Applications of Artificial Neural Network in Predicting Water Quality Indicators: Case Studies from Korean Coastal Waters

Chemical oxygen demand (COD) and dissolved oxygen (DO) and are critical indicators representing water quality in coastal environments. This study attempts to predict COD and DO in two coastal waters of Han River estuary (HRE) and Masan Bay (MB), South Korea, using big data from automatic monitoring of water quality measured every five minutes for years. For HRE, artificial neural network (ANN) model classifier using Random Forest was developed to generate predictions of COD from 29 input parameters in air, river and ocean. We trained the ANN model with 80% of dataset with shuffled sampling and validated the model with the rest 20% of dataset. Accuracy of the model is 91.65% for all parameters and 88.81 % for top six important parameters (chlorophyll-a, pH, month, water temperature, water level, DO). For MB, the feed-forward, multi-layer perceptron and single hidden layer network topology, scaled conjugate gradient learning algorithm and the sigmoid transfer function in the hidden layer were used. 80% of dataset (2013-2016) was used for training the ANN model and the rest 20% dataset (2017) was used for validation. The optimal number of neurons in the hidden layer was determined iteratively based on training set performance. ANN model can predict DO in MB with high performance. Further research will be focusing on hypoxia less than 3.0 mg/L of DO predicted in advance of one or two days, which is beneficial for coastal management.