Researchers have developed a novel artificial intelligence (AI) framework, named RivDepth, that can map river depth across long reaches where high suspended sediment concentrations challenge conventional satellite bathymetry. The method, detailed in a study published in Environmental Science and Ecotechnology, combines Sentinel-2 satellite spectral information with an optically derived suspended sediment concentration (SSC) proxy to retrieve water depth pixel by pixel. Tested on a 786-kilometer stretch of the lower Yellow River, one of the world's most sediment-laden rivers, RivDepth demonstrated high accuracy in capturing complex relationships among water depth, reflectance, and sediment load.
RivDepth's core innovation is an adaptive AI expert module that integrates four machine learning models: parallel random forest (PRF), extreme gradient boosting (XGBoost), support vector regression (SVR), and multilayer perceptron (MLP). Instead of applying a single model uniformly, RivDepth performs preliminary prediction, inference, and decision-making to select the most suitable strategy for each pixel based on water conditions. This pixel-level adaptability is crucial for rivers like the Yellow River, where suspended sediment, flow structure, and optical signals vary sharply over long distances.
The study, conducted by researchers from the State Key Laboratory of Hydroscience and Engineering at Tsinghua University and other Chinese institutes, used Sentinel-2 Level-2A imagery, field-measured cross-sectional elevation data, water-level records, and in situ SSC observations to train and validate the model. Shapley additive explanations (SHAP) analysis identified key predictors including shortwave infrared bands, red and red-edge bands, the water vapor band, the aerosol/blue band, and the SSC proxy. Cloud-affected pixels were reconstructed to improve coverage, and water bodies, channel centerlines, thalwegs, and river widths were extracted from satellite imagery.
By learning different depth–reflectance–SSC patterns and choosing prediction strategies at the pixel level, RivDepth can adapt to spatially changing sediment and channel conditions. This capability turns routine satellite observations into actionable depth information for river science and management. More frequent and continuous bathymetric information could help track channel change, identify thalweg migration, improve sediment-transport modeling, and support flood-risk and habitat assessments. The approach offers a scalable tool for integrated watershed monitoring and management.
RivDepth can be further improved as higher-resolution satellite imagery and more accurate spatial SSC indicators become available. With broader validation, the workflow may be adapted to other turbid river systems. The study was accepted for publication on May 20, 2026, with DOI: 10.1016/j.ese.2026.100711. The research was supported by the Team Key Project of the State Key Laboratory of Hydroscience and Engineering and the National Natural Science Foundation of China.
