Predicting in-stream water quality constituents at the watershed scale using machine learning.

Area of Science:

• Environmental Science
• Water Quality Monitoring
• Machine Learning Applications

Background:

• Accurate in-stream water quality prediction is crucial for effective waterbody management and restoration.
• Data-driven modeling, particularly machine learning, offers a powerful approach for water quality assessment.
• Publicly available datasets present an opportunity for scalable water quality prediction frameworks.

Purpose of the Study:

• To evaluate the efficacy of machine learning algorithms in predicting key in-stream water quality constituents using various data inputs.
• To assess the performance of models using only publicly available data versus those incorporating antecedent conditions and additional water quality observations.
• To determine the feasibility of predicting total phosphorus (TP), total nitrogen (TN), total suspended solids (TSS), dissolved oxygen (DO), and fecal coliform bacteria (FC).

Main Methods:

• Five machine learning algorithms were implemented: Support Vector Machines, Random Forest (RF), eXtreme Gradient Boost (XGB), ensemble RF-XGB, and Artificial Neural Network (ANN).
• Four modeling scenarios were tested, varying input data from publicly available datasets to include antecedent conditions and in-stream water quality observations (e.g., pH, turbidity).
• Model performance was evaluated using Nash-Sutcliffe efficiency (NSE) and percent bias (PBIAS) metrics in Bullfrog Creek, Florida.

Main Results:

• Models using only publicly available data satisfactorily predicted TP, TN, DO, and TSS (NSE > 0.75, PBIAS < 10%).
• Fecal coliform bacteria (FC) predictions were unsatisfactory using only public data (NSE < 0.49, PBIAS > 25%), indicating a need for more specific data.
• Antecedent conditions slightly improved predictions and reduced uncertainty, especially when combined with other water quality observations.
• Model performance showed minimal seasonal dependence for all constituents.

Conclusions:

• Publicly available datasets are sufficient for satisfactory prediction of key water quality parameters like nutrients and TSS.
• Predicting fecal coliform bacteria requires more comprehensive data beyond standard publicly available sources.
• The developed machine learning framework is adaptable and can serve as a valuable tool for routine water quality monitoring and prediction.