Author(s): Cosimo Russo; Alberto Castro; Angela Gorgoglione

Linked Author(s): Angela Gorgoglione

Keywords: First flush; Urban runoff; Random forest; Hydroinformatics

Abstract: The first flush (FF) has been recognized and investigated as a typical phenomenon of urban areas since it represents one of the most critical non-point source pollutions and, therefore, can negatively impact the quality of receiving water bodies. For this reason, the analysis of urban runoff has become a critical factor in the protection of surface water quality. Based on this, model simulation and assessment represent an essential procedure to estimate the strength of the FF effect in urban regions. With this aim, complex physically-based models have been commonly adopted, but their application is limited by the data availability. In this study, we assess the use of a data-driven technique, random forest (RF), to overcome this issue. Such technique is considered an attractive alternative since it is flexible, “quick learner,” and performs well in multisource data prediction problems. Based on these considerations, the objective of this study is twofold: 1) developing a machine-learning algorithm, based on RF technique, able to predict whether a rainfall event can generate sediment first flush, taking into account variables that characterize the precipitation event (dry and wet period), 2) in case of FF occurrence, predicting the sediment load for such storm event. The results showed that the RF model predicts very well the existence of FF (F1 score = 0.78 and accuracy = 0.87) and the sediment event mean load (NSE = 0.92). The outcomes of this study are expected to contribute to the development of accurate and reliable stormwater-quality models and, consequently, effective stormwater treatment design.

DOI: https://doi.org/10.3850/IAHR-39WC252171192022952

Year: 2022