Evaluating statistical model performance in water quality prediction.

Area of Science:

• Environmental Science
• Public Health
• Statistical Modeling

Background:

• Recreational water activities pose risks of gastrointestinal and respiratory diseases due to contaminated water.
• Rapid water quality fluctuations necessitate accurate and timely predictions to minimize public exposure to pathogens.
• Escherichia coli (E. coli) is a key indicator for freshwater quality, with elevated levels correlating to increased illness risk.

Purpose of the Study:

• To compare the predictive performance of various statistical models for water quality using E. coli levels.
• To identify the most effective model for forecasting E. coli concentrations at a recreational river site.

Main Methods:

• A case study analyzed weekly E. coli data from the Oreti River, Wallacetown, New Zealand (2006-2014).
• Evaluated models included naive, multiple linear regression, dynamic regression, regression trees, Markov chains, classification trees, random forests, multinomial logistic regression, discriminant analysis, and Bayesian networks.
• Model performance was assessed using leave-one-out and k-fold cross-validation.

Main Results:

• The Bayesian network model demonstrated superior performance compared to all other evaluated models.
• The Bayesian network achieved a 21% cross-validation error rate.
• It accurately predicted the majority of E. coli levels classified as unsafe according to New Zealand's Microbiological Water Quality Guidelines.

Conclusions:

• Bayesian networks are a promising tool for predicting recreational water quality due to their accuracy and flexibility.
• Their ability to handle missing data, outliers, and enable real-time updates makes them highly suitable for water quality monitoring.
• Future research will extend this analysis to additional case study sites.