Comparative evaluation of limit-of-detection-based and K-nearest-neighbors-based classification approaches for...

Area of Science:

• Environmental Science
• Analytical Chemistry
• Data Science

Background:

• Wastewater treatment plant (WWTP) effluent contamination poses risks to public health and ecosystems.
• Optical whiteners and fluorescence are used as indicators, but long-term machine learning studies are scarce.
• Effective detection methods are crucial for safeguarding aquatic environments.

Purpose of the Study:

• To compare limit of detection (LOD)-based and k-nearest neighbors (KNN)-based machine learning methods for detecting WWTP effluent.
• To evaluate the effectiveness of fluorescence excitation-emission matrices (EEMs) and physicochemical parameters for water quality monitoring.
• To assess the robustness of machine learning models under varying environmental conditions.

Main Methods:

• Collected water samples over 20 months from river sites, WWTP effluent, and impacted downstream locations in Korea.
• Employed fluorescence excitation-emission matrices (EEMs) and conventional physicochemical parameters.
• Utilized LOD-based and KNN-based classification algorithms for data analysis.

Main Results:

• KNN achieved optimal accuracies of 98% and 91%, outperforming LOD-based methods (71-94%) under variable conditions.
• Tryptophan-like (Peak T) and microbial humic-like (Peak M) fluorescence were key indicators.
• Conductivity was the most influential physicochemical parameter.
• Minimum detectable effluent mixing ratios for 90% accuracy ranged from 13-23%.

Conclusions:

• Machine learning, particularly KNN, offers a robust and accurate approach for detecting WWTP effluent contamination.
• Fluorescence spectroscopy combined with machine learning enhances water quality monitoring capabilities.
• This study highlights the potential for improved environmental protection through advanced analytical techniques.