Features Selection in Water Quality Prediction in Neural Network using Canonical Correspondence Analysis (CCA)

Irrelevant inputs can cause deterioration of the network performance. This paper aims to implement features selection for water quality prediction model as the neural network generalization may be improved when the number of measured variables is reduced. Canonical Correspondence Analysis (CCA) was introduced as a feature selection method, where it choose a subset of input variables by eliminating features with little or no predictive information. Data monitoring and sampling process was carried out at 5 sampling sites of Perak river basin. The water quality data with 28 different parameters was used to determine the relative effect for each parameter to the biological oxygen demand (BOD) and chemical oxygen demand (COD) as they represent the river water quality; determination of the pollutants strength in terms of the oxygen necessity to reduce the domestic and industrial wastes, hence giving an initial value how much biodegradable waste existed in the water. While, for COD, its application is practical since it verify the amount of total oxidizable compounds in water. Results of CCA through the biplot diagram showed that suspended solid (SS), turbidity (TUR), total solid (TS), nitrate (NO3) and zinc (Zn) were most important environmental factors that influencing the BOD and COD. The study continued with a number of neural network approaches implemented for predicting BOD and COD. The results showed comparable performance of water quality prediction with CCA implementation in the network compared to ANN standalone model.