Incorporating a generalised additive model of river nutrient concentrations into a mechanistic receiving water model

Abstract

eReefs is a large, collaborative project that is building catchment and marine models for Australia's Great Barrier Reef Lagoon (GBRL), a world-heritage environmental asset. The eReefs package includes three-dimensional mechanistic biogeochemical, sediment and hydrodynamic models for the entire GBRL on 4 km and 1 km grid scales, along with a relocatable coastal and estuary model (RECOM) that can be nested within the larger-scale models. Source Catchment models developed by the Government of Queensland for each GBRL catchment will be used to run scenarios to predict the effects of management and land use changes on nitrogen, phosphorus and sediment loads reaching each river. For day-to-day near-real-time and forecast-mode running of the marine models, however, another approach is needed to provide the river loads of sediments, dissolved and particulate loads required as boundary conditions. Generalised Additive Models (GAMs) have been shown (e. g. Kuhnert et al., 2012) to be powerful tools for the prediction of suspended sediment and particulate nutrient loads in tropical rivers. Here, we extend previous work to build GAMs that are able to predict concentrations of suspended sediments, dissolved and particulate nutrients in the Fitzroy River (Queensland) on a daily time-step. In developing the GAMs, we tested a number of routinely and frequently measured meteorological and hydrological variables for potential predictive power. The new terms considered included water temperature (which may alter biogeochemical processing rates), air temperature (a more reliably measured proxy for water temperature), electrical conductivity (which may reflect the influence of particular subcatchment sources), barometric pressure (an indicator of local storm activity), wind stress (which may affect resuspension and mixing in the river and its weirs) and flow from river tributaries (a direct measure of the influence of particular subcatchments). The models generated were tested with regard to the validity of key statistical assumptions, and were then validated against a subset of observational data that had been held back from the original calibration. The strongest models included flow in the Fitzroy River, flow in one or more tributaries, and a discounted flow term that reflected flow in the preceding days and weeks. Models that did not include tributary flow were able to predict concentrations of particulate, but not dissolved materials. Neither meteorological terms nor electrical conductivity proved to be useful predictors, while water temperature was of marginal value. The final GAM provide more accurate predictions on a daily time-step than previously available methods, for both dissolved and particulate materials, and is being used to provide time-series input (e. g. Figure 1) to mechanistic marine models.