Mapping spatiotemporal dynamics of offshore targets using SAR images and deep learning

Accurate understanding of the dynamic distribution of maritime targets is pivotal for optimizing marine spatial planning and advancing the sustainable utilization of marine resources. However, existing algorithms for ship and offshore wind turbine (OWT) detection face three critical challenges: (1) limited accuracy in the collaborative detection of heterogeneous targets; (2) compromised robustness under complex maritime environments; and (3) insufficient spatiotemporal correlation analysis during long-term monitoring. To address these issues, this study proposes an enhanced YOLOv8n architecture. First, the backbone network incorporates an LSKNet module that simultaneously captures local contextual features and global structural information through adaptive receptive field selection and multi-scale feature aggregation. Second, a ResBlock-CBAM hybrid module is cascaded after the C2f layer in the neck network to enhance discriminative feature extraction and mitigate gradient degradation in deeper layers. Third, the detection head integrates an Asymptotic Feature Pyramid Network (AFPN) to establish a cross-level direct interaction mechanism, thereby preserving critical semantic information. Utilizing decadal satellite monitoring data (2015–2024), the proposed framework reveals significant spatiotemporal differentiation patterns in marine activity distributions through ship trajectory hotspot analysis and OWT density mapping. Experimental results indicate that the method achieves a mean average precision (mAP) of 94.6% in complex marine environments (a 7.8% improvement over the baseline YOLOv8n), a 4% increase in recall for small targets (<30 pixels), and a 6% reduction in the false detection rate. This study offers technical and data-driven insights to support marine spatial governance.