Ship detection in SAR imagery is crucial for maritime surveillance and safety, offering high-resolution imaging regardless of weather or lighting conditions. While deep learning has enhanced detection capabilities, challenges persist in nearshore environments where ships exhibit higher mis-detection rates compared to offshore scenarios. Existing approaches typically employ general detection models for both contexts, overlooking nearshore-specific challenges. In this paper, we propose a novel two-stage training strategy for SAR ship detection that specifically addresses nearshore detection limitations. Our approach leverages a foundation model trained on diverse maritime conditions as a pre-trained baseline, then develops a nearshore-specialized model through transfer learning. To optimize performance, we implement comprehensive data augmentation techniques and evaluate various architectural enhancements including Feature Pyramid Network (FPN), Bidirectional FPN (BiFPN), Ghost Bottleneck (GCB), and Path Aggregation Network (PANet) for improved multi-scale feature detection. Through extensive ablation studies, we demonstrate consistent performance improvements of the nearshore-specialized model over the foundation model across all architectures. Our optimal configuration, YOLOv8n + PANet, achieves significant performance improvements in nearshore scenarios with a 10% increase in recall and a 3.97% improvement in mAP50, while maintaining robust offshore detection capabilities. This research advances SAR ship detection technology and establishes a framework for specialized detection in challenging maritime environments.