I. Introduction
The deterioration of water quality (WQ), caused by human activities and global warming, is a major environmental issue worldwide. The WQ directly affects human health, social and economic development. Monitoring the continuous changes of WQ has been a priority for the meaningful decision making [1, 2]. Traditionally, the existing WQ assessment collects discrete water samples at monthly or weekly intervals, because analyzing them in the laboratory would be tediously time-consuming. Such low-frequency data makes it hard to develop a subsequent model for early warning. Currently, sensor-based buoys and boats can autonomously measure water-related information in real-time, including different physical, chemical, and biological parameters. In large water bodies, a network of sensors with GIS systems has also been used to collect in-situ high-frequency WQ data, which offers a comprehensive insight into the spatiotemporal variability of the target ecosystem. However, the issue of missing data is relatively unavoidable due to problems from equipment ageing, routine maintenance, insufficient collection, and power outage of the monitor. Hence, WQ data from sensors network is commonly high-dimensional and incomplete (HDI) nature. Handling HDI WQ data would create more opportunities to understand the data-intensive processes in the studied ecosystem. [3–5]. Choosing appropriate methods for imputing HDI WQ data is still challenging [6, 7].