Leveraging Supervised Learning to Optimize Urban Greening Strategies for Combined Sewer Overflow Pollution Reduction

Many cities adopt greening strategies to reduce contamination from combined sewer overflows (CSOs). Nonetheless, quantifying the impact of urban greening on CSO-affected water quality at the city scale remains challenging. To address this challenge, this work leveraged supervised learning to link water swimmability with the greening of a CSO shed (the drainage area of a CSO outfall), using New York City (NYC) as a case study. Random forest classification models were built to predict water swimmability after rainfall at 46 sites in NYC water bodies impacted by CSOs. A 14-feature model (AUROC =0.81, accuracy = 0.78) revealed that greening improved local water quality. However, water flow speed, antecedent rain depth, and CSO shed area were also influential. A simplified four-feature model (AUROC = 0.8, accuracy = 0.75) explored links between levels of greening and the probability of non-swimmable waters (Pns) following different 18 h rainfall depths. Increased greening was found to be most impactful in reducing Pns for CSO sheds discharging to water bodies with flow speeds < 6 cm/s. For CSO sheds discharging to water bodies with flow speeds ≥ 14.7 cm/s, urban greening had no impact on Pns. The work illustrates the utility of supervised learning in supporting citywide decisions regarding urban greening investments.