A BAYESIAN HIERARCHICAL SPATIO-TEMPORAL APPROACH TO MODEL PHOSPHORUS LOADING IN EIGHT OHIO WATERSHEDS

Abstract

Accurately estimating riverine nutrient loads remains an imperative step towards mitigating and managing eutrophication-related impairments. However, load estimation is often affected by the infrequent monitoring of riverine nutrient concentrations. In this work, two novel Spatio-temporal Bayesian Hierarchical models (BHM), the ratio estimator approach and the regression approach, are proposed, developed and validated. The models are assessed with regards to their performance in estimating total phosphorus loads between 2005 and 2020 across eight intensively monitored watersheds in Ohio, USA. The two proposed BHMs incorporate watershed-level predictors that allow for the partial-pooling of data collected from multiple watersheds over time. The proposed models include two modelling hierarchies, namely one spatial and another temporal. The spatial hierarchy attempts to model inter-station variability in loading by accounting for watershed-level differences in land cover/land use (LULC), while the temporal hierarchy attempts to capture the inter-annual variabilities in loads due to changes in flows. To assess the skill of the two proposed models, the model estimates are compared to the "true loads" and to loads estimated using typical load estimation methods. Moreover, the performance of the models is tested as a function of two different sampling frequencies. Overall, the results showed that the BHM ratio estimator approach outperformed all other load estimation methods. Moreover, its results proved to be invariant with the changes to the sampling frequency which will allow for more efficient monitoring of watersheds, as it can predict loads with a high degree of accuracy using as few as one sample every 10 days. This significantly reduces the cost and labor associated with monitoring, as well as the potential for errors or inaccuracies. Finally, the results highlighted that accounting for the spatial variabilities in LULC that existed between watersheds within the model hierarchical was significantly more important than accounting for the temporal differences in flow.