Geomorphology-constrained bathymetric inversion of turbid tidal channels using centerline-network priors and machine learning

Bathymetric inversion in turbid tidal channels remains challenging because of degraded bottom-related optical signals and complex-branched channel networks. This study investigates whether geomorphic structure can act as a stable prior for bathymetric reconstruction when the spectral-depth relationship breaks down. We develop a geomorphology-constrained framework that integrates UAV multispectral imagery, suspended sediment concentration, RTK-derived channel depths, and adjacent flat DEM constraints. A branching centerline network is first extracted to characterize channel organization, from which geometric and network-structural predictors, including normalized width, width-to-depth ratio, downstream accumulation, tributary order, and distance to junctions, are derived and used to construct a random-forest depth prior. Then, these predictors are refined through constrained optimization with anisotropic smoothing along channel directions and boundary consistency with exposed terrain. The framework is tested in a turbid tidal-channel system along the southern Yellow Sea and further evaluated through a transfer experiment in the Yangtze Estuary. Among all predictors, the width-to-depth ratio provides the greatest improvement in inversion accuracy, reducing RMSE from 0.30 m to 0.18 m (38 %), highlighting the dominantly geomorphological control on channel cross-sectional geometry. Overall, The provided framework achieves RMSE values of 0.15 m (RTK-based) and 0.23 m (DEM-based), outperforming empirical spectral, conventional machine-learning, and deep-learning approaches, particularly in scour zones, narrow tributaries, and channel confluences. In the Yangtze Estuary, the model attains an RMSE of 3.46 m (R² = 0.70), suggesting preliminary transferability across more than three orders (from ∼ 10 m to 10 km) of magnitude in channel scale. The results show that the main bottleneck in turbid tidal-channel inversion is not only optical degradation, but also the lack of geomorphic organization in model design. By coupling data-driven learning with structural and boundary constraints, this study advances a transferable inversion paradigm for optically complex coastal environments.