross entropy loss and trained over twenty epochs with the Adam optimizer, incorporating weight decay and gradient clipping to ensure stable convergence. On an independent test set, our approach attains a Dice Coefficient of 0.87, indicating high overlap between predicted segmentation masks and manual annotations across diverse imaging scenarios. Furthermore, when deployed on a standard NVIDIA A100-SXM4-40 GB GPU, the model processes approximately twenty $1024 \times 1024-$ pixel SAR scenes per second, demonstrating the real-time capability required for operational applications. Compared to traditional ground moving target indication (GMTI) techniques, this low-complexity segmentationbased method offers a compelling alternative for vessel tracking in cross-polarized configurations. Overall, our findings establish a reliable and efficient pipeline for bright shadow segmentation, paving the way for enhanced SAR-based vessel monitoring in the maritime domain.