Abstract
Reliable shoreline monitoring is essential for understanding how river deltas respond to changing sediment supply, marine processes, and human activities, and for supporting coastal management. However, accurately locating complex land–water boundaries and maintaining consistency across satellite sensors remain challenges that conventional segmentation metrics alone cannot adequately assess. This study developed MFGR-DeepLabV3+, a lightweight deep-learning model combining progressive feature fusion with boundary refinement to improve shoreline delineation in the Yellow River Delta. Landsat and Sentinel-2 imagery were used to reconstruct instantaneous waterlines for eight epochs between 1990 and 2025, followed by positional validation, cross-sensor comparison, and change analysis using the Digital Shoreline Analysis System (DSAS). Compared with the baseline, the model increased boundary intersection over union by 4.08–4.44 percentage points and reduced average symmetric surface distance by 18.9–43.2%, with benefits concentrated in complex shoreline settings. Validation against manually interpreted shorelines yielded root mean square errors of 18.98–38.13 m, while near-synchronous cross-sensor comparisons showed small mean biases but persistent local discrepancies. Across the study area, 60.09% of transects exhibited net retreat, with a median displacement of −503.51 m despite a positive regional mean. Strong local progradation at the active river mouth contrasted with persistent erosion along Laizhou Bay and the abandoned Diaokou River sector, while engineered shorelines showed mixed responses. By linking boundary-focused learning, positional reliability assessment, and long-term geospatial analysis, this framework strengthens AI-supported shoreline monitoring and provides evidence for identifying priority erosion-monitoring sectors.
Showing the abstract — retrieve the full paper via the Exa API.