Dual Polarized SAR Imaging Techniques for Ship Classification

Classifying ships in synthetic aperture radar (SAR) imagery is a challenging task because the targets are often small and the differences between ship categories are subtle. Traditional deep learning methods frequently disregard the polarimetric characteristics of SAR data, which becomes a limitation when working with dual-polarized images. The model was tested on a comprehensive dataset that included both three-class and six-class ship categories. Several deep learning architectures were evaluated, such as ConvNext, ConvNext with attention, VGG16, VGG19, InceptionV3, ResNet, and InceptionResNetV2. For ship detection tasks, different variants of the YOLO architecture were employed, including YOLOv5x6, YOLOv5s6, YOLOv8n, and YOLOv9n.Among the classification methods, the ensemble of Xception and NasNet achieved the highest accuracy, demonstrating its effectiveness in distinguishing between ship types in dual-polarized SAR imagery. On the other hand, the YOLO-based models proved to be highly efficient for real-time detection, highlighting their potential for maritime surveillance and monitoring applications.