Liu 2021

Environmental Pollution 288 (2021) 117337
Contents lists available at ScienceDirect
Environmental Pollution
journal homepage: www.elsevier.com/locate/envpol
A multi-model approach to assessing the impacts of catchment

characteristics on spatial water quality in the Great Barrier
Reef catchments☆
Shuci Liu a, *, Dongryeol Ryu a, J. Angus Webb a, Anna Lintern a, b, Danlu Guo a, David Waters c,
Andrew W. Western a
a
Department of Infrastructure Engineering, The University of Melbourne, Parkville, VIC, 3010, Australia
b
Department of Civil Engineering, Monash University, VIC, 3800, Australia
c
Queensland Department of Resources, Toowoomba, QLD, 4350, Australia
A R T I C L E I N F O A B S T R A C T
Keywords: Water quality monitoring programs often collect large amounts of data with limited attention given to the
Water quality assessment of the dominant drivers of spatial and temporal water quality variations at the catchment scale. This
Catchment characteristics study uses a multi-model approach: a) to identify the influential catchment characteristics affecting spatial
Statistical model
variability in water quality; and b) to predict spatial variability in water quality more reliably and robustly.
Multi-model inference
Tropical catchments in the Great Barrier Reef (GBR) area, Australia, were used as a case study. We developed
Model averaging
statistical models using 58 catchment characteristics to predict the spatial variability in water quality in 32 GBR
catchments. An exhaustive search method coupled with multi-model inference approaches were used to identify
important catchment characteristics and predict the spatial variation in water quality across catchments. Boot
strapping and cross-validation approaches were used to assess the uncertainty in identified important factors and
robustness of multi-model structure, respectively. The results indicate that water quality variables were generally
most influenced by the natural characteristics of catchments (e.g., soil type and annual rainfall), while anthro
pogenic characteristics (i.e., land use) also showed significant influence on dissolved nutrient species (e.g., NOX,
NH4 and FRP). The multi-model structures developed in this work were able to predict average event-mean
concentration well, with Nash-Sutcliffe coefficient ranging from 0.68 to 0.96. This work provides data-driven
evidence for catchment managers, which can help them develop effective water quality management strategies.
1. Introduction Schaffelke et al., 2012; Waterhouse et al., 2017).

The effectiveness of improved management practices heavily de
Fresh water resources are key to agricultural, industrial, and envi pends on a sound understanding of pollutant processes (i.e., source,
ronmental activities (Kundzewicz et al., 2007). However, there is a mobilization and delivery) in catchments (Granger et al., 2010).
worldwide concern that water quality in rivers and streams is deterio Riverine water quality is highly variable across space and time and is
rating (Booth et al., 2016; Hiscock et al., 2002; Zhao et al., 2019). affected by a wide range of natural and anthropogenic factors in
Elevated levels of pollutants in streams can lead to substantial economic catchments (Chang, 2008; Zhang et al., 2016). Therefore, it is important
and environmental losses, particularly in coastal and estuarine ecosys to understand the key factors affecting spatial and temporal variability
tems (De Valck et al., 2018; Pickering et al., 1987). To address riverine of riverine water quality.
water quality degradation, improved management in many coastal re Generally, spatial and temporal variation in water quality is driven
gions has been implemented in recent decades, e.g., Chesapeake Bay in by three key processes in catchments: 1) sources – the amount of pol
the US (NOE et al., 2020; Preston et al., 1999; Zhang et al., 2018) and the lutants available within a catchment; 2) mobilization – detachment of
Great Barrier Reef catchments in Australia (Brodie et al., 2017; pollutants from the source by processes such as erosion and weathering;
☆
This paper has been recommended for acceptance by Sarah Harmon.
* Corresponding author.
E-mail address: shucil@student.unimelb.edu.au (S. Liu).
https://doi.org/10.1016/j.envpol.2021.117337
Received 14 December 2020; Received in revised form 3 April 2021; Accepted 6 May 2021
Available online 10 May 2021
0269-7491/© 2021 Elsevier Ltd. All rights reserved.
S. Liu et al. Environmental Pollution 288 (2021) 117337
and 3) delivery – the transport of the detached pollutants to the important factors affecting riverine water quality and to predict water
receiving waters via surface or subsurface flow (Granger et al., 2010). quality response across space. Specifically, the objectives of this study
Within a catchment, water quality exhibits substantial temporal vari include: (1) identifying the influential catchment characteristics
ability, including at daily (Brainwood et al., 2004; Meybeck et al., 2012), affecting spatial variability in different water quality constituents; and
seasonal (Ouyang et al., 2006; Xiaolong et al., 2010; Xu et al., 2019) and (2) developing a robust statistical modeling approach to predict the
inter-annual (Fabricius et al., 2013; Zhuo et al., 2016) scales. Similarly, average water quality responses, using key catchment characteristics.
riverine water quality can vary markedly between catchments. Natural We focused on the Great Barrier Reef region (Queensland, Australia),
and anthropogenic characteristics of a catchment can influence the three due to: (1) its high ecological and economic values (De Valck et al.,
key catchment processes, and thus lead to large spatial variation in 2018) that are threatened by water quality deterioration from inland
water quality. The relationship between water quality and anthropo catchments (Brodie et al., 2013b; Waterhouse et al., 2017); and (2)
genic factors (e.g., land use) has been extensively studied and identified deficient understanding of water quality spatial variation in tropical and
as one of the key controlling factors that affect spatial variation in water subtropical zones (Piazza et al., 2018). We used a long-term event-based
quality (Bramley et al., 2002; Jiang et al., 2015; Lintern et al., 2018a; water quality monitoring dataset of nine constituents, including: total
Nash et al., 2011). For instance, land clearing and any associated suspended solids (TSS), particulate nitrogen (PN), oxidized nitrogen
intensification of agricultural activities post clearing can result in an (NOX), ammonium nitrogen (NH4), dissolved organic nitrogen (DON),
increase in nutrient loads from fertilizer application, as well as sus filterable reactive phosphorus (FRP), dissolved organic phosphorus
pended sediment caused by altering surface soil properties (e.g., tillage) (DOP), particulate phosphorus (PP) and electrical conductivity (EC).
and sediment budgets (Blevins et al., 2018; Smith et al., 2013). In Monitoring data for all nine constituents were collected from the 32 GBR
addition, the natural conditions of catchments (climate, hydrology, catchments. Fifty-eight catchment-scale natural and anthropogenic
vegetation cover, geology and topography) have a potential impact on characteristics were investigated to assess their relative effects on water
the spatial variation in water quality (Donohue et al., 2006; Ye et al., quality spatial variability. These characteristics came from six cate
2009). For instance, catchment geology and soil type determine the gories: catchment topography, land cover, land use, geology, climate
source of sediment and naturally-derived nutrients in catchments and hydrology.
(Grayson et al., 1997).
In this study, we focus on the spatial variability in long-term average 2. Materials and methods
water quality and how these spatial patterns vary with catchment
characteristics, acknowledging that temporal dynamics are also impor 2.1. Study area
tant (Guo et al., 2019). Previous studies have highlighted a range of
modeling techniques that can be used to explore the relationship be The Great Barrier Reef (GBR) is an iconic Australian coral reef
tween catchment characteristics and water quality spatial responses (Fu ecosystem, with substantial environmental and economic value (De
et al., 2019; Soranno et al., 1996). However, these studies have certain Valck et al., 2018; Whitten et al., 2004). However, it has experienced a
limitations. Firstly, these studies have mainly focused on a small number drastic decline in coral cover – 50% for the entire GBR – since 1985
of catchment characteristics, mostly hydroclimatic and land uses char (Brodie et al., 2013a; Kroon et al., 2016). This deterioration is thought to
acteristics (Afed Ullah et al., 2018; Young et al., 1996), and the relative be driven in part by poor riverine water quality discharging from the
importance of catchment natural and anthropogenic characteristics is adjacent catchments (Waterhouse et al., 2017).
rarely evaluated. More importantly, past investigations have often The GBR catchments (an approximate total area of 432,000 km2,
identified a single ‘best model’ using forward or backward stepwise Fig. 1) consist of six Natural Resource Management (NRM) regions that
variable selection to interpret complex processes (Juahir et al., 2011; discharge into the GBR lagoon. These NRM regions can be further
Sangani et al., 2015). However, because multiple controlling factors can divided into 35 major river basins (Pratchett et al., 2019). These
result in a number of plausible models with comparable predictive catchments are the most climatically, geologically and topographically
power (Whittingham et al., 2006), applying a single-best model pro diverse natural landscapes on the Australian continent (Gilbert et al.,
vides: 1) limited understanding of key drivers affecting spatial vari 2001). Rainfall in GBR catchments occurs predominantly in the summer
ability in water quality; and 2) limited capacity for predicting water wet season (November to April), with an uneven spatial distribution of
quality across space. annual rainfall ranging from less than 500 mm to more than 8000 mm
Multi-model inference overcomes the limitations of the single best (Figure S1-a) (Davis et al., 2017; Kuhnert et al., 2009). As a result, runoff
model approach. This approach considers evidence from multiple of most rivers in the GBR catchments exhibits high spatial and seasonal
plausible models by linearly combining these competing models’ out variations, with a few large events in the wet season contributing to the
puts into new averaged model predictions (Burnham et al., 2004; Cade, majority of annual runoff and low or no flow dominating during the dry
2015; Parrish et al., 2012). Compared to the single model inference season. Differences in geology (e.g., lithology, Figure S1-c) and bioge
approach, the multi-model inference approach provides more robust ography (e.g., bioregion, Figure S1-d) across space also lead to the sig
predictions (Burnham et al., 2002; Poeter et al., 2005; Saft et al., 2016). nificant heterogeneity in the natural landscape characteristics in the
The multi-model approach has gained increasing popularity in the water GBR catchments. In this study, we used water quality monitoring data
resources community in recent years, including ensemble hydrologic collected from 32 monitoring sites within the GBR catchments (Fig. 1,
forecasting (Duan et al., 2007; Raftery et al., 2005), groundwater hy detailed information of these catchments is in Table S1, in Supplemen
drology (Chen et al., 2006; Foglia et al., 2013), catchment functioning tary Material). The selection of these catchments was based on: 1) sites
(Beck et al., 2013; Saft et al., 2016) and climate change impact assess that have continuous discharge monitoring; and 2) sites that have a
ment based on global climate model outputs (Deb et al., 2018; Stoll water quality monitoring record of greater than 5 years, for the selected
et al., 2011). Surprisingly, this approach has been rarely used to un water quality variables (see Section 2.2.1), and with a good coverage of
derstand and predict riverine water quality responses. The only previous samples under both high and low flows. Detailed land use in the areas
application of multi-model inference in evaluation of stream water upstream of the sampling sites is included in Table S2 in Supplementary
quality is Lintern et al. (2018b), where multi-model inference was used Material.
to investigate key factors affecting water quality in temperate catch
ments. However, in that study, the predictions were still derived from a 2.2. Data collection and preparation
single best model structure.
To evaluate the utility of multi-model prediction for spatial vari 2.2.1. Water quality data collection
ability in water quality, this study applies the approach to identify the The water quality samples used in this study were obtained from the
2
Fig. 1. The Great Barrier Reef and GBR catchments,

monitoring sites, land uses and the six NRM regions
(Data sources - Great Barrier Reef Marine Park Au
thority (2004); Queensland Government (2017)). Land
uses have the following characteristics: (1) Conserva
tion: forest, woodland, savannah, etc for conservation
purpose; (2) Dryland: rainfed agriculture but excluding
grazing and sugarcane; (3) Grazing: grazing native
vegetation; (4) Intensive: urban areas, roads, etc. (5)
Irrigated: irrigated cropping excluding sugar cane; (6)
Sugar: rain-fed and irrigated sugar cane, and (7) Water:
water bodies including lake, river, and marsh.
Great Barrier Reef catchment loads monitoring program, established that other constituents of interest were not included because they are not
and maintained by the Queensland Department of Natural Resources, monitored (organic carbon) or there was insufficient data for these an
Mines and Energy and Department of Environment and Science alyses (pesticides).
(Australian and Queensland governments, 2020; Bartley et al., 2017), Unlike many conventional operational water quality sampling pro
who collect and analyze all samples according to American Public grams, event-based samples were dominant in the water quality moni
Health Association protocols (see details in Table S3) (APHA, 2005). The toring data. This dataset contained both the intensive samples (e.g.,
nine constituents studied covered a range of key indicators of stream daily or every few hours by automatic samplers) that were taken during
sediments, nutrients and salinity, including: total suspended solids runoff events, as well as grab samples (e.g., monthly) which were taken
(TSS), particulate nitrogen (PN), oxidized nitrogen (NOX), ammonium under baseflow conditions. The Great Barrier Reef catchment loads
nitrogen (NH4), dissolved organic nitrogen (DON), filterable reactive monitoring program (more details in Huggins et al. (2018)) was
phosphorus (FRP), dissolved organic phosphorus (DOP), particulate designed to capture the pollutant export during both high and low flow
phosphorus (PP) and electrical conductivity (EC), from 2006 to 2016. conditions, and was part of the Paddock to Reef Integrated Monitoring,
These nine water quality variables were selected because they pose a Modeling and Reporting Program (Paddock to Reef program) (Shaw
significant threat to the coral reef ecosystem (McCloskey et al., 2017), et al., 2014). In that program, the monitoring results were used to
and there was sufficient data available to support our analyses. Analyses calibrate a conceptual catchment water quality model – Source Catch
of these constituents will provide a useful comprehensive picture on the ments, which enabled estimation of the annual pollutant loads to the
overall water quality status and its key spatial drivers. It is worth noting reef lagoon (Orr et al., 2014; Waters et al., 2007, 2013).
3
S. Liu et al.
Table 1
Summary of 58 catchment characteristics and their abbreviations used in this paper. Among these categories, only land use characteristics are considered as anthropogenic influences. Detailed description, source and
summary statistics of these characteristics can be found in Tables S7 and S8.
Catchment characteristic Abbreviation Catchment characteristic Abbreviation
Topography maximum catchment elevation (m) Max_Elevation Land use conservation use (%) PerConservation
mean catchment elevation (m) Mean_Elevation dryland farming (%) PerDrylandAgri
catchment area (km2) Area irrigated farming (%) PerIrrigated
stream density (km/km2) StreamDensity water (%) PerWater
catchment relief ratio CatRelifRatio grazing (%) PerGrazing
valley bottoms (%) Valley_Bottoms sugar cane (%) PerSugar
mean catchment slope (%) MeanCatSlope intensive use (%) PerIntensiveUses
total catchment length (km) CatLength cropping (%) PerCropping
mean channel slope (%) ChannelSlope horticulture (%) PerHorti
Land cover grasses (%) Grasses catchment fertilized (%) PerFertilized
forest (%) Forest forestry (%) PerForestry
shrubs (%) Shrubs urbanized (%) PerUrbanized
4
woodland (%) Woodland maximum barrier free flow path length upstream (reservoirs) (km) UpstreamReser
bare (%) Bare maximum barrier free flow path length upstream (damwalls) (km) UpstreamDam
average width of vegetated riparian zone MeanVegW_m Climate average annual radiation (MJ/m2/day) AnnRad
average catchment riparian zone fragmentation FraRipaZone average temperature (◦ C) AnnTemp
Geology catchment underlain by regolith (%) PerUnconsolidated minimum temperature of coldest month (◦ C) ColdMonthTemp
igneous rock (%) PerIgneous maximum temperature of hottest month (◦ C) HotMonthTemp
mixed igneous and sedimentary (%) PerMixIgSed annual average rainfall (mm) AnnRain
sedimentary rock (%) PerSedimentary average rainfall of the warmest quarter (mm) WarmQRain
cation exchange capacity MeanCaExCap average rainfall of the coldest quarter (mm) ColdQRain
acid sulphate level B (%) PerAcidS_B annual average catchment rainfall erosivity (MJ mm/ha hr yr) Erosivity
mean TN levels in soil (mg/kg) MeanTN Hydrology average annual runoff (mm) AnnRun
mean TP levels in soil (mg/kg) MeanTP maximum annual runoff (mm) MaxRun
Clay Content (%) Clay_30 pereniality of runoff (%) RunPereniality
mean soil erodibility MeanSoilEro mean base flow index Mean_BFI
soil pH pH runoff ratio Mean_RR
plant available water capacity (mm) PAWC mean number of days where there is no flow annually (days/year) Cease_to_Flow
bulk density (mg/m3) Bulk_density mean 7-day low flow (ML/d) Sevendaylowflow
Environmental Pollution 288 (2021) 117337

2.2.2. Runoff event delineation and event-mean concentration calculation the Geofabric tool provided by the Australian Bureau of Meteorology
Individual runoff events were delineated based on the continuous (Bureau of Meteorology, 2012). We obtained the 58 catchment charac
discharge record extracted from the Water Monitoring Information teristics that represent conditions upstream of the monitoring sites, from
Portal (DNRME, 2018). To identify each event, we used an automated publicly available data sets (category and abbreviation are as detailed in
hydrograph analysis tool – HydRun (Tang et al., 2017). The start and end Table 1). More detailed description, source and summary statistics can
points of a specific event were determined by using a local-minimum be found in Tables S7 and S8, Supplementary Material. ArcGIS 10.5 was
method through calculating the first derivative of streamflow (first used to extract the catchment average from gridded raster (e.g., catch
separated from baseflow). The event-mean concentration (EMC) was ment rainfall) data or proportion of catchment coverage for polygon (e.
then calculated for each event that had at least two samples on both the g., land use) data. Prior to the analyses, both response variables (i.e.,
rising limb and falling limbs of the hydrograph. This ensured sufficient EMCs in Section 2.2.1) and explanatory variables (i.e., 58 catchment
samples taken over the runoff hydrograph, and the reliability of derived characteristics) were standardized to a mean of 0 and standard deviation
EMC (Waters et al., 2007). It is worth noting that uncertainty in baseflow of 1 to enable comparison between explanatory variable regression co
estimation has a direct impact on pollutant load estimation (Binns et al., efficients (i.e., relative importance of predictors) (Cade, 2015). Some
2018), but a preliminary sensitivity analysis (Figure S2) indicated the catchment characteristics were strongly cross-correlated (e.g., Spear
influence on calculated event-mean concentration was small (only a man’s rank correlation coefficient ρ = − 0.77 for stream density and
9.15% change in EMC under different filter coefficient β across all grazing in Figure S3). The collinearity between predictors will likely
catchments). For each event, the EMC of a constituent was calculated as result in multiple models having similar performance. We will discuss
the total load per unit flow volume within the event (Joo et al., 2012), as the issue of collinearity between predictors in the analysis of the results
follows: (Section 4.1.1).
∑
n
cj +cj+1
Event Load 2
× qj+1/2 × tj+1/2 2.3. Statistical modeling
(1)
j=0
EMC = = ∑
n
Event Flow Volume
j=0
qj+1/2 × tj+1/2 We first mapped the time-averaged EMC values for the nine con
stituents across the GBR catchments. In addition, we undertook corre
where n is the total number of samples for a given event, cj is concen lation analysis between each pair of water quality constituents that
tration of the jth sample, qj+1/2 and tj+1/2 are the inter-sample mean aimed to evaluate whether common patterns exist for different constit
discharge and time interval between jth and (j + 1)th samples. We uents. To assess the effect of each catchment characteristic, we used
identified the concentration at the start and end of all events (c0 and multi-model inference to investigate the key controlling catchment
cn+1) by assuming they were the average concentration of all baseflow characteristics. Once the key catchment characteristics driving spatial
samples. The EMCs we developed were essentially a flow-weighted differences in riverine water quality were identified, predictive models
mean concentrations over different runoff events, which allowed us to were built. Multi-model inference in this study involves three steps,
compare water quality across catchments with contrasting flow regimes namely, (1) identification of multiple plausible models using linear
(Cooke et al., 2000). regression, (2) predictions using a model-averaging approach and (3)
The site-level average of constituent EMCs (i.e., average of all model assessment. Fig. 2 shows the schematic diagram of methods that
available EMCs for each constituent at each site) were calculated were used in these three steps. The analyses were performed in MATLAB
(summary statistics in Table S4) for use in statistical modeling. Prior to version R2017b by MathWorks, Inc. (MATLAB and Statistics Toolbox,
analysis, the site-level average EMCs for each constituent were Box-Cox 2017).
transformed using the car package in R (Fox et al., 2012), to improve the
data symmetry (Box et al., 1964). For each constituent, the Box-Cox 2.3.1. Identification of plausible models
parameter λ was estimated individually (λ can be found in Table S5, To identify the plausible models to predict average EMCs for each
Supplementary Material). All the transformed variables were normally constituent, and reduce the computational burden, we adopted a two-
distributed based on the Shapiro-Wilk’s test (Table S5) (Shapiro et al., round exhaustive search approach (i.e., two stages in Analysis in
1965; Steinman et al., 2018). While the focus of this study is the spatial Fig. 2) (Guyon et al., 2003; Lintern et al., 2018b; May et al., 2011; Saft
variability in water quality (between-site), we acknowledge that high et al., 2016). Adopting a two-round exhaustive search approach had two
temporal variability (within-site) would result in large uncertainty in advantages: 1) results were more interpretable; and 2) model complexity
time-averaged water quality. To test for this, we further decomposed the was controlled by information criterion, thus avoided overfitting (Gre
spatial and temporal variance in EMCs (Figure S3). The spatial compo gorutti et al., 2017; Vatcheva et al., 2016). We used ordinary least
nent of the variance was derived from the deviation of site-level average squares to fit the all candidate models in both rounds of the exhaustive
EMC and globally-averaged EMC for specific constituents. The temporal search.
component was derived from the deviation of individual EMCs and In the first round of the Exhaustive Search, we aimed to truncate the
site-level average EMC for specific constituents. The partitioning of number of predictors to a more manageable level to reduce computa
these two components (Figure S3) indicated that spatial variation ac tional overheads. All possible combinations of predictors - up to a
counts for the majority of variability for all constituents except PP and maximum of five predictors - were used to construct linear additive
DOP. Given this result, we focused the analysis on analyzing spatial models to predict EMCs. The selection of a maximum of five predictors
differences in water quality across different catchments and did not was based on previous exhaustive studies of similar dimensions, to limit
consider temporal variation further in this study. the computational demand and to avoid overfitting (Lintern et al.,
2018b; Saft et al., 2016).
2.2.3. Catchment characteristics data collection We assessed and compared all possible models derived from the first
Based on previous studies (Chang, 2008; Kleinman et al., 2004; round using the Corrected Akaike Information Criterion (AICc), which is
Lintern et al., 2018a, 2018b), there is a wide range of natural and preferred for small sample applications (Hurvich et al., 1989). The
anthropogenic catchment factors that affect spatial variability in water model weights wi are calculated as follows (Burnham et al., 2002; Ye
quality. Among these factors, catchment topography, land cover, land et al., 2008),
use, geology, climate and hydrology most commonly affect key pro /
∑
N
cesses (i.e., source, mobilization and delivery) in catchments (Table S6). wi = e− 0.5ΔAICc i e− 0.5ΔAICc n (2)
Catchment boundaries of the 32 monitoring sites were delineated using n=1
5
Fig. 2. The schematic diagram of overall method used in this study. The two-round exhaustive search (Analysis) was applied to:1) entire catchments as full model
results; and 2) resampling of catchment with replacement (bootstrap) 500 times as uncertainty assessment results for model assessment.
where N is total number of models, and ΔAICc is the difference in AICc The predictions of the individual plausible models were averaged
between model i and the minimum AICc. The AICc quantifies both the using the weighting coefficients defined by ΔAICc, which was the same
model performance and model complexity, which allowed us to method we used to calculate the model weights wi (Equation (2)). The
compare and rank the candidate models. The model weight wi is a model weights were then used to compute model-averaged predictions,
transformation of AICc that provides the evidence/likelihood of ith defined as:
model being the best model with the minimum information loss (Poeter ∑
N
et al., 2005). y= wi y i (3)

Model weights were used to estimate the relative importance of in i=1
dividual explanatory variables by summing wi for each model in which

where yi is a vector of the estimated EMCs across all sites, wi is the
that explanatory variable appeared. This is defined as the Proportion of
weighting coefficients for model i, N is total number of plausible models
Evidence (PoE) for each predictor (Mohan et al., 2018; Saft et al., 2016).
(ΔAICci < 2) for each constituent and y is a vector of the weighted
If a predictor appears more frequently in models with small Δi (i.e.,
predictions.
higher relative performance), then the PoE of that predictor is close to 1.
To compare the effects of predictors selected in the plausible models
This allowed us to consider the relative importance of individual
on the response variable, the weighted parameter coefficients (βj ,
exploratory variables across all models, and hence identify key pre
dictors. We retained the 20 catchment characteristics with the highest Equation (4)) can be calculated as follows,
PoE for the second round of the exhaustive search. This truncated ∑
N
number of potential predictors allowed us to explore more possible βj = wi βi,j (4)

combinations of explanatory variables without excessive computational i=1
requirement.
where βi,j is the fitted model coefficients of predictor j in a given model i,
In the second round of the exhaustive search, we aimed to establish
and wi is the weight of model i. βj was only averaged over the models that
the multi-model modeling averaging structure. We fitted all possible
models with up to 10 predictors for each constituent. This ensured that included the predictor of interest (Lukacs et al., 2010).
the final models all had a minimum subjects per variable (SPV, ratio of
number of observations to number of predictors) larger than 3. Previous 2.3.3. Model assessment
studies have shown that this results in adequate estimates of regression The averaged model predictions were evaluated using the Nash-
coefficients (Austin et al., 2015). Sutcliffe coefficient (NSE) (Nash et al., 1970). A residual assessment
was performed to check: (1) normality of residuals; and (2) hetero
2.3.2. Model averaging scedasticity in residuals (i.e., no clear relationship between residual and
We used the same information criterion (i.e., AICc) to assess each predictors that were included in the model averaging structure).
model (Bozdogan, 1987) and to identify the multi-model averaging Three additional assessments were performed. Firstly, we used a
structure to predict average water quality conditions. We identified all bootstrap approach to quantify the uncertainty in the relative impor
plausible models with AICc difference (ΔAICci), compared to the best tance of the catchment characteristics (i.e., PoE). We assessed this un
model (i.e., the lowest AICc). All models with ΔAICci less than 2 were certainty based on the statistical assumption that individual models of
identified as plausible models as per previous recommendations the multi-model ensemble are from the same population and that the
(Burnham et al., 2002; Foglia et al., 2013). central limit theorem applies (Fischer, 2010). The entire set of catch
ments was sampled randomly with replacement to obtain a bootstrap
6
data set, which was used as input for the identification of plausible the weighted parameter coefficients across the 500 replicates of
models and model-averaging (Sections 2.3.1 and 2.3.2, Analysis in bootstrapping (i.e., 95% CI does not cross zero).
Fig. 2). This was repeated 500 times for each constituent to ensure a
sufficient number of bootstrap samples to obtain an estimate of the Only predictors with selected frequencies larger than 50% and with
bootstrap distribution (Carpenter et al., 2000). We used two indices to significant coefficients were designated as important factors to be dis
assess the robustness of the selection of important catchment cussed further. In addition, the model performance derived from the
characteristics: multi-model averaging approach was compared with that of the single
best model. Specifically, for each bootstrap resampling run, we
1) the distribution (median and 95% bootstrap confidence interval (CI)) compared the difference in each pair of NSEs between the multi-model
of PoE of all 58 catchment characteristics derived from the first averaging structure and the single best model (i.e., the model with
round; and lowest AICc identified in the second round of the exhaustive search). We
2) the significance of selected catchment characteristics, which was assigned a 1 if the NSE of the multi-model averaging structure was
evaluated using the frequency of catchment characteristics selected greater than the single best model, and zero otherwise. The probability
in the final model-averaging structures, as well as the distribution of of the multi-model approach providing an improved performance was
Fig. 3. Proportion of Evidence (PoE) from the first-round of exhaustive search derived from full model (red triangle) and 500 runs of bootstrapping (colored dots and
bars represent the median and 95% CI, respectively) of each catchment characteristics for: (a) total suspended solids (TSS); (b) oxidized nitrogen (NOX) and (c)
filterable reactive phosphorus (FRP). Catchment characteristics within each category (represented by six different colors) were plotted in descending order of the
corresponding median PoE. The definition of abbreviation of each catchment characteristic can be found in Table 1. (For interpretation of the references to color in
this figure legend, the reader is referred to the Web version of this article.)
7
Fig. 4. Results from 500 runs of bootstrapping: upper panel indicates the frequency of significant catchment characteristics selected in the final plausible models;
lower panel indicates the median (dot) and 95% CI (error bar) of the weighted coefficient for each significant catchment characteristic for: (a) TSS; (b) NOX and (c)
FRP. Different colors represent categories. The definitions of abbreviations for each catchment characteristic can be found in Table 1. (For interpretation of the
references to color in this figure legend, the reader is referred to the Web version of this article.)
estimated by averaging this statistic over the 500 bootstrap runs. 3. Results
Secondly, to compare the relative importance of the natural and
anthropogenic (i.e., land use) catchment characteristics, the averaged 3.1. Spatial pattern of averaged EMCs
model structure was re-calibrated using only natural catchment char
acteristics (i.e., land use characteristics were excluded). This allowed us The average EMC of the nine constituents showed different spatial
to understand the relative effect of land use when predicting the spatial patterns across the GBR catchments (Figure S5, Supplementary Mate
variation in water quality. rial). Generally, the particulate constituents, i.e., TSS, PN and PP,
Finally, to evaluate the robustness of model performance, a cross- exhibited a similar pattern, where averaged EMCs were lower in the
validation was performed for the full model. Because this assesses the northern region and increased towards the southern sites (Spearman’s
robustness of model performance rather than variable selection, we only Rank cross-correlations among these three constituents ρ > 0.76, p <
focus on the final multi-model structure identified in Section 2.3.1 (i.e., 0.05, Figure S6). A similar pattern was observed for the dissolved species
full model). For each constituent, 80% of catchments were randomly (e.g., NH4, DON, FRP, relationships ρ > 0.45, p < 0.05, Figure S6);
selected to calibrate the multi-model structure, with all held-out however, averaged NOX showed a contrasting spatial pattern, with sites
catchments used for validation of how well the calibrated models per in the coastal regions (e.g., Wet Tropics, Mackay-Whitsunday, Burdekin
formed (i.e., NSE). This was repeated 100 times to obtain an ensemble of and Burnet Mary) having much higher averaged EMCs compared with
model performance on independent datasets and eliminate the effect of other sites.
sampling.
8
Table 2 3.2.1. Key factors identified in plausible models

Model averaging performance (NSE) and weighted predictor coefficients. NSE is To check the consistency between round 1 and round 2 variable se
calculated based on averaged predictions across all identified plausible models. lection, the PoE for each catchment characteristic derived from the full
The third and fourth columns are averaged model performance when consid model was compared to the 95% bootstrap CI from the first round of the
ering only natural catchment characteristics (i.e., land use characteristics are exhaustive search (Fig. 3, S7 and S8), and was found to be within that CI.
excluded), and the change in NSE compared to full model, respectively.
We note that the 95% CIs were much wider for the catchment charac
Constituent NSE - NSE - only % Number of Shapiro- teristics with high PoEs than for those with lower PoEs. This indicated
full natural change plausible Wilk’s test
that while there was large uncertainty in relative importance derived
model catchment in NSE models p-value in
characteristics residuals
from the first round of the exhaustive search, PoE from the two rounds
included were highly consistent (compare red dots to CI in Fig. 3). We infer from
this that the key drivers identified and the relative importance of those
TSS 0.75 0.73 − 2% 21 0.45
PN 0.68 0.66 − 2% 34 0.92 drivers are largely independent of input data. In addition, the distribu
NOX 0.82 0.64 − 22% 3 0.13 tion of weighted parameter coefficients and frequency of significant
NH4 0.85 0.80 − 6% 22 0.32 catchment characteristics in the final plausible models from the boot
DON 0.96 0.96 0% 19 0.81
strapping varied between different constituents (TSS, NOX and FRP in
FRP 0.92 0.91 − 1% 7 0.71
DOP 0.78 0.78 0% 5 0.29 Fig. 4, for the other constituents see Figures S9 and S10, Supplementary
PP 0.83 0.82 − 1% 11 0.36 Material). Note that significant predictors identified in Fig. 4, S8 and S9
EC 0.95 0.87 − 8% 2 0.67 have all been identified based on the full models (i.e., after round 2) for
Note: Shapiro-Wilk’s test p-value > 0.05 indicates the acceptance of the null individual constituents.
hypothesis that data comes from a normal distribution at the 5% significant
level. 3.2.2. Model-averaging performance
The number of likely models determined from the second round of
3.2. Multi-model inference on modeling of spatial variability in EMCs the exhaustive search for each constituent ranged from 2 for EC to 34 for
PN (a detailed summary of each likely model is available in Table S9).
Results presented below will focus mainly on three constituents (i.e., NSE ranged from 0.68 for PN to 0.96 for DON (Table 2). In addition, the
TSS, NOX and FRP), due to their relatively high risk to the receiving exclusion of anthropogenic characteristics resulted in a large decrease in
marine environment (Bartley et al., 2017; Waterhouse et al., 2017). NSE for NOX (− 22%, Table 2) but had a small (<10%) to minimal effect
Results of the other six constituents are presented in the Supplementary on the other constituents. The residual analysis showed no clear heter
Material. oscedasticities in model residuals (Figures S11 to S19), and both visual
check (Figure S20) and the Shapiro-Wilk’s tests (all p-value > 0.05,
Table 2) indicated that residuals were normally distributed.
Table 3 The averaged models demonstrated a good fit between the pre
Comparison of model performance between multi-model and single-best model dictions and observations, for both full model and bootstrap analysis
structures for nine constituents, in terms of median NSE and probability of the results (Figure S21, Supplementary Material). In addition, based on the
multi-model structure providing better performance (i.e., higher NSE) across
bootstrap results, median NSEs from multi-model structure are higher
500 bootstrap subsamples.
than single-best models for all nine constituents, and the probability that
Constitute Median NSE Probability of the multi-model averaging structure performs better than the single best
the multi-model
Multi-model Single-best model model ranged from 76% to 92% (with an average of 83% across nine
structure
providing better constituents) (Table 3), indicating a higher predictive capacity of the
performance multi-model averaging structure.
TSS 0.51 0.34 0.85 The calibration and validation results for 100 runs of cross-validation
PN 0.32 0.28 0.92 are summarized in Table 4. For all constituents, the performance of
NOX 0.54 0.52 0.77 calibrated models was consistent with the full models. In addition,
NH4 0.71 0.62 0.89 expect for NOX, the validation runs showed only a slight decrease in
DON 0.96 0.93 0.85
median NSEs compared with calibration results (e.g., less than 0.2 dif
FRP 0.82 0.81 0.76
DOP 0.45 0.32 0.77 ference in NSE). The larger drop in validation performance for NOX
PP 0.50 0.44 0.83 might be explained by the strong effects of anthropogenic factors
EC 0.56 0.33 0.80 affecting the spatial variability in NOX – we further discuss this impact in
Section 4.1.1. Overall, the cross-validation analysis indicates that the
multi-model structures are robust and less senstitve to the selection of
input data.
4. Discussion
Table 4
Comparison of NSEs of the full model and 100 runs of cross-validation models
for each constituent. 4.1. Influential factors affecting spatial variation in stream water quality
Constitute Full model Median NSE - 100 Cross-validation
4.1.1. The relative importance of natural and anthropogenic landscape
Calibration Validation characteristics
TSS 0.75 0.74 0.63 For most constituents, natural catchment characteristics were more
PN 0.68 0.65 0.52 important than anthropogenic landscape factors for predicting spatial
NOX 0.82 0.84 0.47
differences in water quality. Except for NOX, exclusion of anthropogenic
NH4 0.85 0.86 0.73
DON 0.96 0.97 0.89 predictors in model building did not influence the prediction perfor
FRP 0.92 0.93 0.73 mance for any constituents markedly (Table 2). It is worth noting that
DOP 0.78 0.74 0.61 the change in NSE measures the effect of anthropogenic variables rela
PP 0.83 0.80 0.62 tive to the background variation and that this background variation is
EC 0.95 0.91 0.82
large for some constituents, such as the particulate species (e.g., TSS, PN
9
and PP), where there is strong natural variation in sediments due to rainfall and grazing land use (ρ = − 0.89, p < 0.05, Figure S4) shows that
climate, vegetation and soils (Lintern et al., 2018a). grazing agriculture is mostly conducted in dry catchments (e.g., Fitzroy
In contrast, land use measures (e.g., proportion of sugarcane and Burdekin regions in Fig. 1). Vegetation cover also tends to be lower
farming) had high predictive power for NOX, NH4 and FPR, demon in these catchments. These catchments contribute a large proportion of
strating that in the GBR catchments specific land use was an important sediments to the GBR, due to a higher potential for gully and hillslope
driving factor in spatial variability of EMCs of dissolved inorganic nu erosion (Bartley et al., 2018; Kuhnert et al., 2012).
trients across catchments. These results are consistent with previous
findings that land use changes are related to sources of dissolved forms 4.1.3. Topography
of nutrients in the GBR catchments (Brodie et al., 2003; Hunter et al., Catchment topographic characteristics were identified as important
2008; Mitchell et al., 2009). factors for dissolved nutrients. Catchment elevation had a significant
In general, natural catchment characteristics exhibited higher pre effect on DON and NH4, with negative weighted regression coefficients.
dictive power than anthropogenic catchment characteristics. Moreover, Meanwhile, there was a negative effect of catchment slope on DOP.
it is worth noting that natural catchment characteristics can have a These factors are relevant to pollutant mobilization and delivery. In the
direct impact on anthropogenic drivers, such as agricultural activities GBR catchments, low-elevation, small coastal catchments are relatively
which are strongly influenced by climate (Hatfield et al., 2011; Thor steeper with a high relief ratio (Figure S4). The topographic features also
burn et al., 2013). This leads to a cross-correlation between natural and influence aquifer dynamics that control the fraction of groundwater in
anthropogenic drivers which might reduce the anthropogenic signal surface runoff (Skoulikidis et al., 2006). Lower-lying land is typically
detected in our analysis because excluding anthropogenic effects might associated with higher groundwater contribution, leading to an increase
be compensated for through correlation with natural drivers. Indeed, we in the concentration of dissolved nutrients in streams (Kratz et al., 1997;
found that correlated predictors do not necessarily have similar pro McKergow et al., 2003).
portions of evidence (e.g., stream density and grazing have − 0.77 The large effects of stream density for NOX (Fig. 3(b)) can be
Spearman’s ρ, but proportions of evidence are 0.43 and 0.04 for NOX, explained by the catchment export processes for inorganic nitrogen,
respectively). This is because multi-model inference can handle the which are strongly controlled by density of stream network (Alexander
collinearity by shrinking the PoE of one of the correlated variables to et al., 2002; Prasad et al., 2005). Catchments with denser stream net
ward zero (Nakagawa et al., 2011; Posch et al., 2020; Walker, 2019). works are more likely to have shorter runoff pathways to receiving
This shrinkage effect leads to a lower weight of the more complex model, waters. This may lead to more rapid delivery of NOX with lower losses
if this more complex model only differs from a simpler model by having from denitrification (Young et al., 1996).
additional correlated variables. This is because the more complex model
receives a higher penalty in AICc. The more complex model would only 4.1.4. Geology
be favored if the benefit in higher predictive capability outweighs the Catchment geology and soil type were important for all constituents,
cost in higher model complexity. This is often not the case as including except for DON. Catchment lithology was one of the most frequently
correlated variables often does not increase the model predictive ca selected predictors. The percentage of catchment underlain by uncon
pacity (Daoud, 2017; Hinne et al., 2020; Kruschke, 2014). Collectively, solidated regolith materials (e.g., colluvium and alluvial deposits)
the results demonstrate that spatial variation in water quality tends to be (PerUnconsolidated) has a negatively weighted regression coefficient for
better explained by catchment natural landscape characteristics. particulate species, i.e., TSS, PN and PP. However, the positive regres
sion coefficients of mixed sedimentary and igneous rock (PerMixIgSed)
4.1.2. Climate for PP demonstrate that sedimentary and igneous deposits may act as a
Climatic variables had high PoEs and consequently were included in source of phosphorus. Phosphate minerals can be derived from sedi
the plausible models for the majority of constituents (Fig. 3, S6 and S7). mentary and igneous deposits (mainly in forms of phosphorites and
Air temperature was identified as a key factor among these climatic apatite, respectively), and the release of dissolved phosphorus in phos
characteristics. It is not surprising since temperature affects almost all phate is enhanced due to weathering and hydrological transport (Holtan
physio-chemical processes and biological reactions for nutrients (Huang et al., 1988; Pufahl et al., 2017). During the wet season in the GBR
et al., 2003; Lintern et al., 2018a; Sardans et al., 2008). The maximum catchments, the increased water availability is likely to enhance the
temperature of hottest month (HotMonthTemp) was identified as signif erosion and chemical weathering of the bedrock (Bouchard et al., 2000).
icant predictor for TSS, PN and PP, with a strong positive effect on The weathered material tends to be transported via surface and sub
constituent concentration. On the other hand, the average lowest min surface runoff, leading to an increase in sediment-bound phosphorus
imum temperature in each year (ColdMonthTemp) had a strong negative (Hattanji et al., 2004; Pelletier et al., 2011). There is negative feedback
effect on NOX. In our study region, climate and hydrology (e.g., between this process and the thickness of regolith (e.g., unconsolidated
discharge) have a clear seasonal pattern with most high flow events in materials), such that catchments underlain by deeper unconsolidated
the wet/hot summer (typical time when the EMCs are derived), and few materials might experience lower levels of particulate and dissolved
events in the relatively cold/dry winter. High sediment and particulate nutrient species (Strudley et al., 2006).
nutrient concentrations in streams can be expected during hot/wet
seasons due to high erosion (Kronvang et al., 1997; Sherriff et al., 2016). 4.1.5. Land cover
During the cold/dry periods, nutrients are more likely to accumulate in The impact of land cover characteristics in the GBR catchment were
the soil, such that there are increases in nutrient availability for the not as evident as geological and topographic characteristics. Land cover
subsequent summer wet season (Edwards et al., 2008; Houser et al., was only identified as an important predictor for dissolved nutrients (e.
2010; Pionke et al., 1999). It is worth noting that such seasonal patterns g., NOX, NH4, DON and FRP). We note that the land cover had a con
are not evident for the catchments in the Wet Tropics, due to continuous trasting influence on NOX (Forest, positive) compared to other constit
high flow events throughout the year, reducing the likelihood of nutri uents, e.g., FRP and DON (average catchment riparian zone
ents building-up in the soil. fragmentation (FraRipaZone), negative). Generally, nutrients might be
Annual rainfall (AnnRain) was another controlling factor with a expected to be inversely correlated with riparian vegetation cover due to
consistently negative impact on TSS, PN, and PP. This result contrasts their reduction via biogeochemical processes (e.g., plant uptake and
with previous studies in other catchments (Cavelier et al., 1997; Granger denitrification) and sedimentation of particulate forms (Johnson et al.,
et al., 2010; Perona et al., 1999). The contrast might be due to the high 1997; Varanka et al., 2015). However, our results for the effect of ri
interaction between rainfall and land use/land cover in the GBR parian vegetation on oxidized nitrogen is the opposite to such findings.
catchments. For example, an inverse relationship between annual In the GBR catchments, a large proportion of NOX export occurs in the
10
wet regions, where rainfall and subsequent flow events are frequent. In data and study area characteristics. Firstly, the water quality monitoring
such areas, the riparian trapping effect for NOX is limited due to shorter data in our study focused on runoff events, rather than the monthly
residence time and greater contribution from subsurface flow (McKer sampling used in Lintern et al. (2018b). We considered the variability in
gow et al., 2003; Meynendonckx et al., 2006). One caveat on this result streamflow when developing EMCs using event-based water quality
is the data source we used for riparian vegetation. The riparian vege samples. This reduced the uncertainty associated with concentrations of
tation information used in the study was measured using Landsat sat monthly samples, and our samples are more reflective of high flow
ellite imagery. A buffer of 100 m on both sides of rivers and riverine conditions. This might result in the averaged EMCs of dissolved nutri
wetlands was considered to encompass the riparian area (Clark et al., ents being more strongly influenced by catchment natural characteris
2015; Queensland Government, 2016). This might introduce bias when tics (e.g., hydrology for NOX and geology for FRP) than was the case for
the actual width of riparian zone is less than the spatial resolution of the Lintern et al. (2018b). Secondly, these two studies also have the
Landsat imagery (i.e., 30 m) (Woodcock et al., 1994). following specific differences in key processes driving dissolved and
particulate pollutants: 1) land use (e.g., sugarcane) has a clear linkage to
4.1.6. Land use dissolved nutrients in our study area but there is no sugar cane and less
Agricultural activity (e.g., sugarcane farming and horticulture) had a intensive cropping in the Victorian catchments, and 2) the two study
significant and positive relationship with NOX, NH4 and FRP, indicating areas have contrasting climates (tropical and temperate), leading to,
these anthropogenic land uses act as sources of instream pollutants. This among other differences, flow regimes and rainfall being more variable
is in accordance with previous studies that concluded sediment, nutri in the GBR catchments compared to Victoria. Both differences can lead
ents and salts in rivers are likely to be sourced from agricultural activ to higher sensitivity of dissolved nutrients to catchment characteristics
ities (Afed Ullah et al., 2018; Liu et al., 2018; Teixeira et al., 2014). and thus the higher predictive capacity of dissolved nutrients in the GBR
The effect of land use is particularly evident for dissolved inorganic catchments.
nitrogen (i.e., NOX) and dissolved phosphorus (i.e., FRP). There is a clear Our modeling was developed using a multi-model inference
association between dissolved inorganic nitrogen and sugarcane (Per approach. Fig. 3, S6 and S7 demonstrated that there is large uncertainty
Sugar) production in the GBR catchments. The majority of the GBR in relative importance of individual catchment characteristics (as indi
sugarcane is concentrated in the coastal regions of the Wet Tropics, cated by the wide 95% CIs). Thus, we cannot only rely on the single best
Mackay-Whitsunday, Burdekin and Burnet Mary (Fig. 1) (Bainbridge model structure when we aim to better understand the effect of catch
et al., 2009; Hunter et al., 2008; Mitchell et al., 2009). Higher inorganic ment characteristics and provide predictions. Our results indicate that:
nitrogen (Figure S5) is linked with application of urea-based nitrogenous 1) the multi-model approach was better able to consistently identify
fertilizer, which transforms to NH4 and NO3 and enters waterways important factors among a large number of candidate predictors; and 2)
through surface and subsurface pathways (Connolly et al., 2015; Davis compared to the conventional single best model approach (Ekholm
et al., 2016; Thorburn et al., 2017). et al., 2000; Sangani et al., 2015; Varanka et al., 2015), multi-model
inference provided greater predictive ability (Table 3). This is due to
4.1.7. Hydrology the model (variable) selection uncertainty being considered inherently
Catchment hydrological characteristics showed significant negative within the multi-model approach (Cade, 2015; Parrish et al., 2012; Ye
effects on most dissolved nutrient species (i.e., FRP, NH4 and DON). This et al., 2008).
can be linked to the strong dilution effect of high flow on dissolved This research was limited to catchments in the north-eastern part of
species, especially in the wet regions (e.g., Wet Tropics) (Orr et al., Queensland, Australia. Even though the study catchments featured
2014). Runoff ratio and runoff pereniality (percentage contribution to diverse landscapes (ranging from humid to semiarid tropical catch
mean annual discharge by the six driest months of the year) appear as ments) (Bell, 2001; Gilbert et al., 2001), there is still an issue of whether
predictors with strong explanatory power for DON and NH4, respec the understanding gained in this research could be transferred to other
tively. The result is not surprising since catchment hydrology is highly catchments in Australia and elsewhere. The controlling factors driving
correlated with other catchment characteristics (e.g., ρ = - 0.85 for the spatial differences in water quality might vary region by region.
runoff ratio and grazing agriculture, and ρ = 0.82 for baseflow index and These differences in catchment processes could be attributed to: 1)
soil TN level, Figure S4). Prathumratana et al. (2008) found that the different baseline hydroclimatic and ecosystem conditions (Aubert et al.,
inter-correlation between catchment runoff, temperature and rainfall 2013; Nilsson et al., 2008); or 2) different processes (e.g., large floods or
could explain the spatial variation of sediments and nutrients in the prolonged drought events) that lead to trends and changes in water
lower Mekong River. A possible reason is that catchment hydrology is quality conditions (Elchyshyn et al., 2018; Li et al., 2018; Richards et al.,
more likely to affect the temporal variation in water quality than spatial 2002). Hydrological conditions could be a more important factor in
variability (Chen et al., 2007). However, the negative effect of the runoff other catchments (e.g., the Lower Murray River catchments in South
ratio and runoff perennially on DON and NH4 implies that EMCs of Australia) compared with the wetter GBR catchments considered in this
dissolved nitrogen are strongly affected by the temporal changes in study (Kingsford et al., 2011; Mosley et al., 2012). Therefore, the gen
runoff volume contributing to the annual runoff due to the dilution ef erality of findings from this research should be considered carefully, and
fect (Joo et al., 2012). the local geographic conditions of any new application should be taken
into account. Nevertheless, the results of this study provide an indication
4.2. Predicting spatial variation in averaged EMCs of what might be important in other tropical regions such as north
eastern coast of Brazil (de Arruda-Santos et al., 2018; Maciel et al.,
The statistical modeling framework proposed in this study was 2015), and the eastern coastal region of India (Damodharan et al., 2012;
pragmatic and provided a simple approach for assessing average water Govindaraj et al., 2011), where understanding of the drivers of water
quality conditions during runoff events across a large tropical region. quality is currently lacking.
The weighted prediction derived from the model averaging approach
performed well and captured a large proportion of spatial variability in 4.3. Management implications and future research
water quality. For all constituents we studied, we had better ability to
predict the dissolved nutrient species than particulate pollutants. This is Catchment pollutant management requires us to recognize the
in contrast to an earlier study (Lintern et al., 2018b), which investigated importance of a wide range of catchment characteristics, especially how
the linkage between spatial variability in average water quality and the effects of changes (e.g., in land use, land management, land cover or
catchment characteristics in 102 catchments in Victoria, Australia. climatic condition under climate change) may alter pollutant export
The contrasting results may be caused by difference in water quality during runoff events. Our analyses show that catchment anthropogenic
11
characteristics are more pertinent to dissolved nutrients (e.g., NOX, NH4 the work reported in this paper.
and FRP). Therefore, continuous monitoring of changes in land use and
management, as well as water quality responses in these intensively used Acknowledgments
catchments would provide improved insight into managing nutrient
sources. This is in line with the current ‘best management practice’ This study was supported by the Australian Research Council
adopted in the GBR catchments (Lintern et al., 2020; Star et al., 2018; (LP140100495), the Environment Protection Authority Victoria, the
Thorburn et al., 2013). Our results also indicate that spatial variability in Victorian Department of Environment, Land, Water and Planning, Bu
particulate pollutants is more directly influenced by a catchment’s reau of Meteorology and Queensland Department of Resources. The
natural characteristics. This does not necessarily imply that land use is author would like to acknowledge the efforts of the Queensland
unimportant for particulate constituents. For example, grazing is the Department of Environment and Science who provided the water quality
dominant land use in the GBR catchments. Hence, departures from this monitoring data. The authors would also like to offer sincere gratitude to
land use (e.g., presence of sugarcane) may be more useful in predicting Ms. Jie Jian for her assistance in geospatial database compilation. Dr.
water quality spatial variation. More detail on the grazing intensity (e.g., Paul Leahy, Mr. Malcolm Watson, Dr. Ulrike Bende-Michl, Mr. Paul
number of cattle per area) could potentially be useful in predicting Wilson, and Ms. Belinda Thompson all provided valuable advice in the
variations in grazing dominated catchments (Smith et al., 2013); how preparation of this manuscript. Water quality data (derived site-level
ever, such detailed survey data are currently unavailable for this region, averaged EMC) and catchment characteristics data used for the statis
and so, future efforts are needed to quantify grazing intensity. tical analyses in this paper has been uploaded on the University of
Our study focused solely on the averaged water quality conditions for Melbourne research data repository (melbourne.figshare.com) with the
runoff events. While this might be useful for the long-term planning of following DOI: https://doi.org/10.26188/5de8d8f2da817. Sources of
improved management practice, it does not provide insight into the these catchment characteristics are provided in Table S7 in Supple
temporal variation in water quality. Thus, future analyses will investi mentary Material.
gate the driving factors (e.g., changes in discharge and land cover) that
influence the temporal variability in water quality (e.g., deviation of Appendix A. Supplementary data
individual EMCs from site-level averaged EMC – the temporal variation
component in Figure S3), which is also of great importance (Brodie Supplementary data to this article can be found online at https://doi.
et al., 2010; Guo et al., 2019; Liu et al., 2021). The incorporation of org/10.1016/j.envpol.2021.117337.
spatial and temporal modeling variability into a single modeling
framework would provide a comprehensive understanding of how water
References
quality changes across space and over time (Guo et al., 2020). In addi
tion, it is noted that pesticides, which pose a direct threat to the GBR Afed Ullah, K., Jiang, J., Wang, P., 2018. Land use impacts on surface water quality by
lagoon ecosystem (Haynes et al., 2000; Lewis et al., 2009), were not statistical approaches. Global Journal of Environmental Science and Management 4
included in this study. Further investigation could extend our existing (2), 231–250.
Alexander, Richard B., Elliott, Alexander H., Shankar, Ude, McBride, Graham B., 2002.
modeling framework to include these emerging chemicals. Estimating the sources and transport of nutrients in the waikato river basin, New
Zealand. Water resources research, 38(12).
5. Conclusions APHA, 2005. Standard methods for the examination of water and wastewater. American
Public Health Association (APHA): Washington, DC, USA.
Aubert, Helene, Alice, Gascuel-Odoux, Chantal, Merot, Philippe, 2013. Annual hysteresis
In this study, a data-driven statistical approach was used to identify of water quality: a method to analyse the effect of intra-and inter-annual climatic
the important factors affecting the spatial differences in water quality in conditions. J. Hydrol. 478, 29–39.
Austin, Peter C., Steyerberg, Ewout W., 2015. The number of subjects per variable
the Great Barrier Reef catchments. We used a multi-model approach to required in linear regression analyses. J. Clin. Epidemiol. 68 (6), 627–636.
identify the influential characteristics for spatial variability in water Australian and Queensland governments, 2020. Methods, Reef Water Quality Report
quality and to make predictions more reliably and robustly, compared to Card 2019. Brisbane. https://www.reefplan.qld.gov.au/__data/assets/pdf_file/0019/
211672/report-card-2019-methods-combined.pdf report.
the single best model that has been often used in previous statistical
Bainbridge, Zoe T., Brodie, Jon E., Faithful, John W., Sydes, Damon A., Lewis, Stephen
modeling of water quality. Our results indicate that natural catchment E., 2009. Identifying the land-based sources of suspended sediments, nutrients and
characteristics explain more variation in water quality than anthropo pesticides discharged to the great barrier reef from the tully–murray basin,
genic characteristics, although land use is strongly related to dissolved Queensland, Australia. Mar. Freshw. Res. 60 (11), 1081–1090.
Bartley, Rebecca, Waters, David, Turner, Ryan, Kroon, Frederieke, Garzon-Garcia, Alex,
nutrient concentrations. The models developed were able to predict Kuhnert, Petra, Lewis, Stephen, Smith, Rachel, Bainbridge, Zoe, Olley, Jon, 2017.
average event-mean concentrations well (NSE ranging from 0.68 to Scientific Consensus Statement: land use impacts on the Great Barrier Reef water
0.96). The multi-model averaging framework could be used to identify quality and ecosystem condition, Chapter 2: sources of sediment, nutrients,
pesticides and other pollutants to the Great Barrier Reef, p. 2017.
potential hot spots of water quality concern, at unmonitored locations. Bartley, Rebecca, Thompson, Chris, Croke, Jacky, Pietsch, Tim, Baker, Brett,
This modeling framework also enables evaluation of water quality re Hughes, Kate, Kinsey-Henderson, Anne, 2018. Insights into the history and timing of
sponses to future changes in climate or land use. With ongoing water post-European land use disturbance on sedimentation rates in catchments draining
to the Great Barrier Reef. Mar. Pollut. Bull. 131, 530–546.
quality monitoring data available at multiple GBR catchments, further Beck, Hylke E., van Dijk, Albert IJM., Miralles, Diego G., de Jeu, Richard AM.,
investigations focusing on temporal variability in water quality are Bruijnzeel, L.A., Sampurno, McVicar, Tim, R., Schellekens, Jaap, 2013. Global
essential to advance our understanding of water quality dynamics. patterns in base flow index and recession based on streamflow observations from
3394 catchments. Water Resour. Res. 49 (12), 7843–7863.
Bell, L.C., 2001. Establishment of native ecosystems after mining—Australian experience
Author statement across diverse biogeographic zones. Ecol. Eng. 17 (2–3), 179–186.
Binns, P., Waters, D., 2018. Baseflow separation. Refinement of the Lyne & Hollick baseflow
separation methodology using historical water quality data from Great Barrier Reef
All authors contributed to the design of the research. SL carried out
catchments. Brisbane.
data collation, performed the simulations and prepared the manuscript Blevins, R.L., Lal, R., Doran, J.W., Langdale, G.W., Frye, W.W., 2018. Conservation
with contributions from all co-authors. All authors contributed to the Tillage for Erosion Control and Soil Quality Advances In Soil and Water Conservation
interpretation of the results and provided feedback. (Pp. 51–68. Routledge.
Booth, Derek B., Roy, Allison H., Smith, Benjamin, Capps, Krista A., 2016. Global
perspectives on the urban stream syndrome. Freshw. Sci. 35 (1), 412–420.
Declaration of competing interest Bouchard, Mireille, Jolicoeur, Serge, 2000. Chemical weathering studies in relation to
geomorphological research in southeastern Canada. Geomorphology 32 (3–4),
213–238.
The authors declare that they have no known competing financial Box, George EP., Cox, David R., 1964. An analysis of transformations. J. Roy. Stat. Soc. B
interests or personal relationships that could have appeared to influence 26 (2), 211–243.
12
Bozdogan, Hamparsum, 1987. Model selection and Akaike’s information criterion (AIC): Duan, Qingyun, Newsha K, Ajami, Gao, Xiaogang, Sorooshian, Soroosh, 2007. Multi-
the general theory and its analytical extensions. Psychometrika 52 (3), 345–370. model ensemble hydrologic prediction using Bayesian model averaging. Adv. Water
Brainwood, Meredith A., Burgin, Shelley, Maheshwari, B., 2004. Temporal variations in Resour. 30 (5), 1371–1386.
water quality of farm dams: impacts of land use and water sources. Agric. Water Edwards, A.C., Withers, P.J.A., 2008. Transport and delivery of suspended solids,
Manag. 70 (2), 151–175. nitrogen and phosphorus from various sources to freshwaters in the UK. J. Hydrol.
Bramley, R.G.V., Roth, C.H., 2002. Land-use effects on water quality in an intensively 350 (3), 144–153.
managed catchment in the Australian humid tropics. Mar. Freshw. Res. 53 (5), Ekholm, P., Kallio, K., Salo, S., Pietiläinen, O.-P., Rekolainen, S., Laine, Y., Joukola, M.,
931–940. 2000. Relationship between catchment characteristics and nutrient concentrations in
Brodie, Jon E., McKergow, Lucy A., Prosser, Ian P., Furnas, Miles Hughes, Andrew, O., an agricultural river system. Water Res. 34 (15), 3709–3716.
Hunter, Heather, 2003. Sources of sediment and nutrient exports to the great barrier Fabricius, Katharina E., De’ath, Glenn, Craig, Humphrey, Irena, Zagorskis,
reef world heritage area. Schaffelke, Britta, 2013. Intra-annual variation in turbidity in response to terrestrial
Brodie, Jon, Schroeder, Thomas, Rohde, Ken, Faithful, John, Masters, Bronwyn, runoff on near-shore coral reefs of the Great Barrier Reef. Estuar. Coast Shelf Sci.
Dekker, Arnold, Brando, Vittorio, Maughan, Mirjam, 2010. Dispersal of suspended 116, 57–65.
sediments and nutrients in the Great Barrier Reef lagoon during river-discharge Fischer, Hans, 2010. A history of the central limit theorem: from classical to modern
events: conclusions from satellite remote sensing and concurrent flood-plume probability theory. Springer Science & Business Media.
sampling. Mar. Freshw. Res. 61 (6), 651–664. Foglia, L., Mehl, S.W., Hill, M.C., Burlando, P., 2013. Evaluating model structure
Brodie, J., Waterhouse, J., Schaffelke, B., Johnson, J., Kroon, F., Thorburn, P., Rolfe, J., adequacy: the case of the Maggia Valley groundwater system, southern Switzerland.
Lewis, S., Warne, M., Fabricius, K., 2013a. Reef water quality scientific consensus Water Resour. Res. 49 (1), 260–282.
statement 2013. Fox, John, Weisberg, Sanford, Adler, Daniel, Bates, Douglas, Baud-Bovy, Gabriel,
Brodie, J., Waterhouse, J., Schaffelke, B., Kroon, F., Thorburn, P., Rolfe, J., Johnson, J., Ellison, Steve, Firth, David, Friendly, Michael, Gorjanc, Gregor, Graves, Spencer,
Lewis, S., Devlin, M., Warne, M., McKenzie, L., 2013b. Scientific consensus 2012. Package ‘car’. Vienna: R Foundation for Statistical Computing.
statement: land use impacts on the great barrier reef water quality and ecosystem Gilbert, M., Brodie, J.E., 2001. Population and major land use in the Great Barrier Reef
condition. Department of the Premier and Cabinet, Queensland Government, catchment area spatial and temporal trends. Townsville.
Brisbane 2013. Govindaraj, P., Sindhu, R., Balamurugan, A., Appunu, C., 2011. Molecular diversity in
Brodie, Jon E., Lewis, Stephen E., Collier, Catherine J., Wooldridge, Scott Bainbridge, sugarcane hybrids (Saccharum spp. complex) grown in peninsular and east coast
Zoe, T., Waterhouse, Jane Rasheed, Michael, A., Honchin, Carol, Holmes, Glen, zones of tropical India. Sugar Tech 13 (3), 206–213.
Fabricius, Katharina, 2017. Setting ecologically relevant targets for river pollutant Granger, S.J., Bol, R., Anthony, S., Owens, P.N., White, S.M., Haygarth, P.M., 2010.
loads to meet marine water quality requirements for the Great Barrier Reef, Towards a holistic classification of diffuse agricultural water pollution from
Australia: a preliminary methodology and analysis. Ocean Coast Manag. 143, intensively managed grasslands on heavy soils. Adv. Agron. 105, 83–115 (Elsevier).
136–147. Grayson, R.B., Gippel, C.J., Finlayson, Brian L., Hart, Barry T., 1997. Catchment-wide
Bureau of Meteorology, 2012. Geofabric V2. Retrieved 02/09/2016. ftp://ftp.bom.gov. impacts on water quality: the use of ‘snapshot’sampling during stable flow.
au/anon/home/geofabric/. J. Hydrol. 199 (1–2), 121–134.
Burnham, K.P., Anderson, D.R., 2002. Model Selection and Multi-Model Inference: a Great Barrier Reef Marine Park Authority, 2004. Great Barrier Reef Marine Park Zoning.
Practical Information-Theoretic approach.: Springer-Verlag, New York, USA. Retrieved from. http://www.gbrmpa.gov.au/geoportal.
Burnham, Kenneth P., Anderson, David R., 2004. Multimodel inference: understanding Gregorutti, Baptiste, Michel, Bertrand, Saint-Pierre, Philippe, 2017. Correlation and
AIC and BIC in model selection. Socio. Methods Res. 33 (2), 261–304. variable importance in random forests. Stat. Comput. 27 (3), 659–678.
Cade, Brian S., 2015. Model averaging and muddled multimodel inferences. Ecology 96 Guo, Danlu, Lintern, Anna, Webb, J Angus, Ryu, Dongryeol, Liu, Shuci, Bende-
(9), 2370–2382. Michl, Ulrike, Leahy, Paul, Wilson, Paul, Western, A.W., 2019. Key factors affecting
Carpenter, James, Bithell, John, 2000. Bootstrap confidence intervals: when, which, temporal variability in stream water quality. Water Resour. Res. 55 (1), 112–129.
what? A practical guide for medical statisticians. Stat. Med. 19 (9), 1141–1164. https://doi.org/10.1029/2018WR023370.
Cavelier, Jaime, Jaramillo, Marı ́a, Solis, Daniel, Doris, de León, 1997. Water balance and Guo, Danlu, Lintern, Anna, Webb, J Angus, Ryu, Dongryeol, Bende-Michl, Ulrike,
nutrient inputs in bulk precipitation in tropical montane cloud forest in Panama. Liu, Shuci, Western, Andrew William, 2020. A data-based predictive model for
J. Hydrol. 193 (1–4), 83–96. spatiotemporal variability in stream water quality. Hydrol. Earth Syst. Sci. 24 (2),
Chang, Heejun, 2008. Spatial analysis of water quality trends in the Han River basin, 827–847.
South Korea. Water Res. 42 (13), 3285–3304. Guyon, Isabelle, Elisseeff, André, 2003. An introduction to variable and feature selection.
Chen, Yen-chuan, Ma, Hwong-wen, 2006. Model comparison for risk assessment: a case J. Mach. Learn. Res. 3 (Mar), 1157–1182.
study of contaminated groundwater. Chemosphere 63 (5), 751–761. Hafizan, Juahir, Zain, Sharifuddin M., Yusoff, Mohd Kamil, Hanidza, TI Tengku,
Chen, Hua, Guo, Shenglian, Chong-yu, Xu, Singh, Vijay P., 2007. Historical temporal Armi, AS Mohd, Toriman, Mohd Ekhwan, Mokhtar, Mazlin, 2011. Spatial water
trends of hydro-climatic variables and runoff response to climate variability and quality assessment of Langat River Basin (Malaysia) using environmetric techniques.
their relevance in water resource management in the Hanjiang basin. J. Hydrol. 344 Environ. Monit. Assess. 173 (1–4), 625–641.
(3–4), 171–184. Hatfield, Jerry L., Boote, Kenneth J., Kimball, Bruce A., Ziska, L.H., Izaurralde, Roberto
Clark, Andrew, Tindall, Dan, Healy, Al, 2015. Riparian Vegetation Levels in the C., Donald Thomson, Ort, Allison, M., Wolfe, D., 2011. Climate impacts on
Queensland Murray-Darling Basin and Bulloo Catchments for 2013. Brisbane. htt agriculture: implications for crop production. Agron. J. 103 (2), 351–370.
ps://trove.nla.gov.au/version/244291037. Hattanji, Tsuyoshi, Onda, Yuichi, 2004. Coupling of runoff processes and sediment
Connolly, N.M., Pearson, R.G., Loong, D., Maughan, M., Brodie, J., 2015. Water quality transport in mountainous watersheds underlain by different sedimentary rocks.
variation along streams with similar agricultural development but contrasting Hydrol. Process. 18 (4), 623–636.
riparian vegetation. Agric. Ecosyst. Environ. 213, 11–20. Haynes, David, Müller, Jochen, Carter, Steve, 2000. Pesticide and herbicide residues in
Cooke, Sandra E., Ahmed, Said M., MacAlpine, Neil, 2000. Introductory guide to surface sediments and seagrasses from the great barrier reef world heritage area and
water quality monitoring in agriculture: conservation and development branch, Queensland coast. Mar. Pollut. Bull. 41 (7–12), 279–287.
alberta agriculture, food and rural …. Hinne, Max, Gronau, Quentin F., van den Bergh, Don, Wagenmakers, Eric-Jan, 2020.
Baihua, Fu, Wendy, Merritt, Croke, S., Barry, F.W., Weber, Tony, Jakeman, Anthony J., A conceptual introduction to bayesian model averaging. Advances in Methods and
2019. A review of catchment-scale water quality and erosion models and a synthesis Practices in Psychological Science 3 (2), 200–215. https://doi.org/10.1177/
of future prospects. Environ. Model. Software 114, 75–97. 2515245919898657.
Damodharan, Usha, Reddy, M Vikram, 2012. Impact of sugar industrial treated effluent Hiscock, Kevin M., Grischek, Thomas, 2002. Attenuation of groundwater pollution by
on the growth factor in sugarcane–cuddalore, India. J. Sustain. Bioenergy Syst. 2 (3), bank filtration. J. Hydrol. 266 (3–4), 139–144.
43. Holtan, H., Kamp-Nielsen, L., Stuanes, A.O., 1988. Phosphorus in Soil, Water and
Daoud, Jamal I., 2017. Multicollinearity and Regression Analysis. Paper Presented at the Sediment: an Overview Phosphorus In Freshwater Ecosystems (Pp. 19–34. Springer.
Journal of Physics: Conference Series. Houser, Jeffrey N., Richardson, William B., 2010. Nitrogen and phosphorus in the Upper
Davis, Aaron M., Tink, Michelle, Rohde, Ken, Brodie, Jon E., 2016. Urea contributions to Mississippi River: transport, processing, and effects on the river ecosystem.
dissolved ‘organic’nitrogen losses from intensive, fertilised agriculture. Agric. Hydrobiologia 640 (1), 71–88.
Ecosyst. Environ. 223, 190–196. Huang, X.P., Huang, L.M., Yue, W.Z., 2003. The characteristics of nutrients and
Davis, Aaron M., Pearson, Richard G., Brodie, Jon E., Butler, Barry, 2017. Review and eutrophication in the Pearl River estuary, South China. Mar. Pollut. Bull. 47 (1–6),
conceptual models of agricultural impacts and water quality in waterways of the 30–36.
Great Barrier Reef catchment area. Mar. Freshw. Res. 68 (1), 1–19. Huggins, R., Wallace, Rohan Orr, David, N., Smith, Rachael A., Taylor, O., King, Olivia
de Arruda-Santos, Helen, Roxanny, Schettini, Augusto França, Carlos, Yogui, C., Gardiner, Richard, Wallace, S., Ferguson, Ben, Preston, S., 2018. Total suspended
Takeshi, Gilvan, Maciel, Claudino, Daniele, Zanardi-Lamardo, Eliete, 2018. Sources solids, nutrient and pesticide loads (2015–2016) for rivers that discharge to the great
and distribution of aromatic hydrocarbons in a tropical marine protected area barrier reef–great barrier reef catchment loads monitoring program.
estuary under influence of sugarcane cultivation. Sci. Total Environ. 624, 935–944. Hunter, Heather M., Walton, Richard S., 2008. Land-use effects on fluxes of suspended
Deb, Proloy, Mukand S, Babel, Denis, Anjelo Francis, 2018. Multi-GCMs approach for sediment, nitrogen and phosphorus from a river catchment of the Great Barrier Reef,
assessing climate change impact on water resources in Thailand. Modeling Earth Australia. J. Hydrol. 356 (1–2), 131–146.
Systems and Environment 4 (2), 825–839. Hurvich, Clifford M., Tsai, Chih-Ling, 1989. Regression and time series model selection in
DNRME, 2018. 07/10/2016). Water Monitoring Information Portal. From. https://wat small samples. Biometrika 76 (2), 297–307.
er-monitoring.information.qld.gov.au/. Jiang, Penghui, Cheng, Liang, Li, Manchun, Zhao, Ruifeng, Duan, Yuewei, 2015. Impacts
Donohue, Ian McGarrigle, Martin, L., Mills, Paul, 2006. Linking catchment of LUCC on soil properties in the riparian zones of desert oasis with remote sensing
characteristics and water chemistry with the ecological status of Irish rivers. Water data: a case study of the middle Heihe River basin, China. Sci. Total Environ. 506,
Res. 40 (1), 91–98. 259–271.
13
Johnson, Lucinda, Richards, Carl, Host, George, Arthur, John, 1997. Landscape Meybeck, Michel, Moatar, Florentina, 2012. Daily variability of river concentrations and
influences on water chemistry in Midwestern stream ecosystems. Freshw. Biol. 37 fluxes: indicators based on the segmentation of the rating curve. Hydrol. Process. 26
(1), 193–208. (8), 1188–1207.
Kingsford, Richard T., Walker, Keith F., Lester, Rebecca E., Young, William J., Meynendonckx, J., Heuvelmans, G., Muys, Bart, Feyen, Jan, 2006. Effects of watershed
Fairweather, Peter G., Sammut, Jesmond, Geddes, Michael C., 2011. A ramsar and riparian zone characteristics on nutrient concentrations in the River Scheldt
wetland in crisis–the coorong, lower lakes and Murray mouth, Australia. Mar. Basin. Hydrol. Earth Syst. Sci. Discuss. 3 (3), 653–679.
Freshw. Res. 62 (3), 255–265. Ming, Ye, Meyer, Philip D., Neuman, Shlomo P., 2008. On model selection criteria in
Kleinman, Peter JA., Sharpley, Andrew N., Veith, Tamie L., Maguire, Rory O., multimodel analysis. Water Resour. Res. 44 (3).
Vadas, Peter A., 2004. Evaluation of phosphorus transport in surface runoff from Mitchell, A., Reghenzani, J., Faithful, J., Furnas, M., Brodie, J., 2009. Relationships
packed soil boxes. J. Environ. Qual. 33 (4), 1413–1423. between land use and nutrient concentrations in streams draining a ‘wet-tropics’
Kratz, Timothy, Webster, Katherine, Bowser, Carl, Maguson, John, Benson, Barbara, catchment in northern Australia. Mar. Freshw. Res. 60 (11), 1097–1108.
1997. The influence of landscape position on lakes in northern Wisconsin. Freshw. Mohan, Chinchu, Western Andrew, W., Wei, Yongping, Saft, Margarita, 2018. Predicting
Biol. 37 (1), 209–217. groundwater recharge for varying land cover and climate conditions–a global meta-
Kronvang, Brian, Laubel, Anker, Grant, Ruth, 1997. Suspended sediment and particulate study. Hydrol. Earth Syst. Sci. 22 (5), 2689–2703.
phosphorus transport and delivery pathways in an arable catchment, Gelbaek Mosley, Luke M., Zammit, Benjamin, Leyden, Emily, Heneker, Theresa M.,
stream. Denmark. Hydrological Processes 11 (6), 627–642. Hipsey, Matthew R., Skinner, Dominic, Aldridge, Kane T., 2012. The impact of
Kroon, Frederieke J., Thorburn, Peter, Schaffelke, Britta, Whitten, Stuart, 2016. Towards extreme low flows on the water quality of the Lower Murray River and Lakes (South
protecting the Great Barrier Reef from land-based pollution. Global Change Biol. 22 Australia). Water Resour. Manag. 26 (13), 3923–3946.
(6), 1985–2002. Nakagawa, Shinichi, Freckleton, Robert P., 2011. Model averaging, missing data and
Kruschke, John, 2014. Doing Bayesian Data Analysis: A Tutorial with R, JAGS, and Stan. multiple imputation: a case study for behavioural ecology. Behav. Ecol. Sociobiol. 65
Academic Press. (1), 103–116.
Kuhnert, Petra, Wang, You-Gan, Henderson, Brent, Stewart, Lachlan, Wilkinson, Scott, Nash, Maliha S., Chaloud, Deborah J., 2011. Partial least square analyses of landscape
2009. Statistical methods for the estimation of pollutant loads from monitoring data. and surface water biota associations in the Savannah River Basin. ISRN Ecology
Final Project Report. Report to the Marine and Tropical Sciences Research Facility, 2011.
Reef and Rainforest Research Centre Limited, Cairns. Nash, J Eamonn, Sutcliffe, Jonh V., 1970. River flow forecasting through conceptual
Kuhnert, Petra M., Henderson, Brent L., Lewis, Stephen E., Bainbridge, Zoe T., models part I—a discussion of principles. J. Hydrol. 10 (3), 282–290.
Wilkinson, Scott N., Brodie, Jon E., 2012. Quantifying total suspended sediment Nilsson, Christer, Malm-Renöfält, Birgitta, 2008. Linking flow regime and water quality
export from the Burdekin River catchment using the loads regression estimator tool. in rivers: a challenge to adaptive catchment management. Ecol. Soc. 13 (2), 18.
Water Resour. Res. 48 (4). Noe, Gregory B., Cashman, Matthew J., Skalak, Katie, Gellis, Allen, Hopkins, Kristina G.,
Kundzewicz, Zbigniew W., Mata, Luis Jose, Arnell, N.W., Doll, Petra, Kabat, Pavel, Moyer, Doug, Webber, James, Benthem, Adam, Maloney, Kelly, Brakebill, John,
Jimenez, Blanca, Miller, Kathleen, Taikan, Oki, Zekai, S., Shiklomanov, Igor, 2007. 2020. Sediment dynamics and implications for management: state of the science
Freshwater resources and their management. from long-term research in the Chesapeake Bay watershed, USA. Wiley
Leanne, Elchyshyn, Jean-Olivier, Goyette, Talbot, Saulnier, Maranger, Émilie, Interdisciplinary Reviews: Water 7 (4), e1454.
Nozais, Roxane, Solomon, Christian, Christopher, T., Gregory-Eaves, Irene, 2018. Orr, D., Turner, R.D.R., Huggins, R., Vardy, S., Warne, J., St, M., 2014. Wet Tropics water
Quantifying the effects of hydrological changes on long-term water quality trends in quality statistics for high and base flow conditions. Brisbane.
temperate reservoirs: insights from a multi-scale, paleolimnological study. Ouyang, Y., Nkedi-Kizza, P., Wu, Q.T., Shinde, D., Huang, C.H., 2006. Assessment of
J. Paleolimnol. 60 (3), 361–379. seasonal variations in surface water quality. Water Res. 40 (20), 3800–3810.
Lewis, Stephen E., Brodie, Jon E., Bainbridge, Zoë T., Rohde, Ken W., Davis, Aaron M., Parrish, Mark A., Moradkhani, Hamid, DeChant, Caleb M., 2012. Toward reduction of
Masters, Bronwyn L., Maughan, Mirjam, Devlin, Michelle J., Mueller, Jochen F., model uncertainty: integration of Bayesian model averaging and data assimilation.
Schaffelke, Britta, 2009. Herbicides: a new threat to the great barrier reef. Environ. Water Resour. Res. 48 (3).
Pollut. 157 (8–9), 2470–2484. Pelletier, Jon D., Baker, Victor R., 2011. The role of weathering in the formation of
Lintern, A., Webb, J.A., Ryu, D., Liu, S., Bende-Michl, U., Waters, D., Leahy, P., bedrock valleys on Earth and Mars: a numerical modeling investigation. J. Geophys.
Wilson, P., Western, A.W., 2018a. Key factors influencing differences in stream water Res.: Plan 116 (E11).
quality across space. Wiley Interdisciplinary Reviews: Water 5 (1), e1260. Perona, E., Bonilla, I., Mateo, P., 1999. Spatial and temporal changes in water quality in a
Lintern, A., Webb, J.A., Ryu, D., Liu, S., Waters, D., Leahy, P., Bende-Michl, U., Spanish river. Sci. Total Environ. 241 (1–3), 75–90.
Western, A.W., 2018b. What are the key catchment characteristics affecting spatial Piazza, Antonio, Gustavo, , Dupas, Rémi, Gascuel-Odoux, Chantal, Grimaldi, Catherine,
differences in riverine water quality? Water Resour. Res. 54 (10), 7252–7272. Pinheiro, Adilson, Kaufmann, Vander, 2018. Influence of hydroclimatic variations on
Lintern, Anna, McPhillips, Lauren, Winfrey, Brandon, Duncan, Jonathan, Grady, Caitlin, solute concentration dynamics in nested subtropical catchments with heterogeneous
2020. Best management practices for diffuse nutrient pollution: wicked problems landscapes. Sci. Total Environ. 635, 1091–1101.
across urban and agricultural watersheds. Environ. Sci. Technol. 54 (15), Pickering, A.D., Pottinger, T.G., 1987. Poor water quality suppresses the cortisol
9159–9174. response of salmonid fish to handling and confinement. J. Fish. Biol. 30 (3),
Liu, Shuci, Ryu, Dongryeol, Webb, J. Anugs, Lintern, Anna, Guo, Danlu, Waters, David, 363–374.
Western, Andrew W., 2021. A Bayesian approach to understanding the key factors Pionke, H.B., Gburek, W.J., Schnabel, R.R., Sharpley, A.N., Elwinger, G.F., 1999.
influencing temporal variability in stream water quality: a case study in the Great Seasonal flow, nutrient concentrations and loading patterns in stream flow draining
Barrier Reef catchments. Hydrol. Earth Syst. Sci. Discuss. https://doi.org/10.5194/ an agricultural hill-land watershed. J. Hydrol. 220 (1–2), 62–73.
hess-2020-681. In press. Poeter, Eileen, Anderson, David, 2005. Multimodel ranking and inference in ground
Liu, S., Ryu, D., Webb, J.A., Lintern, A., Waters, D., Guo, Danlu, Western, A.W., 2018. water modeling. Groundwater 43 (4), 597–605.
Characterisation of spatial variability in water quality in the Great Barrier Reef Posch, Konstantin, Arbeiter, Maximilian, Pilz, Juergen, 2020. A novel Bayesian approach
catchments using multivariate statistical analysis. Mar. Pollut. Bull. 137, 137–151. for variable selection in linear regression models. Comput. Stat. Data Anal. 144,
https://doi.org/10.1016/j.marpolbul.2018.10.019. 106881.
Lukacs, Paul M., Burnham, Kenneth P., Anderson, David R., 2010. Model selection bias Prasad, V Krishna, Ortiz, Ariel, Stinner, Ben, Mccartney, David, Parker, Jason,
and Freedman’s paradox. Ann. Inst. Stat. Math. 62 (1), 117. Hudgins, Deana, Hoy, Casey, Moore, Richard, 2005. Exploring the relationship
Maciel, Claudino, Daniele, de Souza, Roberto Botelho, José, Taniguchi, Satie, between hydrologic parameters and nutrient loads using digital elevation model and
Bícego, Márcia Caruso, Zanardi-Lamardo, Eliete, 2015. Sources and distribution of GIS–a case study from Sugarcreek headwaters, Ohio, USA. Environ. Monit. Assess.
polycyclic aromatic hydrocarbons in a an urbanized tropical estuary and adjacent 110 (1–3), 141–169.
shelf, Northeast of Brazil. Mar. Pollut. Bull. 101 (1), 429–433. Pratchett, Morgan S., Bridge, Tom CL., Brodie, Jon, Cameron, Darren S., Day, Jon C.,
Marianna, Joo, Myriam Aa, Raymond, McNeil, Vivienne H., Raethea, Huggins, Emslie, Michael J., Grech, Alana, Hamann, Mark, Heron, Scott F., Hoey, Andrew S.,
Turner, Ryan DR., Choy, Satish, 2012. Estimates of sediment and nutrient loads in 10 2019. Australia’s Great Barrier Reef World Seas: an Environmental Evaluation (Pp.
major catchments draining to the Great Barrier Reef during 2006–2009. Mar. Pollut. 333–362. Elsevier.
Bull. 65 (4–9), 150–166. Prathumratana, Lunchakorn, Sthiannopkao Suthipong, Kim, Kyoung Woong, 2008. The
Matlab and Statistics Toolbox, 2017. The MathWorks, Inc., natick, Massachusetts, United relationship of climatic and hydrological parameters to surface water quality in the
States. lower Mekong River. Environ. Int. 34 (6), 860–866.
May, Robert, Dandy, Graeme, Maier, Holger, 2011. Review of input variable selection Preston, Stephen D., Brakebill, John W., 1999. Application of spatially referenced
methods for artificial neural networks. Artificial neural networks-methodological regression modeling for the evaluation of total nitrogen loading in the Chesapeake
advances and biomedical applications (InTech). Bay watershed: USGS.
McCloskey, G., Waters, D., Baheerathan, R., Darr, S., Dougall, C., Ellis, R., Fentie, B., Pufahl, Peir K., Groat, Lee A., 2017. Sedimentary and igneous phosphate deposits:
Hateley, L., 2017. Modelling reductions of pollutant loads due to improved formation and exploration: an invited paper. Econ. Geol. 112 (3), 483–516.
management practices in the great barrier reef catchments: updated methodology Queensland Government, 2016. Great Barrier Reef Report Card 2016. Brisbane:
and results-technical report for reef report card 2015. Queensland Department of Retrieved from. https://www.reefplan.qld.gov.au/tracking-progress/reef-report-
Natural Resources and Mines, Brisbane, Queensland. card/2016.
McKergow, Lucy A., Weaver, David M., Prosser, Ian P., Grayson, Rodger B., Reed, Adrian Queensland Government, 2017. The Queensland Land Use Mapping Program (QLUMP).
EG., 2003. Before and after riparian management: sediment and nutrient exports From Department of Science, Information Technology and Innovation. https://www.
from a small agricultural catchment, Western Australia. J. Hydrol. 270 (3–4), qld.gov.au/environment/land/vegetation/mapping/qlump.
253–272. Raftery, Adrian E., Gneiting, Tilmann, Balabdaoui, Fadoua, Polakowski, Michael, 2005.
Using Bayesian model averaging to calibrate forecast ensembles. Mon. Weather Rev.
133 (5), 1155–1174.
14
Richards, R Peter, Baker, David B., 2002. Trends in water quality in LEASEQ rivers and Valck, De, Jeremy, Rolfe, John, 2018. Linking water quality impacts and benefits of
streams (Northwestern Ohio), 1975–1995. J. Environ. Qual. 31 (1), 90–96. ecosystem services in the Great Barrier Reef. Mar. Pollut. Bull. 130, 55–66.
Saft, Margarita, Peel, Murray C., Western, Andrew W., Zhang, Lu, 2016. Predicting shifts Varanka, Sanna, Hjort, Jan, Luoto, Miska, 2015. Geomorphological factors predict water
in rainfall-runoff partitioning during multiyear drought: roles of dry period and quality in boreal rivers. Earth Surf. Process. Landforms 40 (15), 1989–1999.
catchment characteristics. Water Resour. Res. 52 (12), 9290–9305. Vatcheva, Kristina P., Lee, MinJae, McCormick, Joseph B., Rahbar, Mohammad H., 2016.
Sangani, Hasani, Mohammad, Amiri, Jabbarian, Bahman, Shabani, Alizadeh, Afshin, Multicollinearity in regression analyses conducted in epidemiologic studies.
Sakieh Yousef, Ashrafi, Sohrab, 2015. Modeling relationships between catchment Epidemiology 6 (2).
attributes and river water quality in southern catchments of the Caspian Sea. Walker, Jeffrey A., 2019. Model-averaged regression coefficients have a straightforward
Environ. Sci. Pollut. Control Ser. 22 (7), 4985–5002. interpretation using causal conditioning. BioRxiv 133785. https://doi.org/10.1101/
Sardans, Jordi, Peñuelas, Josep, Estiarte, Marc, 2008. Changes in soil enzymes related to 133785.
C and N cycle and in soil C and N content under prolonged warming and drought in a Waterhouse, J., Schaffelke, B., Bartley, R., Eberhadr, R., Brodie, J., Star, M.,
Mediterranean shrubland. Appl. Soil Ecol. 39 (2), 223–235. Thorburn, P., Rolfe, J., Ronan, M., Taylor, B., Kroon, F., 2017. 2017 Scientific
Schaffelke, Britta, Carleton, John, Michele, Skuza, Irena, Zagorskis, Furnas, Miles J., Consensus Statement: A Synthesis of the Science of Land-Based Water Quality
2012. Water quality in the inshore Great Barrier Reef lagoon: implications for long- Impacts on the Great Barrier Reef. Brisbane.
term monitoring and management. Mar. Pollut. Bull. 65 (4–9), 249–260. Waters, D., Packett, R., 2007. Sediment and Nutrient Generation Rates for Queensland
Shapiro, Sanford, Samuel, Wilk, Martin B., 1965. An analysis of variance test for Rural Catchments-An Event Monitoring Program to Improve Water Quality
normality (complete samples). Biometrika 52, 591–611, 3/4. Modelling. Paper Presented at the Proceedings of the 5th Australian Stream
Shaw, M., Silburn, D.M., 2014. Paddock to Reef integrated monitoring, modelling and Management Conference. Australian Rivers: Making a Difference. Charles Sturt
reporting program, Paddock scale modelling technical report. Brisbane. University, Thurgoona, New South Wales.
Sherriff, Sophie C., Rowan, John S., Fenton, Owen, Jordan, Philip, Melland, Alice R., Waters, D.K., Carroll, C., Ellis, R., Hateley, L., McCloskey, J., Packett, R., Dougall, C.,
Mellander, Per-Erik, Huallachain, Daire O., 2016. Storm event suspended sediment- Fentie, B., 2013. Modelling reductions of pollutant loads due to improved
discharge hysteresis and controls in agricultural watersheds: implications for management practices in the great barrier reef catchments-whole of GBR, volume 1
watershed scale sediment management. Environ. Sci. Technol. 50 (4), 1769–1778. department of natural resources and Mines: technical report (ISBN: 978-1-7423-
Skoulikidis, N Th, Amaxidis, Y., Bertahas, I., Laschou, S., Gritzalis, K., 2006. Analysis of 0999).
factors driving stream water composition and synthesis of management tools—a case Whitten, S., Bennett, J., 2004. Economics for natural resources management:
study on small/medium Greek catchments. Sci. Total Environ. 362 (1–3), 205–241. bioeconomic modeling, policy threshold analysis and transaction costs. In: Paper
https://10.1016/j.scitotenv.2005.05.018. Presented at the Sixth Annual BioEcon Conference. Kings College, Cambridge.
Smith, Andrew P., Western, Andrew W., Hannah, Murray C., 2013. Linking water quality Whittingham, Mark J., Stephens, Philip A., Bradbury, Richard B., Freckleton, Robert P.,
trends with land use intensification in dairy farming catchments. J. Hydrol. 476, 2006. Why do we still use stepwise modelling in ecology and behaviour? J. Anim.
1–12. Ecol. 75 (5), 1182–1189.
Soranno, P.A., Hubler, S.L., Carpenter, S.R., Lathrop, R.C., 1996. Phosphorus loads to Woodcock, Curtis E., Collins, John B., Gopal, Sucharita, Jakabhazy, Vida D., Li, Xiaowen,
surface waters: a simple model to account for spatial pattern of land use. Ecol. Appl. Macomber, Scott, Ryherd, Soren, Harward, V Judson, Levitan, Jack, Wu, Yecheng,
6 (3), 865–878. 1994. Mapping forest vegetation using Landsat TM imagery and a canopy reflectance
Star, Megan, Rolfe, John, McCosker, Kevin, Smith, Rachael, Ellis, Robin, Waters, David, model. Rem. Sens. Environ. 50 (3), 240–254.
Waterhouse, Jane, 2018. Targeting for pollutant reductions in the Great Barrier Reef Xiaolong, Wang, Jingyi, Han, Ligang, Xu, Qi, Zhang, 2010. Spatial and seasonal
river catchments. Environ. Sci. Pol. 89, 365–377. variations of the contamination within water body of the Grand Canal, China.
Steinman, Alan, Hassett, Michael, Oudsema, Maggie, 2018. Effectiveness of best Environ. Pollut. 158 (5), 1513–1520.
management practices to reduce phosphorus loading to a highly eutrophic lake. Int. Xu, Guoce, Li, Peng, Lu, Kexin, Tantai, Zhan, Zhang, Jiaxin, Ren, Zongping,
J. Environ. Res. Publ. Health 15 (10), 2111. Wang, Xiukang, Yu, Kunxia, Shi, Peng, Cheng, Yuting, 2019. Seasonal changes in
Stoll, Sebastian, Hendricks Franssen, Harrie-Jan, Michael, Butts, Kinzelbach, Wolfgang water quality and its main influencing factors in the Dan River basin. Catena 173,
KH., 2011. Analysis of the impact of climate change on groundwater related 131–140.
hydrological fluxes: a multi-model approach including different downscaling Ye, Lin, Cai, Qing-hua, Liu, Rui-qiu, Cao, Ming, 2009. The influence of topography and
methods. Hydrol. Earth Syst. Sci. 15 (1), 21–38. land use on water quality of Xiangxi River in Three Gorges Reservoir region.
Strudley, Mark W., Murray, A Brad, Haff, P.K., 2006. Emergence of pediments, tors, and Environ. Geol. 58 (5), 937–942.
piedmont junctions from a bedrock weathering–regolith thickness feedback. Geology Young, William J., Marston, Frances M., Davis, Richard J., 1996. Nutrient exports and
34 (10), 805–808. land use in Australian catchments. J. Environ. Manag. 47 (2), 165–183.
Tang, Weigang, Carey, Sean K., 2017. HydRun: a MATLAB toolbox for rainfall–runoff Zhang, Qian, Blomquist, Joel D., 2018. Watershed export of fine sediment, organic
analysis. Hydrol. Process. 31 (15), 2670–2682. carbon, and chlorophyll-a to Chesapeake Bay: spatial and temporal patterns in
Teixeira, Zara, Teixeira, Heliana, Marques, João C., 2014. Systematic processes of land 1984–2016. Sci. Total Environ. 619, 1066–1078. https://doi.org/10.1016/j.
use/land cover change to identify relevant driving forces: implications on water scitotenv.2017.10.279.
quality. Sci. Total Environ. 470, 1320–1335. Zhang, Zhaoyong, Juying, Li, Mamat, Zulpiya, QingFu, Ye, 2016. Sources identification
Thorburn, P.J., Wilkinson, S.N., Silburn, D.M., 2013. Water quality in agricultural lands and pollution evaluation of heavy metals in the surface sediments of Bortala River,
draining to the Great Barrier Reef: a review of causes, management and priorities. Northwest China. Ecotoxicol. Environ. Saf. 126, 94–101.
Agric. Ecosyst. Environ. 180, 4–20. Zhao, C.S., Yang, Y., Yang, S.T., Xiang, H., Wang, F., Chen, X., Zhang, H.M., Yu, Q., 2019.
Thorburn, Peter J., Biggs, Jody S., Palmer, Jeda, Meier, Elizabeth A., Verburg, Kirsten, Impact of spatial variations in water quality and hydrological factors on the food-
Skocaj, Danielle M., 2017. Prioritizing crop management to increase nitrogen use web structure in urban aquatic environments. Water Res.
efficiency in Australian sugarcane crops. Front. Plant Sci. 8, 1504. Zhuo, La, Mekonnen, Mesfin M., Hoekstra, Arjen Y., Wada, Yoshihide, 2016. Inter-and
Tianyang, Li, Siyue, Li, Chuan, Liang, Bush, Richard T., Xiong, Lihua, Jiang, Yongjun, intra-annual variation of water footprint of crops and blue water scarcity in the
2018. A comparative assessment of Australia’s Lower Lakes water quality under Yellow River basin (1961–2009). Adv. Water Resour. 87, 29–41.
extreme drought and post-drought conditions using multivariate statistical
techniques. J. Clean. Prod. 190, 1–11.
15

Liu 2021

Uploaded by

Copyright:

Available Formats

Liu 2021

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Liu 2021

Uploaded by

Copyright:

Available Formats

Environmental Pollution 288 (2021) 117337

Contents lists available at ScienceDirect

A multi-model approach to assessing the impacts of catchment

1. Introduction Schaffelke et al., 2012; Waterhouse et al., 2017).

Fig. 1. The Great Barrier Reef and GBR catchments,

Environmental Pollution 288 (2021) 117337

et al., 2005). y= wi y i (3)

dividual explanatory variables by summing wi for each model in which

number of potential predictors allowed us to explore more possible βj = wi βi,j (4)

Table 2 3.2.1. Key factors identified in plausible models

You might also like