Using machine learning to reveal spatiotemporal complexity and driving forces of water quality changes in Hong Kong marine water

A self-organizing map (SOM), which is an unsupervised neural network, is used to explore the spatiotemporal variability of water quality in Hong Kong marine water areas based on 31 years of monitoring data (1986–2016). In this study, three types of SOMs, referred to as s-v, s-t and t-s SOM, respectively, are applied to the multivariate marine water quality data, and principal component analysis (PCA) is used to help the clustering of SOM neurons and component planes. The major findings revealed by the spatiotemporal SOM analyses include the following: a) Hong Kong marine water areas can be classified into five regions with distinctive water quality characteristics over the long term, b) the spatiotemporal variations in chlorophyll-a (Chl-a) and NO are greatly affected by ocean currents and nutrients from Pearl River discharge, c) marine water quality is significantly affected by water control projects, e.g., the Hong Kong Harbor Area Treatment Scheme (HATS), and d) nitrogen (N) is the limiting nutrient for phytoplankton growth in winter, west and south of the Hong Kong marine regions, while phosphorus (P) is the limiting nutrient in summer, owing to the massive discharge of the Pearl River that contain NO. Furthermore, the severe pollution in Deep Bay with high levels of Chl-a, nutrients and fecal coliform (FC) deserves immediate attention from the governments. The results of this study prove that combined spatiotemporal SOM analyses provide a powerful tool to identify spatiotemporal patterns in water quality data and to reveal the driving forces behind water quality changes. This study presents the first attempt to apply SOMs to water quality analysis at both spatial and temporal domains.