Bayesian network modelling of phosphorus pollution in agricultural catchments with high-resolution data

Abstract

A Bayesian Belief Network was developed to simulate phosphorus (P) loss in an Irish agricultural catchment. Septic tanks and farmyards were included to represent all P sources and assess their effect on model performance. Bayesian priors were defined using daily discharge and turbidity, high-resolution soil P data, expert opinion, and literature. Calibration was done against seven years of daily Total Reactive P concentrations. Model performance was assessed using percentage bias, summary statistics, and visually comparing distributions. Bias was within acceptable ranges, the model predicted mean and median P concentrations within the data error, with simulated distributions more variable than the observations. Considering the risk of exceeding regulatory standards, predictions showed lower P losses than observations, likely due to simulated distributions being left-skewed. We discuss model advantages and limitations, the benefits of explicitly representing uncertainty, and priorities for data collection to fill knowledge gaps present even in a highly monitored catchment.