ABSTRACT
The water quality index (WQI) is a widely used tool for comprehensive assessment of river environments. However, its calculation involves numerous water quality parameters, making sample collection and laboratory analysis time-consuming and costly. This study aimed to identify key water parameters and the most reliable prediction models that could provide maximum accuracy using minimal indicators. Water quality from 2020 to 2023 were collected including nine biophysical and chemical indicators in seventeen rivers in Yancheng and Nantong, two coastal cities in Jiangsu Province, China, adjacent to the Yellow Sea. Linear regression and seven machine learning models (Artificial Neural Network (ANN), Self-Organizing Maps (SOM), K-Nearest Neighbor (KNN), Support Vector Machines (SVM), Random Forest (RF), Extreme Gradient Boosting (XGB) and Stochastic Gradient Boosting (SGB)) were developed to predict WQI using different groups of input variables based on correlation analysis. The results indicated that water quality improved from 2020 to 2022 but deteriorated in 2023, with inland stations exhibiting better conditions than coastal ones, particularly in terms of turbidity and nutrients. The water environment was comparatively better in Nantong than in Yancheng, with mean WQI values of approximately 55.3-72.0 and 56.4-67.3, respectively. The classifications "Good" and "Medium" accounted for 80 % of the records, with no instances of "Excellent" and 2 % classified as "Bad". The performance of all prediction models, except for SOM, improved with the addition of input variables, achieving R2 values higher than 0.99 in models such as SVM, RF, XGB, and SGB. The most reliable models were RF and XGB with key parameters of total phosphorus (TP), ammonia nitrogen (AN), and dissolved oxygen (DO) (R2 = 0.98 and 0.91 for training and testing phase) for predicting WQI values, and RF using TP and AN (accuracy higher than 85 %) for WQI grades. The prediction accuracy for "Medium" and "Low" water quality grades was highest at 90 %, followed by the "Good" level at 70 %. The model results could contribute to efficient water quality evaluation by identifying key water parameters and facilitating effective water quality management in river basins.
ABSTRACT
Lake water-level fluctuation is a complex and dynamic process, characterized by high stochasticity and nonlinearity, and difficult to model and forecast. In recent years, applications of machine learning (ML) models have yielded substantial progress in forecasting lake water-level fluctuations. This paper presents a comprehensive review of the applications of ML models for modeling water-level dynamics in lakes. Among the many existing ML models, seven popular ML model types are reviewed: (1) artificial neural network (ANN); (2) support vector machine (SVM); (3) artificial neuro-fuzzy inference system (ANFIS); (4) hybrid models, such as hybrid wavelet-artificial neural network (WA-ANN) model, hybrid wavelet-artificial neuro-fuzzy inference system (WA-ANFIS) model, and hybrid wavelet-support vector machine (WA-SVM) model; (5) evolutionary models, such as gene expression programming (GEP) and genetic programming (GP); (6) extreme learning machine (ELM); and (7) deep learning (DL). Model inputs, data split, model performance criteria, and model inter-comparison as well as the associated issues are discussed. The advantages and limitations of the established ML models are also discussed. Some specific directions for future research are also offered. This review provides a new vision for hydrologists and water resources planners for sustainable management of lakes.