Validating a Bayesian network model to characterise faecal indicator organism loss from septic tank systems in rural catchments

Validating model predictions with observed data is crucial for fostering confidence in model results, yet it is often overlooked in Bayesian Network (BN) studies. This research validated a BN model designed to predict faecal indicator organism (FIO) loss from septic tank systems (STS) in rural catchments. Both a hybrid model (combining continuous and discrete variables) and a fully discretised model were assessed in two test catchments. Our approach to model validation employed four methods: (1) comparing probability distributions of simulated and observed FIO loads in the hybrid model, (2) sensitivity analysis in the discrete model to identify key variables influencing results, (3) estimating percentage bias (PBIAS) to evaluate the average difference between predicted and observed FIO loads in the hybrid model, and (4) applying Shannon entropy to measure uncertainty in the spatial application of the discrete model. Predicted FIO loads per STS were consistent across models, with the hybrid network estimating 4.63 × 10¹⁰ cfu/yr in the Cessnock catchment and 4.36 × 10¹⁰ cfu/yr in the Mein catchment, while the discrete network predicted 3.85 × 10¹⁰ cfu/yr and 3.65 × 10¹⁰ cfu/yr, respectively, closely aligning with observed values of 6.17 × 10¹⁰ cfu/yr and 5.10 × 10¹⁰ cfu/yr. Sensitivity analysis identified STS condition and treatment level as critical factors influencing FIO loss. Shannon entropy values (1.60-1.85) revealed significant uncertainty in model predictions in the catchment where STS were associated with a variability of Hydrology of Soil Types (HOST)-derived risk factors. When applied at national scale, greater confidence in model results was associated with Central, East and West Scotland where most STS were associated with a moderate to high HOST-derived risk classification. Our research is the first to show how BN models can predict FIO pollution from STS to watercourses and the findings suggest that refining model predictions requires more accurate data on STS treatment levels and maintenance, as well as access to good quality high-resolution water quality monitoring data.