Transferability of a Bayesian Belief Network across diverse agricultural catchments using high-frequency hydrochemistry and land management data

Biogeochemical catchment models are often developed for a single catchment and, as a result, often generalize poorly beyond this. Evaluating their transferability is an important step in improving their predictive power and application range. We assess the transferability of a recently developed Bayesian Belief Network (BBN) that simulated monthly stream phosphorus (P) concentrations in a poorly-drained grassland catchment through application to three further catchments with different hydrological regimes and agricultural land uses. In all catchments, flow and turbidity were measured sub-hourly from 2009 to 2016 and supplemented with 400-500 soil P test measurements. In addition to a previously parameterized BBN, five further model structures were implemented to incorporate in a stepwise way: in-stream P removal using expert elicitation, additional groundwater P stores and delivery, and the presence or absence of septic tank treatment, and, in one case, Sewage Treatment Works. Model performance was tested through comparison of predicted and observed total reactive P (TRP) concentrations and percentage bias (PBIAS). The original BBN accurately simulated the absolute values of observed flow and TRP concentrations in the poorly and moderately drained catchments (albeit with poor apparent percentage bias scores; 76 % ≤ PBIAS≤94 %) irrespective of the dominant land use, but performed less well in the groundwater-dominated catchments. However, including groundwater total dissolved P (TDP) and Sewage Treatment Works (STWs) inputs, and in-stream P uptake improved model performance (-5 % ≤ PBIAS≤18 %). A sensitivity analysis identified redundant variables further helping to streamline the model applications. An enhanced BBN model capable for wider application and generalisation resulted.