Testing Bayesian Network transferability to diverse agricultural catchments with high phosphorus saturation

A Bayesian Network (BN) aimed at calculating stream phosphorus (P) concentrations in agricultural catchments was previously parametrized with high-frequency data in a pilot study. To test model transferability, the BN was applied to three further agriculture-dominated catchments in Ireland with varying land use, hydrology, and P pressures, all monitored through the Agricultural Catchments Programme (ACP) of the Irish Agriculture and Food Development Authority. While the pilot catchment Ballycanew was dominated by poorly drained grassland, the further three catchments were dominated by well-drained grassland (Timoleague), well-drained arable (Castledockrell), and moderately-drained arable (Dunleer), respectively. In all four catchments, the main P source came from agriculture and (minimal) domestic inputs, whilst the well-drained arable catchment also contained Sewage Treatment Works (STWs).

To best fit the characteristics of the catchments, a total of six different BN structures were developed. The models were parametrized using a range of methods, including bootstrapping of high-frequency data to obtain fitted distributions, distribution fitting of literature data, and expert elicitation to quantify in-stream P uptake processes. Model transferability and fit were evaluated using a suit of approaches, including 1) calculating percentage bias between simulated and observed distributions fitted to the observed stream Total Reactive P (TRP) concentration, 2) comparing modelled concentration quantiles and means to the observed, and 3) visually comparing the posterior distributions by plotting them against daily observations.

The original BN structure developed in the pilot study was found to best fit the poorly and moderately drained catchments, irrespective of the dominant land use (78% ≤ PBIAS ≤ 81%), not as well in the groundwater-dominated catchments. This confirms that the initial BN represents the catchment-specific process understanding whereby transport via quick-flow dominates P processes in these catchments. In contrast, the well-drained catchments required more complex BN structures to perform well. The additional processes included groundwater Total Dissolved P (TDP) loads, derived from observed concentrations from piezometer data, STWs loads, and in-stream P uptake calculations. These more complex model implementations yielded good results in Castledockrell and Timoleague (-5% ≤ PBIAS ≤ 14%). In all four catchments, the additional in-stream P removal process improved the model performance, however, it remains a second-order mechanism.

Overall, the unique monitoring programme allowed pilot-testing BN transferability, a research avenue that needs to be further explored across catchment typologies and scales.