Maritime Surveillance in SAR Data Using Multiscale Attention Models

Abstract

A novel hybrid deep learning framework for ship detection in Synthetic Aperture Radar (SAR) satellite imagery is proposed, integrating MobileNetV3 for lightweight and efficient feature extraction, Feature Pyramid Networks (FPNs) for multi-scale feature fusion, and an optimized YOLOv8n object detection module for precise and real-time ship localization. This framework addresses key challenges in SAR-based maritime detection, including varying ship scales, low-contrast targets, and complex noisy sea backgrounds. To enhance model generalization under limited annotated data, the system leverages transfer learning using pre-trained MobileNetV3 weights and applies robust data augmentation techniques, such as geometric transformations and noise simulations. Experimental evaluations conducted on a dedicated SAR ship detection dataset demonstrate that the proposed model achieves an average precision of approximately 90.4% at an IoU threshold of 0.5, while significantly reducing false alarm rates compared to conventional methods. The model's real-time inference ability and minimal computing overhead make it ideal for deployment on platforms with limited resources, like UAVs and onboard satellite systems. The findings demonstrate the framework's promise as a dependable and effective solution for automated vessel monitoring, coastal security, and marine surveillance applications.