Large-scale nonpoint source (NPS) pesticide pollution is a growing concern in urban areas; however, modeling of such pollution is constrained by challenges in acquiring urban pipeline data and the scarcity of pollutant monitoring data. This study presents a hybrid model comprising a rainfall runoff module based on a modified gated recurrent unit and a pesticide concentration module grounded in physical process equations to assess NPS pesticide pollution in large urban areas, adopting Guangzhou City as a case study. The model parameters were calibrated and validated using monitored runoff volumes and pesticide concentrations, employing a stochastic gradient descent algorithm. The results indicated that the developed model performed well, matching or exceeding the performance of traditional NPS models in small urban areas. NPS pesticide pollution in this area exhibited spatiotemporal characteristics impacted by meteorological conditions. Washoff loads were positively correlated with maximum pesticide concentrations and runoff volumes but not when they were preceded by dry periods. The initial rainfall intensity, rather than the total rainfall volume, affected pesticide washoff amounts. The findings of the study provide insight into urban NPS pesticide pollution and its causes, while the model shows promise for modeling emerging pollutants in any urban area.
Keywords: hybrid model; large-scale modeling; machine learning; pesticide pollution; urban nonpoint source.