Water Quality Evaluation of the Yangtze River in China Using Machine Learning Techniques and Data Monitoring on Different Time Scales

Abstract

Unlike developed countries, China has a nationally unified water environment standard and a specific watershed protection bureau to perform water quality evaluation. It is a major challenge to assess the water quality of a large watershed at a wide spatial scale and to make decisions in a scientific way. In 2016, weekly and real-time data for four monitoring indicators (pH, dissolved oxygen, permanganate index, and ammonia nitrogen) were collected at 21 surface water sections (sites) of the Yangtze River Basin, China. Results showed that one site had a relatively low Site Water Quality Index and was polluted for 12 weeks meanwhile. By using expectation-maximization clustering and hierarchical clustering algorithms, the 21 sites were classified. Variable spatiotemporal distribution characteristics for water quality and pollutants were found; some sites exhibited similar water quality variations on the weekly scale, but had different yearly grades. The results revealed polluted water quality for short periods and abrupt anomalies, which imply potential pollution sources and negative effects on water ecosystems. Potential spatio-temporal water quality characteristics, explored by machine learning methods and evidenced by time series and statistical models, could be applied in environmental decision support systems to make watershed management more objective, reliable, and powerful. and to evaluate alternative hypotheses regarding important contaminant sources and watershed properties that control transport over large spatial scales. However, there is no unified model available for the Chinese government to make decisions regarding large watershed management [12]. Local governments have a variety of environmental models that they can choose from, and may select diverse models that do not allow meaningful comparisons with the results of models chosen in other areas [13]. Even in the same area, different departments of the same local government use different models with diverse data to assess the water quality of the same river basin, resulting in a huge amount of variability in the water quality evaluation reports, which often fail to reach the same conclusion [3,4].Thanks to the increased collection and use of data, data-driven approaches have been playing an increasingly important role in water management [14]. Statistical and numerical models enable environmental decision support systems (EDSS) to be more reliable and powerful in coping with real-world environmental systems [15]. Real-time data are widely used in urban water management and by water utilities in developed countries [16][17][18][19][20], but rarely in rural watershed management, especially in large watershed management [21,22]. In China, the rapidly growing economy and population is generating widely distributed polluted surface water throughout the country. Thus, there is an increasing need for online data for large watershed management to meet the objectives of early warning monitoring of surface water quality, and for monitoring and control of tot...

Using Real-Time Data and Unsupervised Machine Learning Techniques to Study Large-Scale Spatio–Temporal Characteristics of Wastewater Discharges and their Influence on Surface Water Quality in the Yangtze River Basin

“…The YRB sites were clustered based on the yearly means of COD, NH 3 -N, pH, and DO in surface water in 2016 (EM_SA Method) and 2017 (EM_SB Method). The EM_SA Method was the same as the EM_Y Method in our previous study [11]. The YRB sewage outlets were clustered based on the yearly means of COD, NH 3 -N, and pH in wastewater discharges and the EM models used were named EM_A Method (COD, NH 3 -N, and pH data in 2016 were input), EM_B Method (COD and NH 3 -N data in 2016 were input), EM_C Method (COD, NH 3 -N, and pH data in 2017 were input), or EM_D Method (COD and NH 3 -N data in 2017 were input).…”

“…Because of a lack of spatial and temporal data, the relationships between point and non-point pollutant sources and water quality have only been studied at the microscale in the past [7,8]. Luckily, at present, more data are available and data-driven approaches and statistical (or numerical) models are now playing an increasingly important role in water management, so that environmental decision support systems (EDSSs) are more reliable and are capable of coping with real-world environmental systems [9][10][11]. Numerous researchers have analyzed real-time data to support the management of urban water and water supplies in developed countries [12][13][14][15], but this approach has not often been used to manage large rural watersheds or wastewater [16,17].…”

“…Clustering algorithms, as established unsupervised machine learning models, have been used to analyze data from a wide range of disciplines, such as gene expression data in biology and stock market financial data [18,19], yet have been rarely applied to the water environment due to the lack of data [11,[20][21][22][23]. The partition-based, hierarchical, and density-based algorithms are all popular spatial clustering methods [24].…”

“…Its steps of calculation and assignment repeat iteratively until the objective function obtains the required precision [28]. Low-dimensional data can be analyzed successfully with the EM clustering algorithm, and this approach is particularly useful when the only data available for training a probabilistic model are incomplete [11,18,28,29]. Therefore, the EM algorithm, as an improvement of k-means, is an ideal unsupervised machine learning method for clustering pollution conditions [30,31] and has been proven as objective and efficient in evaluating surface water quality of the Yangtze river in China in our earlier study [11].…”

Soft Sensor with Adaptive Algorithm for Filter Gain Correction in the Online Monitoring System of a Polluted River

“…With this approach, we were not able to observe the dynamic changes in water quality resulting from random extortions. A similar approach was presented in [5], in which the authors developed and proposed methods for assessing water quality for management purposes of local governments, with data obtained on a weekly or annual basis. However, this solution did not take into account the impact of anomalous pollution changes, which indicates the need for continuous online measurements.In [6], the authors proposed real-time monitoring based on online measurements using mobile measurement stations.…”