Estimation of sodium adsorption ratio in a river with kernel-based and decision-tree models

Area of Science:

• Environmental Science
• Water Resource Management
• Agricultural Science

Background:

• Surface water quality control is vital for effective water resource management.
• Sodium adsorption ratio (SAR) is a key water quality parameter for agricultural productivity.
• Traditional chemical analysis for SAR can be time-consuming and impractical for large-scale river systems.

Purpose of the Study:

• To evaluate machine learning models for estimating SAR in the Aladag River, Turkey.
• To identify the most effective model and minimal input parameters for SAR estimation.
• To assess the feasibility of using simplified models for water quality assessment.

Main Methods:

• Comparison of Support Vector Regression (SVR), K-Nearest Neighbour (KNN), and decision tree algorithms (Hoeffding tree, Random Forest, Random Tree, REPTree).
• Utilized 15 distinct water quality parameters as inputs, with subsequent analysis focusing on fewer parameters, including only sodium (Na).
• Employed correlation matrix and sensitivity analysis to determine significant model inputs.
• Model performance was evaluated using correlation coefficient, Nash-Sutcliffe efficiency, RMSE, MAE, and Willmott's index.

Main Results:

• The Support Vector Regression (SVR) model with a poly kernel function demonstrated the highest accuracy in SAR estimation.
• No significant difference in accuracy was observed between models using all 15 parameters and those using only the sodium (Na) parameter.
• Machine learning approaches enable accurate SAR determination with minimal input data.

Conclusions:

• SVR with a poly kernel is a highly effective method for estimating SAR.
• Accurate SAR prediction is achievable using machine learning models with sodium (Na) as the sole input parameter.
• This approach offers a practical alternative to traditional chemical analysis for water quality assessment in agricultural contexts.