Abstract: ObjectivesTo address the insufficient accuracy of navigable water-surface region detection for uncrewed surface vehicles in inland waterways caused by strong water-surface reflections and complex shoreline geometries, a segmentation method for navigable water regions in complex inland scenarios is investigated. Methods An encoder–decoder network with feature-selective and shoreline-enhanced network (FSSENet) is proposed. Multi-scale context modeling and feature fusion are adopted to enhance the perception of water-surface obstacles at different scales. A bidirectional feature discriminative enhancement module (BFDEM) is introduced to mitigate inter-class feature confusion caused by water-surface reflections through forward and reverse discriminative matrices. Meanwhile, a shoreline perception probability gradient constraint (SPGC) is incorporated to improve segmentation stability and boundary consistency in shoreline areas. Results Experimental results on the USVLand and LaRS datasets demonstrate that the proposed method achieves significant improvements over several mainstream semantic segmentation approaches, with a precision (Pr) of 75.2% and coastline segmentation accuracy μ_r of 66.5% on the LaRS dataset.Conclusions The results demonstrate that a segmentation framework combining multi-scale feature modeling, discriminative feature enhancement, and shoreline geometric constraints can effectively improve navigable water-region detection performance for unmanned surface vehicles in complex inland waterway environments.