Improving the sediment and nutrient first-flush prediction and ranking its influencing factors: An integrated machine-learning framework

Pollutant first flush (FF) is a critical phenomenon in urban regions driven by multiple factors. Therefore, having a robust mathematical FF definition and modeling tool along with thorough knowledge of its main influencing variables, is essential. Machine-learning techniques have been widely exploited in the water-quality field; however, their use to predict FF occurrence and pollutant load is limited. This study developed a generic machine-learning framework based on the two most commonly adopted FF definitions (30/80 and M(V) curve) and implemented for three of the most detected pollutants in urban areas (TSS, TN, and TP). This framework is able to accurately predict if a rainfall event generates FF and, in case it does, the correspondent pollutant load (”very good” model performance). It can also quantify and rank the most influencing factors using a non-model-biased method based on game theory (Shapley Additive exPlanations (SHAP)), showing that the rainfall variables are the most significant predictors of sediment and nutrient FF occurrence in urban regions. The findings proved that the machine-learning framework developed in this study can be a valuable modeling tool for improving the accuracy and reliability of pollutant load prediction during FF events in urban areas. Furthermore, they represent one of the first steps towards a robust mathematical FF definition, which is still a topic under debate.