RLGD: A High-Resolution UAV Dataset and Benchmark for Semantic Segmentation and Revetment Extraction in Waterway Scenes

High-resolution unmanned aerial vehicle (UAV) imagery is increasingly used for waterway inspection, but automated scene understanding remains constrained by the lack of pixel-level datasets that explicitly represent revetments and their complex interfaces with water, vegetation, roads, and riverside structures. We introduce the Riverway and Landscape Ground Dataset (RLGD), a dedicated UAV semantic segmentation dataset for waterway scenes. RLGD contains 100 images at 3000×4000 pixels, corresponding to 1.2 billion labeled pixels. Its public RLGD taxonomy explicitly includes revetment. Seven semantic segmentation models are evaluated under a unified protocol. Across five training runs with different random seeds, mean mIoU ranges from 69.27% to 77.02%, while Hierarchical Joint Network with Boundary Guidance (GHJNet) achieves 77.02 ± 0.41% mIoU and 83.51 ± 1.18% revetment IoU. Direct transfer from Varied Drone Dataset (VDD) to RLGD results in relative mIoU reductions of 11.29–49.81%, whereas adaptation using 20% of the RLGD training source images closes 65.5% of the measured domain gap. For the GHJNet run used in the downstream analysis, object-prior refinement using the Segment Anything Model (SAM) increases revetment IoU from 84.10% to 87.25%. A 42-image visual-completeness screening case study further achieves 83.93% macro-F1 and 81.25% balanced accuracy against image-level interpretations by three experts. Overall, RLGD provides a densely annotated, task-specific benchmark for dry-season waterway semantic segmentation, revetment extraction, and image-level inspection prioritization within the surveyed Yangtze River corridor.