Source-Assisted Hierarchical Semantic Calibration Method for Ship Detection Across Different Satellite SAR Images

With the increase of spaceborne synthetic aperture radar (SAR) platforms, numerous SAR images are available for ship detection applications. Traditional deep learning-based detection methods struggle with the distributional disparities in SAR images acquired from different platforms, arising from differences in radar characteristics and data acquisition conditions. Existing approaches employ domain adaptation (DA) techniques to align domain distribution and thus mitigate distribution divergence. However, due to the inherent specificity of SAR images, i.e., ship targets and background environments exhibit highly visual similarity, these methods may inadvertently destroy the discriminative representations of ship targets, resulting in poor cross-domain detection performance. To alleviate this dilemma, we propose a source-assisted hierarchical semantic calibration (SHSC) framework for ship detection across different satellite SAR images. First, a source-assisted semantic calibration module (SSCM) is designed, which performs multilevel semantic calibration by constructing a source-assisted (SA) detector as a guiding mechanism to preserve the discriminative semantics of ship targets. Then, the uncertainty-aware guided feature-level alignment module (UG-FAM) and instance-level alignment module (UG-IAM) are developed, which effectively capture the crucial ship target attributes by emphasizing the learning of those discriminative samples. Extensive experiments are conducted on the datasets obtained from the TerraSAR, Gaofen-3, Sentinel-1, and RadarSat-2 satellites. The experimental results show that the proposed SHSC method outperforms the other UDA approach by an average of more than 3% on AP in ship target detection accuracy across different satellite SAR images.