AIS-Information-Aided Ship Association and Identification in SAR Images

To address the lack of identity attributes for ship targets in synthetic aperture radar (SAR) imagery and the heterogeneous temporal, spatial, and geometric errors between SAR imagery and automatic identification system (AIS) data, a multi-dimensional feature-based stepwise association method is proposed for ship identity identification in SAR imagery. The method is designed to mitigate temporal asynchrony, ship-shape deviation, and locally non-uniform spatial distortion during SAR AIS data fusion. AIS observations within 30 min before and after SAR imaging are used to construct continuous ship trajectories through cubic spline interpolation, and the ship positions corresponding to the SAR imaging time are reconstructed to reduce motion-induced temporal matching errors. An instance segmentation network is used to extract pixel-level ship masks and determine target positions from mask centroids. To address local protrusions and asymmetric shapes caused by side-looking imaging, sidelobe effects, and strong scattering from ship superstructures, principal component analysis (PCA) is employed to estimate the ship principal axis, and an iterative principal-axis correction method is developed to reduce the influence of asymmetric protrusion regions on orientation estimation. The corrected ship principal axis, length, and width are then used as geometric features. After SAR and AIS targets are transformed into a common geographic coordinate system, iterative closest point (ICP) is applied for global coarse registration. K-means is subsequently used to partition the scene into locally distorted regions according to the spatial distribution of targets, while the product of ship length and width is introduced as a scale feature for regional ICP matching. Accurate association is finally performed in a joint position scale feature space. Experiments are conducted using real SAR imagery and AIS data collected in the Wusongkou and Yangtze River estuary waters of Shanghai, including 72 463 AIS records from 1 171 vessels. The instance segmentation network achieves a precision of 0.874, a recall of 0.727, and a mean average precision at an intersection over union threshold of 0.5 (mAP50) of 0.821. The average distance error between SAR and AIS targets decreases from 2.469×10^-3(°) before calibration to 0.368×10^-3(°) after applying the proposed method, corresponding to reductions of 74.7%, 77.2%, and 77.8% compared with conventional ICP, nearest-neighbor matching, and the Hungarian algorithm, respectively. The results demonstrate that the proposed method can jointly exploit AIS trajectory information, SAR target positions, and ship-scale features while mitigating the effects of AIS temporal asynchrony, SAR ship-shape deviation, and locally non-uniform geometric distortion over large-area imagery. The method is particularly suitable for complex waters characterized by dense ship traffic, closely spaced targets, and spatially varying positioning errors, and can provide technical support for ship identity assignment in SAR imagery, maritime traffic supervision, abnormal-target screening, and multi-source maritime situational awareness in emergency search and rescue.