Modeling of microbial contamination in the Marmara Sea, Bursa-Turkey

Aim: The main objective of this study was to design and develop the feed forward neural network (FNN) model structures for forecasting of faecal coliform concentrations and microbial water quality in Gemlik, Karacabey and Mudanya coastal areas alongside the Sea of Marmara, Turkey. Methodology: Artificial neural networks (ANNs) are modeling tools for environmental parameters, especially water quality and provide working of inter-related multi parameters. In this study, 4 model structures were implemented to forecast the faecal coliform concentrations for the sea coasts of "Gemlik, Karacabey and Mudanya" alongside the Marmara Sea. Total coliform and faecal streptococci were input parameters. The Levenberg Marquardt algorithm was applied for training the modeling studies. The results of the models were crosschecked with the real concentrations according to performance functions root mean squared error (RMSE). Results: Comparison of the modeling results with the measured concentrations demonstrated that established model structures provided correct results. (R) Correlation coefficients were determined between 0.57 and 0.98. It was observed that during the trials enhancing the hidden layer counts in the model structures did not increase the model performance in each test. Kind and count of inputs affected the model productivity. The growing rates of the coliform group bacteria were dissimilar because, different types of contaminants in the seawater affect the metabolism. The error values of the forecasting results applied in Gemlik and Mudanya Coasts were larger because there were large quantities of pollution loads and pollutant diversities. Interpretation: The developed model structures could predict the microbial contamination in the coastal environments and provided information on the more effective integrated sea coast management and protection of human health.