Apparent Shoreline-Change Mapping with Frequency-Aware Feature Decoupling and Semantic Guidance

Mapping apparent visible-waterline change from bi-temporal optical imagery requires distinguishing narrow boundary responses from sea-surface texture, radiometric variation, and residual registration error. This study presents a direct binary mapping network that combines frequency-aware feature decoupling with deep semantic guidance. At four encoder scales, fixed 3×3 average pooling separates a locally smooth component from the element-wise absolute temporal feature difference; subtraction yields a complementary residual. Mamba processes the smooth component, spatial attention reweights the residual, and deep semantic features guide decoder skips. The target is the pixel-wise exclusive OR of two independently annotated, finite-width visible-waterline masks. The dataset contains 174 pairs of 512×512-pixel patches with nominal 2 m pixel spacing. Patches do not overlap within each source-image pair, while the manually constructed training, validation, and test subsets can share source scenes. On the reported 122/17/35 partition, the method achieved 91.45% Precision, 89.62% Recall, 90.53% pixel F1, and 93.35% boundary F1 at a 2-pixel tolerance. The reported pixel F1 was 2.68 percentage points higher than ChangeMamba. These are descriptive within-scene comparisons: 33 of the 35 test pairs have an empty shoreline mask at one date, and the references predominantly contain straight or piecewise-linear geometry. The scores consequently characterize the specified annotation-based target rather than geographically independent generalization or physical shoreline displacement. No tidal-datum correction was applied.