Large-scale bathymetry in high-turbidity rivers enabled by remote sensing and artificial intelligence.

Water depth, as a direct indicator of a river's continuous underwater topography, provides crucial information for studies such as investigating channel morphological evolution, modeling sediment transport, and quantifying material flux. Conventional approaches to acquiring localized water depth, involving in situ measurements or methods based on remote sensing imagery, are mostly employed in clear rivers at small scales. However, these techniques encounter substantial limitations when applied to extensive river reaches with high-sediment-concentration flow, where accurately determining water depth is generally only achievable through in situ field measurements. Here we present an intelligent model named RivDepth, designed to obtain water depth distributions in rivers with relatively high suspended sediment concentrations (SSC) exceeding 1 kg m-3. By integrating satellite-acquired spectral variables and an optically derived SSC proxy as inputs, together with an "AI expert" module capable of inference, decision-making, and prediction, the proposed model captures the coupled depth-reflectance-SSC relationship patterns of sample pixels. This enables pixel-wise retrieval of large-scale water depth distributions in high-SSC rivers. RivDepth was trained and tested on the lower Yellow River, one of the rivers with the highest SSC in the world. The proposed model delivered accurate depth estimates, achieving an R 2 of 0.896, RMSE of 0.456 m, MAE of 0.228 m, and ME of -0.020 m. Shapley additive explanations-based feature-importance analysis indicates that shortwave infrared bands, the red band, a red-edge band, the water vapor band, the aerosol/blue band, and the SSC proxy are among the most influential predictors of water depth in this high-SSC river. Their contributions vary spatially in a complex, non-monotonic manner. This study presents a feasible and robust method for large-scale water-depth retrieval in rivers with high SSC. Furthermore, it provides a valuable reference for large-scale hydrological monitoring and basin-scale integrated management by integrating remote sensing imagery with in situ observations.