Prediction of sediment, particulate nutrient and dissolved nutrient concentrations in a dry tropical river to provide input to a mechanistic coastal water quality model

https://doi.org/10.1016/j.envsoft.2014.08.009Get rights and content

Highlights

  • A GAM is applied to predict dissolved and particulate nutrient concentrations.

  • A thorough performance evaluation was conducted.

  • Tributary flows were required to successfully predict dissolved nutrients.

  • The Nogoa catchment is an important source of dissolved phosphorus in the Fitzroy.

  • The models will provide input to a mechanistic receiving waters model.

Abstract

A Generalised Additive Modelling (GAM) approach is applied to prediction of both particulate and dissolved nutrient concentrations in a wet-tropical river (the Fitzroy River, Queensland, Australia). In addition to covariant terms considered in previous work (i.e. flow, discounted flow and a rising-falling limb term), we considered several new potential covariates: meteorological and hydrological variables that are routinely monitored, available in near-real time, and were considered to have potential predictive power. Of the additional terms considered, only flows from three tributaries of the Fitzroy River (namely, the Nogoa, Comet and Isaac Rivers) were found to significantly improve the model. Inclusion of one or more of these additional flow terms greatly improved results for dissolved nitrogen and dissolved phosphorus concentrations, which were not otherwise amenable to prediction. In particular, the Nogoa sub-catchment, dominated by pasture for cattle, was found to be important in determining dissolved inorganic nitrogen and phosphorus concentrations reaching the river mouth. This insight may direct further research, including future refinement of processed-based catchment models. The GAMs described here are used to provide near real-time river boundary conditions for a complex coupled hydrodynamic and biogeochemical model of the Great Barrier Reef Lagoon, and can be coupled with a forecasting hydrological model to allow integrated forecasting simulations of the catchment to coast system.

Graphical abstract

Despite the absence of any obvious direct relationship between flow and dissolved nutrient concentrations (in this case dissolved organic nitrogen, DON) concentrations in the Fitzroy River (left), a Generalised Additive Model is able to predict dissolved nutrient concentrations as a function of flow in the Fitzroy River and one of its tributaries (right).

  1. Download : Download high-res image (142KB)
  2. Download : Download full-size image

Introduction

Sediment and nutrient loads from rivers to estuaries and coastal waters often increase when land use changes, especially as agriculture intensifies and urban development progresses. This is often a concern for environmental managers, as increased sediment and nutrient loads adversely affect water quality and change the trophic status of freshwater and marine systems (Smith, 2003).

In the case of the Great Barrier Reef Lagoon, an environmental asset of internationally recognised importance, increased nutrient loads from catchments over the past 200 years are believed to be the driving force behind the spread of Crown of Thorns Starfish, an invasive species which preys on coral and has been responsible for massive damage to reef ecosystems in recent decades (Brodie et al., 2005, De'ath et al., 2012, Fabricius, 2005). Increased nutrient loads also drive increases in pelagic primary production, which can have a complex range of ecological impacts.

A large project (eReefs) is currently underway to implement a three-dimensional coupled hydrodynamic, sediment dynamic and biogeochemical model for the entire Great Barrier Reef Lagoon (GBRL) (Chen et al., 2011). The marine models are implementations of SHOC (Cetina-Heredia and Connolly, 2011) and EMS (Margvelashvili et al., 2013, Robson et al., 2008, Wild-Allen, 2013). These models operate on small time-steps (typically around 2 min) and require as input daily or sub-daily estimates of concentrations of sediments and nutrients (dissolved and particulate organic and inorganic nitrogen and phosphorus) in each of 22 rivers flowing into the system.

Semi-distributed models that predict average annual loads of sediments, nitrogen and phosphorus in GBR catchments have been implemented in Source Catchments and its predecessors (Armour et al., 2009, Dougall et al., 2006, McCloskey et al., 2011), but have so far not demonstrated the ability to accurately simulate day-to-day variations in the concentrations of nutrients. While more complex, process-based catchment models such as SWAT may (Lai et al., 2006, Richards et al., 2003) or may not (Chahinian et al., 2011, Chu et al., 2004) offer the potential to simulate such nutrients on daily timescales, these detailed process models have very large data requirements and heavy implementation costs.

Sediment rating curves, which estimate sediment concentrations using a simple linear regression of sediment load as a function of flow, often produce adequate estimates of loads on yearly or decadal timescales, but produce very large errors when applied to daily concentrations (Horowitz, 2003). More sophisticated statistical techniques such as maximum likelihood estimation (MLE) and adjusted maximum likelihood estimate (AMLE) applied to parameterisation of a linear regression model have been used successfully in a range of applications (Cohn, 2005). AMLE is particularly appropriate for use with datasets that include “below detection limit” results. The LOADEST package, available online from the USGS, provides a user-friendly tool to implement these methods, optionally incorporating sin and cosine terms in the linear regression to account for seasonal patterns (Aulenbach and Hooper, 2006), and has been widely used, particularly within the USA.

Vecchia and Ballerini (1991) and more recent work involving the same author (e.g. Johnson et al., 2009, Milly et al., 2005) have applied autoregressive time series analyses to detect and quantify temporal trends in water quality data, adjusted for expected concentrations given flow. Hirsch et al. (2010) built on this theme, presenting the WRTDS (Weighted Regression on Time, Discharge and Season) model, a linear modelling approach that pays particular attention to changes and trends over time.

More recently, Generalised Additive Models (GAMs) such as LRE (Loads Regression Estimator) have been developed and implemented for several rivers (Kuhnert et al., 2012, Wang et al., 2011). GAMs are generalised regression models which incorporate smoothing functions, s() of covariates. These functions do not have a pre-defined form, and need not be linear, but attempt to capture the main features of the data, using, for instance, a penalised regression spline function that fits a flexible smoothing term to the data. This makes the approach very flexible as it is capable of combining multiple, nonlinear functional responses. GAMs have been demonstrated as a powerful tool for prediction of sediment loads, requiring much less input data and lower computational costs than process-based models. LRE has been extended to estimate total nutrient as well as sediment loads from Great Barrier Reef catchments (Kroon et al., 2012) to provide a firm comparative basis for management.

The LRE model is simply defined as:log(C)j=β+i=12αkXkj+i=12sk(Zkj)+ɛj.(Kuhnert et al., 2012) where C is the concentration of a constituent at time i, Xkj and Zkj are covariates measured at that time, s() represents a smoothing term (described above), and ɛj is a normally distributed error term. x1i to x3i represent linear and quadratic flow terms and a categorical term to indicate whether flow is rising or falling. Smoothed terms (zkj) include discounted flow terms (discussed below). Additional covariates, which may or may not be log terms, are included on a case-by-case basis.

Although LRE achieves good agreement with observed sediment, total nitrogen and total phosphorus concentrations in both the Burdekin River (Kuhnert et al., 2012) and Fitzroy River (our analysis), when the method is applied to prediction of dissolved organic or inorganic nutrients on daily time-steps, LRE does not achieve satisfactory results if driven by flow and discounted flow alone.

In this paper, we build on the basic LRE GAM, applying the approach with a range of additional covariates to the Fitzroy River, one of the largest rivers flowing into the GBRL and demonstrate models that provides good agreement with observational measurements of dissolved inorganic and organic nitrogen and phosphorus as well as sediment and particulate nutrient concentrations in the Fitzroy River.

Section snippets

Methods

Regular sediment and nutrient monitoring has been conducted in the Fitzroy River since 1999. Samples are taken at ‘the Gap’, just upstream of the barrage at Rockhampton, to avoid the complicating tidal influences downstream.

The dataset used to develop our model includes 102 Total Suspended Solids (TSS) records and 67 nutrient records, from samples taken under varying flow conditions between 2003 and 2008. An event sampling strategy was followed, so most measurements relate to flow events.

Results

All models reported in Table 2, Table 3, Table 4, Table 5 were tested for acceptable agreement with the assumptions of normality, homoscedasticity and linearity (e.g. Fig. 2) and include only significant (p < 0.05) covariate terms. GAMs can only be considered valid if the residual errors are normally distributed (Fig. 2, top left) and show no trends when plotted against the order of observations (Fig. 2, top right) or against covariate terms (Fig. 2, bottom left and right). Performance metrics

Discussion

This study provides further evidence that GAMs can provide greatly improved predictions of concentrations of particulate materials (total suspended solids, particulate nitrogen and particulate phosphorus) at a daily time-step in comparison with simple linear regressions that rely only on flow as a predictive variable. These approaches have been previously demonstrated for Great Barrier Reef rivers, including the Burdekin River, by Wang et al. (2011), Kuhnert et al. (2012) and Kroon et al. (2012)

Conclusions

GAMs presented for similar rivers by previous authors (e.g. the LRE model presented by Kuhnert et al., 2012) have considered the effects of flow, discounted flow (which accounts for recent past flow) and a rising-falling limb term (which indicates whether flow is currently increasing or decreasing). We have extended this approach to consider several additional potential forcing terms using data available from routine monitoring in the Fitzroy River and a nearby meteorological station. The new

Acknowledgements

This work was conducted through eReefs, a collaboration between the CSIRO Oceans and Atmosphere Flagship, the Australian Bureau of Meteorology and the Government of Queensland. eReefs is supported by the Science Industry Endowment Fund (SIEF), the Great Barrier Reef Foundation, the BHP Billiton Mitsubishi Alliance and in-kind contributions from the project partners. The second author (Mr Vincent Dourdet) contributed to this work through a trainee placement at CSIRO as part of his degree at

References (47)

  • I.M. Sobol et al.

    Estimating the approximation error when fixing unessential factors in global sensitivity analysis

    Reliab. Eng. Syst. Saf.

    (2007)
  • R.B. Thomas et al.

    An evaluation of flow-stratified sampling for estimating suspended sediment loads

    J. Hydrol.

    (1995)
  • Y.G. Wang et al.

    Load estimation with uncertainties from opportunistic sampling data – a semiparametric approach

    J. Hydrol.

    (2011)
  • Y.G. Wang et al.

    Sediment concentration prediction and statistical evaluation for annual load estimation

    J. Hydrol.

    (2013)
  • K. Wild-Allen et al.

    Mechanisms driving estuarine water quality: a 3D biogeochemical model for informed management

    Est. Coast. Shelf Sci.

    (2013)
  • F.E. Anderson

    Resuspension of estuarine sediments by small amplitude waves

    J. Sediment. Res.

    (1972)
  • J.D. Armour et al.

    Catchment modelling of sediment, nitrogen and phosphorus nutrient loads with SedNet/ANNEX in the Tully-Murray basin

    Mar. Freshw. Res.

    (2009)
  • S. Arrhenius

    Quantitative Laws in Biological Chemistry

    (1915)
  • B.T. Aulenbach et al.

    The composite method: an improved method for stream-water solute load estimation

    Hydrol. Process.

    (2006)
  • U. Bende-Michl et al.

    A systematic approach to choosing an automated nutrient analyser for river monitoring

    J. Environ. Monit.

    (2010)
  • U. Bende-Michl et al.

    High-frequency nutrient monitoring to infer seasonal patterns in catchment source availability, mobilisation and delivery

    Environ. Monit. Assess.

    (2013)
  • P. Cetina-Heredia et al.

    A simple approximation for larval retention around reefs

    Coral Reefs

    (2011)
  • N. Chahinian et al.

    Flow and nutrient transport in intermittent rivers: a modelling case-study on the Vene River using SWAT 2005

    Hydrol. Sci. J.-J. Des. Sci. Hydrol.

    (2011)
  • Cited by (21)

    • Novel predictors related to hysteresis and baseflow improve predictions of watershed nutrient loads: An example from Ontario's lower Great Lakes basin

      2022, Science of the Total Environment
      Citation Excerpt :

      Many studies that have used GAMs have focused on ecology (Riemann et al., 2016; Yang et al., 2020), lake water quality (Liu et al., 2019; Yang and Moyer, 2020; Zhang and Zhi, 2020), coastal water quality (Haraguchi et al., 2015; Harding et al., 2016; Murphy et al., 2013; Richards et al., 2014, 2013; Riemann et al., 2016), air quality (He et al., 2006), hydrology (Chebana et al., 2014), and physical water parameters (Laanaya et al., 2017; Yang and Moyer, 2020). Applications using GAMs in freshwater nutrient analyses are sparse (Hwang et al., 2016; Murphy et al., 2013) and nutrient load estimations (e.g., LRE: loads regression estimator tool) have been limited to larger tropical rivers using low-frequency nutrient concentration data (Kuhnert et al., 2012; Robson and Dourdet, 2015; Wang et al., 2011). Moreover, GAMs have not been used to predict nutrient loads from watersheds in the Great Lakes basin to the best of our knowledge.

    • Space-time modelling of groundwater level and salinity

      2021, Science of the Total Environment
    • Development of catchment water quality models within a realtime status and forecast system for the Great Barrier Reef

      2020, Environmental Modelling and Software
      Citation Excerpt :

      Water and solutes are routed through the channel before transporting through the river system. These models have not been able to accurately simulate day-to day variations in the concentration of nutrients (Robson and Dourdet, 2015). Statistical models have been widely used to model water quality (Cohn, 1995; Cohn et al., 1992; Colby, 1956; Miller, 1951) and have been used in Australia (Joo et al., 2014; Kroon et al., 2012; Kuhnert et al., 2013, 2012; Robson and Dourdet, 2015; Rustomji and Wilkinson, 2008; Wang et al., 2011).

    View all citing articles on Scopus
    1

    Present address: Ecole des Mines d'Ales, France.

    View full text