Concentration estimation of dissolved oxygen in Pearl River Basin using input variable selection and machine learning...

Area of Science:

• Environmental Science
• Water Resource Management
• Data Science

Background:

• Dissolved oxygen (DO) is a critical indicator of water quality and ecosystem health.
• Accurate DO prediction is essential for effective water environment assessment and management.
• Existing models may not fully capture the complex, nonlinear relationships between environmental factors and DO.

Purpose of the Study:

• To develop and validate a novel modeling approach for estimating DO concentrations.
• To utilize input variable selection to identify key environmental drivers of DO.
• To construct a robust data-driven model for predicting DO in river basins.

Main Methods:

• Maximal Information Coefficient (MIC) was employed for input variable selection to identify primary environmental factors influencing DO.
• Support Vector Regression (SVR), a data-driven model, was used to build a predictive model for DO concentration.
• The methodology was applied and validated using data from the Pearl River Basin, China.

Main Results:

• The MIC technique effectively screened significant local environmental factors affecting DO, with electrical conductivity (EC) showing a high MIC score.
• Variable reduction using MIC improved SVR model performance, decreasing Root Mean Square Error (RMSE) by 28.65% and increasing R-squared and Nash-Sutcliffe Efficiency (NSE).
• The MIC-SVR model demonstrated superior performance in tidal river networks compared to non-tidal networks, with substantial improvements in RMSE and NSE.

Conclusions:

• The proposed MIC-SVR approach effectively handles nonlinear relationships among environmental factors for accurate DO estimation.
• This method provides a robust tool for water quality assessment, particularly in complex tidal river environments.
• The study highlights the utility of advanced data-driven techniques in environmental monitoring and management.