Data-driven prediction for fluvial organic carbon in riverine sediments: A quasi-global scale simulation

Suspended sediments in riverine systems are critical for transporting organic carbon (OC) from terrestrial to aquatic and marine environments, playing a pivotal role in the global carbon cycle. However, predicting sediment-associated OC at large spatial scales is challenging due to complex interactions among geochemical, climatic, and land surface processes. This study introduces an optimized data-driven modeling framework using the Global River Sediment (GloRiSe) database as a proxy to estimate OC concentrations in riverine suspended sediments. A suite of machine learning (ML) models was implemented, and their internal hyperparameters were optimized using a heuristic framework integrated with feature selection, enhanced model parsimony, and interpretability. The results show that ensemble-based models achieved the highest predictive performance, with mean R² values exceeding 0.75 and the lowest root mean square error (RMSE). Feature importance analysis consistently identified iron oxide (Fe₂O₃), aluminum oxide (Al₂O₃), and phosphorus pentoxide (P₂O₅) as dominant predictors, highlighting their roles in stabilizing and sorbing organic carbon in fluvial systems. Monte Carlo simulations confirmed the robustness of ensemble models under input perturbations. This integrative framework demonstrates the potential of optimized ML models to enhance the predictive understanding of sedimentary carbon dynamics globally. The methodology provides a scalable and interpretable approach for informing future assessments of carbon transport and land–ocean connectivity, with implications for climate modeling, watershed management, and biogeochemical forecasting.