Motamarri and Boccelli (2012) proposed the application of learning vector quantization (LVQ), a direct classification approach, for comparison with multivariate linear regression (MLR) and ANN approaches and integrates input selection for model improvement by concerning primary and secondary water quality standards. They showed that ANN performance was more similar to LVQ when a larger number of explanatory variables were utilized, but the ANN performance degraded toward MLR performance as explanatory variables were removed. Overall, the use of LVQ as a direct classifier provided the best overall classification ability to violated/non-violated samples for both standards.
Verma et al. (2013) applied different classification techniques in data mining to create day-ahead, time series forecast models for total suspended solids (TSS) in wastewater. They examined various scenarios of daily average influent carbonaceous biochemical oxygen demand (CBOD) and influent flow rate measured at 15 min intervals. Then, they used five data-mining algorithms, i.e., multi-layered perceptron, KNN, multi-variate adaptive regression spline, SVM, and random forest, to construct day-ahead, time-series prediction models for TSS. Historical TSS values were employed as input parameters to predict current and future values of TSS. They indicated that accurate predictions are feasible if sufficient data are available.
Kovcs et al. (2014) performed an investigation that consolidated the use of a clustering technique and discriminant analysis to identify similar groups of water quality samples from a lake in Austria/Hungary. They combined cluster and discriminant analysis (CCDA) to facilitate deciding whether the further division of classification is necessary or not for searching of homogeneous groups.
Liu and Lu (2014) compared ANN and SVM techniques to predict total nitrogen (TN) and total phosphorus (TP) from a river influenced by agricultural drainages. River flow, water temperature, flow time, rainfall, dissolved oxygen, and upstream TN or TP concentrations were selected as initial inputs of the two models. Their results indicated that the proposed SVM models performed better in generalization due to the lower risk of overfitting and the ability to optimize fewer parameters based on the structural risk minimization (SRM) principle.
Modaresi and Araghinejad (2014) presented a survey to classify the water quality using classification algorithms to reduce computational time. They investigated and compared the performance of three supervised methods of classification including SVM, PNN, and KNN. They considered two performance evaluation statistics of error rate and error value. Their results showed that the SVM algorithm presents the best performance with no errors in calibration and validation phases.
Mohammadpour et al. (2015) predicted water quality index in constructed wetlands using SVM and two types of ANNs, namely feed-forward backpropagation (FFBP) and radial basis function (RBF) techniques. Seventeen observation points in the wetland were monitored twice a month for 14 months, and an extensive dataset was collected for 11 water quality variables. A detailed comparison of the overall performance showed that the prediction of the SVM model was more suitable than ANN. They highlighted that the SVM and FFBP can be successfully employed for the prediction of water quality in a free surface constructed wetland environment. This work pushed Muharemi et al. (2018) to remodel SVM and ANN as it promises to give better results. They evaluated some popular classification algorithms to model a water quality detection system. They had trained a set of models using classification algorithms: logistic regression, SVM, Linear Discriminant Analysis, and ANN algorithm. Muharemi et al. (2018) inferred that past experiments on both algorithms SVM and ANN proved to be very useful in ecologically-related problems, but in their research, the merits of the results of Logistic Regression goes to interaction terms. They found that interaction terms can have a huge impact when we have collinearity between many variables, and we have some prior knowledge on the domain.
Zhang et al. (2017) proposed a novel anomaly detection algorithm for water quality data using dual time-moving windows, which can distinguish anomaly data from historical patterns in real-time. The algorithm was based on a statistical model such as autoregressive linear combination. The algorithm has been tested using 3-months water quality data of pH from a real water quality monitoring station in a river system. Experimental results showed that the algorithms can significantly decrease the rate of false-positive alarms in their system.
Detecting outliers is a difficult task. In contamination detection problem, for example, using an input dataset which includes outliers makes the detection of the anomalies very difficult, as the final model of the used algorithm may not be robust enough (Domingues et al. 2017). So, in line with Kiang’s (2003) suggestion, this paper is conducted by four different algorithms are used for data mining and classification of water quality data of two different aquatic systems: A river and a water distribution network. The examined algorithms are based on different methods, namely statistical methods, nearest-neighbor methods, and data isolation methods. As presented, the KNN algorithm has been used to some extent in the field of aquatic systems, but the other three proposed algorithms in this paper, namely local outlier factor (LOF), Isolation Forest (iForest), and Anomaly Detection, have not been studied in water systems. The performance of these algorithms is also evaluated by Recall, Precision, and F-measure indices. The structure of this paper is as follows: Section 2 introduces the methodology of the research. In Section 3, the results of the used methods are presented and discussed. Section 4 summarizes the results and includes suggestions for future studies.
