Water quality modelling using artificial neural network and multivariate statistical techniques

This study investigates and proposes a reduction in the number of water quality monitoring stations, parameters and develops the best input combination for water quality modelling using artificial neural network and multivariate statistical technique. Fourteen water quality physicochemical parameters acquired from eight monitoring sites for 8 years (2006–2013) were investigated. Hierarchical agglomerative cluster analyses (HACA) classify the eight monitoring sites into two significant clusters. Principal component analysis (PCA) accounted for more than 82% of the total variance and attributes the sources of pollution to critical anthropogenic activities, surface run-off and weathering of parent rocks. Furthermore, sensitivity analyses percentage contribution of pollutants revealed dissolved oxygen as the most significant parameter responsible for the pollution (66.3%), followed by ammonia nitrogen (14.4%), chemical oxygen demand (9.4%) and biochemical oxygen demand (5.3%). The result for source category apportionment assigned 39% to rock weathering, 25% anthropogenic activities, 20% surface run-off, 11% faecal waste, 3.4% human and natural factors and 1.4% erosion of river bank. In addition, three input combination models (model 1, 2 and 3) were developed in order to identify the best that can predict water quality index (WQI) at a very high precision. Model 2 using the principal component scores before varimax rotation appears to have the best prediction capability at node eight with coefficient of determination (R2) = 0.999 and root mean square error (RMSE) = 0.159. These findings justify the use of environmetrics modelling technique to reveal the pattern of water quality for decision making by government and stakeholders.