Dual-Teacher Cross-Modal Ship Detection With Edge-Constrained SAR-Style Generation
Juanping Wu, Chenxing Mao, Weiwei Guo, Zenghui Zhang · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2026
Ship detection is an important research direction in the intelligent interpretation of remote sensing imagery, with significant applications in marine monitoring, maritime traffic management, and search and rescue. Synthetic aperture radar (SAR) imagery plays a crucial role in maritime scenarios due to its all-weather, day-and-night imaging capabilities, making it a vital data source for ship detection. However, single-modality data are often constrained by imaging conditions and insufficient for detection in complex environments. Optical and SAR images each offer unique advantages, and their complementary nature motivates cross-modal fusion to enhance detection performance. To alleviate the reliance on limited annotated SAR data, this paper proposes an edge-constrained CycleGAN to generate structurally consistent SAR-style pseudo-images from unpaired optical images, thereby enhancing cross-modal feature alignment and increasing data diversity. To further address the complexity of learning unified representations across modalities within the teacher-student framework, a dual-teacher student architecture based on the probabilistic ensembling fusion strategy is designed to extract domain-invariant features from both optical and SAR pseudo-samples and to optimize the student network through collaborative distillation. Experimental results demonstrate that the proposed method achieves better overall detection performance compared to representative semi-supervised methods in multi-source remote sensing ship detection, validating its effectiveness for SAR ship detection under complex remote sensing conditions.