First Flush Occurrence Prediction and Ranking of Its Influential Variables in Urban Watersheds: Evaluation of XGBoost and SHAP Techniques

In the last decades, the quality of the surface-water bodies has been increasingly threatened mainly due to anthropogenic activities (i.e., urbanization and the consequent increase of impervious surfaces). This leads to the generation of runoff even more rich in pollutants and difficult to adequately treat in the specific first-flush treatment plant. Consequently, accurate water-quality predictions in urban areas and related detections of events that may generate first flush are the key to enhancing urban water management and pollution control. In this study, the ability of a supervised machine-learning technique (XGBoost) to predict the occurrence of first flush was demonstrated. The model showed outstanding performance in predicting first flush for three of the most detected pollutants in urban areas (total suspended solids, total nitrogen, and total phosphorus) by using rainfall-runoff variables as input. Furthermore, by exploiting a non-model-biased method based on game theory (SHAP), such variables were quantified and ranked based on their level of importance in pollutant first-flush predictions. The findings of this work proved that the XGBoost model is a functional tool for enhancing the accuracy of first-flush predictions in urban watersheds and, thus, contributes to the development of an effective design of first-flush treatment plants.