Improved hydro-epidemiological prediction of faecal indicator organisms using knowledge distillation inspired machine...

Area of Science:

• Environmental Science
• Water Quality Monitoring
• Machine Learning Applications

Background:

• Predicting Faecal Indicator Organisms (FIOs) in coastal waters is challenging due to rapid environmental fluctuations.
• Mechanistic models are data-intensive and computationally expensive.
• Machine learning (ML) offers alternatives for early warning systems but struggles with correlated environmental data.

Purpose of the Study:

• To develop a hybrid, dimension-reduced ensemble framework for accurate and interpretable FIO forecasting.
• To combine Principal Component Analysis (PCA) with stacked ML and a teacher-student model.
• To reduce input data requirements and computational costs for real-time water quality prediction.

Main Methods:

• Applied a hybrid framework combining PCA with stacked ML (KNN-RF-XGBoost) and a knowledge-distillation-inspired teacher-student model.
• Utilized high-frequency hydro-meteorological and microbial data from Swansea Bay, UK.
• Reduced input dimensionality using PCA, identifying key environmental drivers (weather types, wind, rain).

Main Results:

• The hybrid model achieved high accuracy for E. coli (R²=0.72) and Enterococci (R²=0.64).
• Performance surpassed standalone ML/DL models by up to 30%.
• PCA reduced input data by 65%, enhancing interpretability and reducing sensor/computational demands.

Conclusions:

• The PCA-ensemble-teacher-student hybrid model offers an accurate, interpretable, and scalable solution for real-time coastal water quality forecasting.
• This approach is valuable for developing effective early warning systems and digital twins.
• The framework reveals physically meaningful patterns driving coastal contamination dynamics.