Predicting microplastic transport in open channels with different bed types and river regulation with machine learning techniques

Abstract

Microplastic transport in open channels with varying bed (acrylic, medium sand, and cobble) and regulation (weirs and sluice gates) configurations for three types of microplastics (densities: 910–1350 kg/m3 ) with varying discharges were studied in a laboratory setup. In general, bivariate correlations between retention and independent variables (e.g., discharge) were weak. Also, retention was dependent on the type of river regulation. The sluice gate retained statistically fewer microplastics compared to the weir, excluding the influence of bed conditions. Nevertheless, it permitted the escape of two microplastic types to the downstream. These complex observations were well predicted in machine learning models and validation using a gravel bed confirmed model robustness. Decision Tree performed best among individual models, with a co-efficient of correlation (R2 ) of 0.99 and mean square error (MSE) of 2.00. The stacking ensemble model combining Decision Tree, Random Forest, and XGB achieved a similar accuracy (R2 of 0.99, MSE of 2.00) while reducing the likelihood of overfitting. These results clearly demonstrate the strength of stacking in enhancing both predictive performance and generalization capability. SHapley Additive exPlanations (SHAP) analysis, including dependency plots highlighted density and bed roughness as key factors of retention. The developed ML model offers a resilient predictive tool for understanding microplastic transport, suggesting ways to address pollution through river management.