Canonical Variable Selection for Ecological Modeling of Fecal Indicators.

Area of Science:

• Environmental microbiology
• Water quality assessment
• Ecological modeling

Background:

• Over 270,000 km of rivers and streams are impaired by fecal pathogens, posing public health and economic risks.
• Traditional fecal indicator organisms lack specificity and fail to identify human health risks or pollution sources.
• Statistical and machine learning models offer potential to understand ecological mechanisms influencing fecal pollution.

Purpose of the Study:

• To employ canonical correlation analysis (CCorA) and machine learning (Maxent) to predict surface water impairment and bacteriophage detections.
• To identify key chemical and microbial parameters influencing fecal pollution.
• To compare the predictive power of different model parameterization strategies.

Main Methods:

• Canonical Correlation Analysis (CCorA) was used for parameter selection.
• Maxent machine learning model was employed for prediction.
• Bootstrapping cross-validation was utilized for model validation.
• Three model suites were developed: initial, CCorA-selected, and optimized parameters.

Main Results:

• CCorA reduced parameters while maintaining initial model accuracy (84.7%); sensitivity analysis improved accuracy to 86.1%.
• Bacteriophage model accuracies were 79.2% (initial), 70.8% (CCorA), and 69.4% (optimized), indicating complex ecological interactions not fully captured by CCorA.
• Distinct ecological drivers were identified: impairment linked to hardness and microbial activity, while bacteriophage detection was inhibited by sediment coliforms.

Conclusions:

• Canonical correlation analysis effectively reduced model complexity for predicting impairment.
• Machine learning models reveal that ecological drivers differ between fecal indicator organisms.
• Both fecal bacteria and bacteriophages are influenced by organic pollution and phosphorus limitation, highlighting nutrient dynamics in water quality.