Developing a deep learning model for the simulation of micro-pollutants in a watershed

Abstract

In recent years, as agricultural activities and types of crops have become diverse, the occurrence of micro-pollutants has been reported more frequently in rural areas. These pollutants have detrimental effects on human health and ecological systems; thus, it is important to manage and monitor their presence in the environment. The modeling approach could be an effective way to understand and manage these pollutants. This study predicts the concentrations of micro-pollutants (MPs) using deep learning (DL) models, and the results are then compared with simulation results obtained from the soil water assessment tool (SWAT) model. The SWAT model showed an unacceptable performance owing to the resulting negative Nash–Sutcliffe efficiency (NSE) values for the simulations. This may be caused by the limitations of SWAT, which pertains to adopting simplified equations to simulate micro-pollutants. In addition, the ambiguous plan of pesticide application increased the model uncertainty, thereby deteriorating the model result. Here, we developed two different DL models: long short-term memory (LSTM) and convolutional neural network (CNN). LSTM exhibited the highest model performance, with NSE values of 0.99 and 0.75 for the training and validation steps, respectively. In the multi-target MP model, the error decreased as the number of simulated pollutants increased. The simulation of the four pollutants had the highest error, while the six-target simulation had the lowest error. In conclusion, this study demonstrated that the LSTM model has the potential to improve the prediction of MPs in aquatic systems.