Document Type : Research Paper

Authors

• Maryam AriaSadr ¹
• Dariush Rahimi ²
• Mehran Zand ³
• Hadi Amiri ⁴

¹ Ph.D Graduate of Cliamatology, University of Isfahan, Isfahan, Iran

² Professor of Department of Physical Geography, University of Isfahan, Isfahan, Iran

³ Associate Professor, Soil Conservation and Watershed Management Research Institute, Agricultural Research, Education and Extension Organization (AREEO), Tehran, Iran

⁴ Associate Professor of Department of Economics, University of Isfahan, Isfahan, Iran

Abstract

Introduction
The complexity of aquatic systems, their spatiotemporal variations, the high cost and time-consuming nature of traditional testing methods, and the need for continuous monitoring all contribute to the difficulty of monitoring and evaluating water quality. Therefore, artificial intelligence, machine learning, and deep learning approaches are useful for predicting water quality given the diverse parameters involved and are more cost-effective compared to traditional testing methods. There is no uniform algorithm that performs optimally for predicting water quality, and different algorithms exhibit superior performance in different contexts. The quality of rivers in the Zagros mountainous region has decreased due to the erosion of karst formations, their passage through residential areas, changes in land use, drought, and climate change. Some of these rivers, such as the Karkheh, Karun, and Dez, serve as the source of drinking water for a population of over 10 million people. Therefore, changes in water quality pose increased risks and threats to the drinking water supply of these settlements. Assessing the quality of water resources and developing a suitable model will play an effective role in managing the required water supply. The present study aims to identify the best machine learning algorithm for estimating the water quality parameters of the Cham Anjir watershed while facilitating ongoing monitoring of the resource.
 
Materials and methods
Data from 321 samples of discharge and water quality parameters at the Cham Anjir hydrometric station during the period 1969–2021 were used to select a suitable artificial intelligence model and to assessthe quality of surface water resources in the Cham Anjir watershed. These samples included physical and chemical indicators: TDS, TH, SAR, Na, Mg, Ca, and Cl. Machine learning algorithms were employed to model the water quality of the Cham Anjir watershed. Through iterative testing of different models, the Support Vector Machine (SVM) and Classification and Regression Tree (CART) models were selected as the best performers. A correlation matrix was used to evaluate relationships among the variables, and based on these correlations, monthly discharge and monthly water quality indices—including TDS, TH, SAR, Na, Mg, Ca, Cl, EC, %Na, pH, HCO₃, and temporary hardness—were analyzed. A total of 321 monthly samples of the two key indices, TDS and TH, were studied over the statistical period of 1969–2021. To evaluate the accuracy of the machine learning algorithms in water quality modeling, the following performance indices were used: coefficient of determination (R²), mean squared error (MSE), and mean absolute error (MAE).
Results and discussion
The p-value from trend tests confirmed the existence of an increasing trend at the 95% confidence level for the two variables TDS and TH (surface water quality parameters of the Cham Anjir watershed) since 1985. The average TDS in the first period (1969–1984) was 286.6 mg/L, and in the second period (1985–2021) it was 422.08 mg/L. The average TH in the first period (1969–1984) was 181.5 mg/L, and in the second period (1985–2021) it was 278.6 mg/L.
The results of the correlation matrix indicated that TDS has a strong correlation with EC and TH. TH has a positive correlation with TDS, EC, HCO₃, Ca, Cl, and Mg , and an inverse correlation with pH. The machine learning algorithms SVM and CART successfully captured the increasing trend for the two parameters of total hardness and total dissolved solids in the discharge of the Cham Anjir watershed. The SVM algorithm with the linear kernel exhibited the best performance. In addition, the root mean square error (RMSE) validation index also showed lower values for the SVM algorithm compared to the CART algorithm. Therefore, SVM is more accurate in predicting water quality indicators.
 
Conclusion
Machine learning algorithms are effective for water quality modeling. The results indicated an increasing trend in TDS and TH concentrations in the Cham Anjir watershed since 1985. SVM performed better than CART in predicting TDS and TH. The findings suggest that decreasing river flow, increasing water consumption, and karst geology influence TH and TDS concentrations in the quality of this watershed. Water quality monitoring is essential for water resource management. For future estimates of water quality in this context, the SVM algorithm is recommended.

Keywords

• Decision tree
• Support vector machine
• Total hardness
• Validation
• Water quality