Predicting bioretention pollutant removal efficiency with design features: A data-driven approach

The objective of this study is to synthesize previous research findings from bioretention experiments and identify design features that lead to the best performance of bioretention pollutant removal with a data-driven approach. A bioretention database was built from 79 bioretention publications, composed of 182 records of bioretention cells with their design features and the corresponding pollutant removal efficiency data. Non-parametric correlation analysis, multiple linear regression (MLR), and decision tree classifiers were applied to investigate the relationships between bioretention design features and pollutant removal efficiencies. Non-parametric statistics and MLR results indicated that bioretention surface area, media depth, the presence of an internal water storage (IWS) layer, soil composition, and vegetation cover are all significantly correlated with pollutant removal efficiencies. The impacts of design features are significantly different under different climate and inflow conditions. Decision tree classifiers showed that non-vegetated bioretention cells with sand filter media generally have higher than 80% total suspended solid (TSS) mass removal efficiencies; bioretention cells with minimum organic matter and greater than 0.58 m soil media depth tend to remove more than 51% of total nitrogen (TN); and vegetated bioretention cells with minimum organic matter remove more than 67% of total phosphorus (TP). The overall accuracy of decision tree classifiers in the test set is around 70% to predict TSS, TN, and TP mass removal efficiency classes. This study suggests that the data-driven approach provides insights into understanding the complex relationship between bioretention design features and pollutant removal performance.