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

Reliable prediction of Faecal Indicator Organisms (FIOs) in coastal bathing waters remains a major challenge because FIO concentrations could vary rapidly with the flow, meteorological, and hydrological conditions. Mechanistic models require high computational resources, data and could be time consuming to set up. Machine learning (ML) models have been used as alternative solutions particularly for early warning systems and digital twins. However, conventional ML models often underperform when exposed to correlated environmental variables. In this study, we propose a hybrid, dimension-reduced ensemble framework that combines Principal Component Analysis (PCA) with stacked ML and knowledge-distillation-inspired teacher-student model to deliver accurate, interpretable forecasts from a reduced number of inputs. The hybrid approach was applied to the Swansea Bay, UK, where 23 different hydro-meteorological and E. coli and Enterococci data were collected during one bathing season at high-frequency, namely 30-min. The results showed that the first two principal components (PCs), highlighting the warm/sunny or cool/humid weather types (PC1) and the wind-driven or calmer and level of rain (PC2), were the two dominant axes that characterise the Bay. The PCs were fed into a stacked "teacher" ensemble (KNN-RF-XGBoost) model whose predictive signal was partially transferred to a lightweight neural network "student" via blended supervision inspired by knowledge distillation. The framework achieved R² = 0.72, RMSE = 0.11 (∼259 cfu/100 ml) for E. coli, and R² = 0.64, RMSE = 0.14 for Enterococci (∼334 cfu/100 ml), outperforming standalone ML or DL models by up to 30 percent. Beyond accuracy, the PCA layer reduced input by 65 percent, cutting requirements for sensors for early warning systems and digital twins, computational cost and revealing physically meaningful patterns driving contamination dynamics. This PCA-ensemble-teacher-student hybrid modelling provides an interpretable, scalable tool for real-time water quality forecasting in dynamic coastal environments.