Main content area

Artificial neural networks as emulators of process-based models to analyse bathing water quality in estuaries

García-Alba, Javier, Bárcena, Javier F., Ugarteburu, Carlos, García, Andrés
Water research 2019 v.150 pp. 283-295
Escherichia coli, coasts, estuaries, indicator species, neural networks, water quality, Spain
This study aims to provide a method for developing artificial neural networks in estuaries as emulators of process-based models to analyse bathing water quality and its variability over time and space. The methodology forecasts the concentration of faecal indicator organisms, integrating the accuracy and reliability of field measurements, the spatial and temporal resolution of process-based modelling, and the decrease in computational costs by artificial neural networks whilst preserving the accuracy of results. Thus, the overall approach integrates a coupled hydrodynamic-bacteriological model previously calibrated with field data at the bathing sites into a low-order emulator by using artificial neural networks, which are trained by the process-based model outputs. The application of the method to the Eo Estuary, located on the northwestern coast of Spain, demonstrated that artificial neural networks are viable surrogates of highly nonlinear process-based models and highly variable forcings. The results showed that the process-based model and the neural networks conveniently reproduced the measurements of Escherichia coli (E. coli) concentrations, indicating a slightly better fit for the process-based model (R2 = 0.87) than for the neural networks (R2 = 0.83). This application also highlighted that during the model setup of both predictive tools, the computational time of the process-based approach was 0.78 times lower than that of the artificial neural networks (ANNs) approach due to the additional time spent on ANN development. Conversely, the computational costs of forecasting are considerably reduced by the neural networks compared with the process-based model, with a decrease in hours of 25, 600, 3900, and 31633 times for forecasting 1 h, 1 day, 1 month, and 1 bathing season, respectively. Therefore, the longer the forecasting period, the greater the reduction in computational time by artificial neural networks.