Geo-adaptive deep learning for regional-scale bathymetric modeling in optically and morphologically challenging waters: A case study of the Yellow River Estuary

Study region: The Yellow River Estuary, China, is a shallow and highly turbid coastal zone with complex optical conditions that challenge traditional bathymetric inversion methods. Study focus: This study develops a Geo-Adaptive Deep Learning (GADL) framework for multispectral bathymetric inversion, ensuring consistent performance across optically shallow and deep waters. Two physics-informed indices, the substrate resistance index (SRI) and the sediment concentration adjustment index (SAI), are introduced to mitigate the effects of substrate heterogeneity and sediment scattering. A compact multispectral feature set covering visible to near-infrared bands is employed to capture the radiative convergence and divergence effects induced by seabed topography. The model adaptively retrieves depth from seabed reflectance in shallow waters and from spectral reflectance intensity in deep waters, maintaining stability under complex optical conditions. New hydrological insights for the region: The method is particularly effective for the Yellow River Estuary, which is characterized by shallow water depth and high sediment concentration. GADL achieved a root mean square error of 0.62 m, improving accuracy by 39.8 % compared with baseline models. Feature attribution analysis identified SRI as the most influential predictor (SHAP=3.483), while SAI showed nonlinear corrective effects between 0.04 and 0.81. The framework maintained stable accuracy within 0–20 m depths and sediment concentrations up to 500 mg/L, with an RMSE of 0.74 m in optically deep waters, demonstrating strong adaptability in complex coastal environments.