Explainable data-driven modeling of suspended sediment concentration at a deltaic marsh boundary under river regulation and storm events

Abstract

Sediment supply is a critical factor influencing the evolution of river deltas and deltaic marshes. This study presents an innovative approach using explainable data-driven modeling, specifically integrating bagged regression trees with Shapley additive explanations (SHAP), to predict suspended sediment concentration (SSC) at the boundary of an intertidal salt marsh in the Yellow River Delta. The model achieves high predictive accuracy, with an R² of 0.978 and an RMSE of 0.099 kg/m³ for training, and an R² of 0.899 and an RMSE of 0.118 kg/m³ for testing, supported by 5-fold cross-validation and ensemble learning. SHAP analysis identifies significant wave height and SSC from the upper river as the main factors influencing SSC at the marsh boundary. The predominant influence of wave heights over other factors suggests that wave-induced local resuspension governs the sediment supply to the marsh, rather than remote sediment advection from the river mouth, which was previously regarded as the primary source. This may explain the continued expansion of the marshes despite a declining riverine sediment discharge in recent years. Additionally, the developed model links SSC at the marsh boundary to key hydrodynamic parameters, allowing for defining dynamic sediment boundary conditions in modeling marsh evolution under changing environments, instead of using oversimplified, static sediment boundary conditions in common practice. By integrating predictive accuracy with interpretability, this method can provide deeper insights into sediment dynamics of the deltaic marsh, therefore supporting comprehensive management of river regulations and delta resilience building.