Abstract/Summary

Bayesian Belief Networks (BBNs) are a promising but underutilized probabilistic graphical tool for modelling water quality for Environmental Impact Assessment, with their ability to include uncertainty in the predictions being relevant to catchment and water managers. This thesis explores the application of these tools to predict phosphorus (P) losses in terms of total reactive phosphorus (TRP) concentrations in four Irish agricultural catchments, with high P concentrations being a major concern in at least three of them. A hybrid Bayesian Belief Network combining discrete and continuous variables was developed for a surface hydrology-dominated grassland catchment, using daily concentration data to build the BBN priors and assess model performance. The step-wise introduction of different P sources, combined with high-frequency data and detailed catchment understanding improved the first model iteration’s predictive ability. In all model applications, the models’ predictions presented wider distributions than the observations, which was noted in similar work, and remains a property of BBNs. Transferring the BBN across catchments allowed testing the model’s structural uncertainty and showed that the developed BBN could perform well in surface-driven catchments. The BBN was enhanced by improved process representation and catchment-specific parameterization. Model transferability across catchment typologies (surface vs groundwater-dominated, grassland vs tillage land use) is explored, and the BBNs are used to predict future P concentrations under climate change scenarios. The application of the catchment-specific BBNs to predict future P concentrations under climate change revealed the need for further BBN sensitivity analysis to aid result interpretation. The potential for BBNs to be used as a tool to inform compliance with regulatory standards is discussed. The discussion considers learnings from current BBN research, P processes represented in both BBNs and process-based models, and the model application in this study. Limitations of the approach and future research avenues are explored.