Deep neural network for marine water quality classification with the consideration of coastal current circulation effect

Neural Network has been widely used to model the dynamics of chlorophyll-a concentration for over a decade. Previous studies were always based on shallow network structures (i.e. 3–5 layers) and used time-lagged data from the localized region as model inputs. Recent ecological studies have shown that the coastal ocean current circulation is one of the key factors for the formation of algae bloom (red tide), hence the level of chlorophyll-a concentration increases. This suggests that the data from nearby regions should be included in the modeling process along with the localized data. This study investigates the classification performance among models with and without the use of data from nearby regions under deep neural network learning models. Networks with 3 to 12 layers are employed in two distinct structures respectively for conducting the empirical analysis on 1990–2016 monthly marine water quality data obtained from the Hong Kong Environmental Protection Department. The deep networks are shown to be able to extract useful information from 108 input attributes consolidated from 5 nearby coastal regions in Deep Bay water control zone. The 5-fold cross valuation results indicated that the use of additional layers tends to improve the classification performance and the optimal result is achieved by using 11 layers under the two proposed network structures.