Structure-guided cross-site semantic segmentation of shipyard remote-sensing imagery

Semantic segmentation of shipyard remote-sensing imagery is challenging under distribution shifts between sites. Ambiguous class boundaries and fragmented region interiors are particular concerns. We evaluate a structure-guided framework based on SegFormer-B2. A morphology-derived boundary constraint addresses class transitions. A prototype-based region-consistency constraint addresses fragmented predictions within semantic regions. In the main evaluation across five test sites and three random seeds, the proposed model improves on SegFormer-B2. Mean intersection over union increases from 0.6793 to 0.6907. Foreground mean intersection over union rises from 0.6763 to 0.6889, and the mean Dice coefficient rises from 0.8053 to 0.8129. Gains in all three metrics remain positive in all 10 site-assignment comparisons across two validation-site assignments. The boundary F-score at a one-pixel tolerance increases from 0.7190 to 0.7393. The small-component ratio decreases from 0.0775 to 0.0631. Both auxiliary branches are removed at inference, leaving the SegFormer-B2 deployment graph unchanged at 27.35 million parameters and 92.1 billion floating-point operations. The results support complementary improvements in boundary agreement and region coherence across the five evaluated shipyards without additional inference complexity.