A New Approach to Predict Tributary Phosphorus Loads Using Machine Learning- and Physics-Based Modeling Systems.

Area of Science:

• Environmental Science
• Water Quality Modeling
• Machine Learning Applications

Background:

• Tributary phosphorus (P) loads are a primary cause of freshwater lake eutrophication.
• Accurate prediction of P loads is essential for understanding water quality degradation in downstream ecosystems.

Purpose of the Study:

• To develop and evaluate an integrated multimedia modeling system using machine learning (ML) for predicting monthly total phosphorus (TP) and dissolved reactive phosphorus (DRP) loads.
• To assess the performance of ML models trained on meteorological, hydrological, and agricultural management data for P load prediction.

Main Methods:

• Developed two ML models: one for TP loads (10 variables) and one for DRP loads (9 variables).
• Utilized data from the Weather Research and Forecasting (WRF) Model, Variable Infiltration Capacity (VIC) model, and Environmental Policy Integrated Climate (EPIC) model.
• Validated models using data from the Maumee, Sandusky, Portage, and Raisin watersheds discharging into Lake Erie.

Main Results:

• Streamflow was identified as the most significant predictive variable for both TP and DRP loads.
• ML models demonstrated high accuracy in predicting TP and DRP loads both temporally and spatially.
• TP load predictions aligned with or improved upon existing study ranges; DRP load predictions exceeded other studies' performance measures.

Conclusions:

• The integrated multimedia modeling system effectively predicts P loads in freshwater systems.
• The ML-based approach shows potential for improvement with increased data availability.
• This methodology is recommended for studying other freshwater systems and water quality variables.