Using surrogates to predict trace organic contaminant removal in stormwater biofilters

Stormwater biofilters are increasingly implemented as nature-based solutions for mitigating trace-level organic contaminants (TrOCs) in urban stormwater. However, TrOC removal in biofilters is often inconsistent because performance varies with compound-specific properties and rainfall-driven system conditions. Reliable TrOC removal prediction is therefore needed to enhance stormwater management. This study evaluated the feasibility of pollutant-based and sensor-based surrogates for predicting the removal of ten TrOCs in biofilters, using data from 26 rainfall events collected over 1.5 years. Surrogate inputs were combined with general descriptors and molecular fingerprints representing TrOC properties and evaluated using machine-learning algorithms. Tree-based ensemble algorithms paired with molecular fingerprints achieved high predictive accuracy, with validation Nash-Sutcliffe efficiency (NSE) values frequently exceeding 0.85. Reduced pollutant-based surrogate sets retained predictive performance, with median NSE decreasing from approximately 0.91 for the full surrogate set to 0.82 for a single surrogate. Among pollutant-based surrogates, ultraviolet absorbance at 254 nm (UVA254), dissolved oxygen (DO), and total nitrogen (TN) were the strongest individual predictors, achieving median NSE values of 0.90, 0.88, and 0.85, respectively, while their combinations offered reliable performance. Sensor-derived event-level descriptors of oxidation-reduction potential (ORP) and soil moisture in the submerged zone (SZ) emerged as robust sensor-based surrogates. Independent validation using historical-rainfall events further supported the robustness of these surrogates, with SZ_ORP and the UVA254 + DO set achieving NSE > 0.70 and showing strong practical potential. Overall, this work establishes a scalable and cost-effective surrogate-based framework for predicting TrOC removal and supporting performance assessment of stormwater biofilters.