Comparison of machine learning algorithms to predict dissolved oxygen in an urban stream

Area of Science:

• Environmental Science
• Water Resource Management
• Machine Learning Applications

Background:

• Urbanization significantly impacts water quality in urban watersheds.
• Effective monitoring systems are crucial for identifying and mitigating these impacts.
• Low-cost sensors offer a viable solution for widespread water quality data collection.

Purpose of the Study:

• To identify a predictive machine learning model for dissolved oxygen (DO) using data from low-cost sensors.
• To establish a monitoring system for the urban stream network in Hunnicutt Creek, Clemson, SC.
• To compare the performance of multiple linear regression with various machine learning algorithms.

Main Methods:

• Evaluated k-nearest neighbor, decision tree, random forest, and gradient boosting algorithms against multiple linear regression.
• Used water temperature, conductivity, turbidity, and water level change as input parameters.
• Assessed model performance using the Nash-Sutcliffe model efficiency coefficient (NSE).

Main Results:

• The random forest algorithm demonstrated the highest performance in predicting DO across all four study sites.
• NSE scores exceeded 0.9 at three sites and 0.598 at the fourth site, indicating strong predictive accuracy.
• Explainable AI (XAI) revealed temperature as a key predictor for DO, with varying influences based on site-specific land cover and water quality interactions.

Conclusions:

• Machine learning, particularly the random forest model, provides a robust tool for predicting urban stream DO.
• Integrating land cover data enhances understanding of the complex relationships between urbanization and water quality.
• This approach supports informed decision-making for effective urban watershed management.