Leveraging Machine Learning for accurate and interpretable suspended sediment concentration predictions

Suspended sediment concentration (SSC) significantly impacts water quality, aquatic ecosystems, and reservoir capacity, making accurate prediction vital for effective watershed management. Traditional empirical and physically based models often struggle to handle the complexities and non-linear dynamics of sediment transport. Machine learning (ML) techniques, with their ability to model non-linear relationships and process large datasets, offer a promising alternative. This study explores the application of ML models, including extra trees (ET), random forest (RF), categorical boosting (CatBoost), and extreme gradient boosting (XGBoost) and their combination with genetic programming (GP), to predict SSC. Key environmental variables such as precipitation, streamflow, and seasonality are used as inputs, and the models are trained and validated using historical hydrological data. The SHapley Additive exPlanations (SHAP) framework is employed to interpret the models, offering insights into the influence of each input variable on SSC predictions. Results demonstrate that ML models outperform traditional approaches in accuracy and robustness, particularly in capturing peak sediment events. The findings underline the potential of ML in improving SSC prediction and guiding sustainable watershed management practices.

Keywords: Suspended Sediment Concentration (SSC), Machine Learning (ML), SHAP Values, Hydrological Modeling, Sediment Transport, Watershed Management.