Prior Density Map Guided Feature Modulation for Ship Detection in SAR Imagery

Ship detection in synthetic aperture radar (SAR) imagery is a critical task for maritime surveillance, but it still faces challenges such as small-sized targets, strong nearshore clutter, and limited discriminative information from single-modal SAR data. Most existing deep learning methods improve detection bymodifying network structures or introducing attentionmechanisms, while ignoring valuable spatial distribution priors of ship occurrences. In this article, we propose a prior density map guided feature modulation framework for SAR ship detection, which introduces spatial prior knowledge at the input level without extra parameters or architectural changes. First, a hybrid prior density map is constructed via kernel density estimation from training annotations, capturing both local and global ship distribution patterns. Then, two fusion strategies are designed: prior density channel (PDC) and prior density guidedmodulation (PDGM).PDCdirectly uses the normalized density map as a dedicated input channel to provide pure and unentangled prior cues, which is proven more effective than PDGM. Extensive experiments on SSDD and HRSID show that the optimally configured PDC (σ=10, α=0.1) boosts YOLOv8n by +1.6/+15.4 on mAP@0.5/mAP@0.5:0.95 for SSDD and +9.5/+18.3 for HRSID, with real-time inference. Stratified evaluation reveals that gains concentrate in inshore subsets (+18.7% mAP50 on HRSID inshore) and on small targets. We further verify portability across five detector families (CNN and Transformer), with PDC delivering +14.8 mAP50:95 on RTDETR- L, and characterize the operating scope via cross-dataset experiments: PDC is designed for persistent monitoring scenarios with stable spatial distributions.