Application of stacked ensemble machine learning for the prediction of organic matter content in marginal marine sediments: A case study of the southwestern coastal region of the East Sea

• A stacking ensemble model predicts total organic carbon in sediments. • Ten machine learning algorithms are combined for robust ensemble prediction. • 2,100 stacking models are evaluated per target for optimal configuration. • Domain knowledge-based filtering improves bottom-layer model accuracy. • High accuracy is achieved across spatially complex marine environments. Accurate prediction of organic matter (OM) content and related geochemical parameters in marine sediments is essential for understanding sedimentary processes and evaluating hydrocarbon potential, yet remains challenging due to data heterogeneity and post-depositional modification. This study presents a stacked ensemble machine learning (SEML) framework for predicting Rock-Eval parameters (S1, S2, S3, S1 + S2) and total organic carbon (TOC) in marginal marine sediments from the southwestern East Sea, South Korea. The study area represents a coherent depositional setting characterized by limited riverine OM input and stable hydrodynamic conditions, enabling evaluation of model behavior across sedimentary domains. SEML models were constructed using spatial, bathymetric, and grain-size-related predictors, and a total of 2,100 stacking configurations were systematically evaluated to assess model robustness and convergence rather than performance. Clear contrasts emerged between top and bottom sediment layers. Surface sediments exhibited stronger relationships between fine-grained components and geochemical targets, reflecting limited diagenetic overprinting. In contrast, bottom sediments showed weaker and more heterogeneous predictor–target relationships, consistent with enhanced sediment mixing and post-depositional alteration. Exclusion of anomalous samples from site 23MLSC improved model performance and clarified feature relevance, highlighting the importance of geological context in data preprocessing. Beyond predictive accuracy, ensemble-level uncertainty diagnostics were used to assess model stability and spatial prediction reliability. These diagnostics provide insight into where geochemical relationships are well constrained versus ambiguous. Overall, this study demonstrates that SEML advances beyond conventional stacking approaches by integrating uncertainty-aware evaluation with geological interpretation, offering a robust and interpretable framework for geochemical prediction in heterogeneous sedimentary environments.