Architecture-preserving post-training boundary refinement for RGB sea–land segmentation: cross-architecture transfer and supervision-scope trade-offs
Boundary errors can persist in RGB water/nonwater segmentation even when region overlap is high. Occupancy-routed boundary calibration (ORBC) refines a trained model through a single epoch of updates to its decoder and prediction head. Ground-truth water occupancy restricts a morphological boundary Dice loss to patches with 10% to 90% water, whereas region supervision remains active for every patch. The encoder is frozen, and the input, inference graph, parameter count, and operation count remain unchanged. On SWED, mean boundary F1 gains at a two-pixel tolerance range from 0.0046 to 0.0069 for SegFormer-B0, U-Net-R34, and DeepLabV3+-R34 relative to matched control continuations from the same base within each architecture, under the respective official-test or validation protocols. In matched SegFormer comparisons, boundary gains accompany degraded low-water segmentation and increased false-water predictions on pure nonwater patches. For mixed Coast Train records, higher pooled boundary F1 coexists with lower site- and tile-macro means. ORBC provides a short refinement procedure for retained models when boundary improvements justify the associated region errors.
Authors
- Wei Sun (ORCID: https://orcid.org/0000-0003-1166-5836)
- Yuting Li
Institutions
- Shanghai Maritime University (CN)
Publication Details
- Journal
- Journal of Electronic Imaging
- Published
- 2026-11-14
- DOI
- https://doi.org/10.1117/1.jei.35.6.063010
- Primary Topic
- Remote-Sensing Image Classification
- Type
- article
- Field-Weighted Citation Impact
- 0.00