2. Methods

2.1: Question 1: Assessment of the methods used by the regions to monitor the quality of freshwater

We obtained information (metadata) from the four regional councils (Waikato, Taranaki, Horizons and Southland) that described their State of Environment (SoE) water quality monitoring programs for rivers, lakes and groundwater including; the locations and the details of sites in their networks, the frequency of monitoring, the variables analysed and the QA/QC and data storage procedures. From this information we reviewed the following aspects of their monitoring programmes:

1. Network design. We considered the design of the regional council networks including; the number of sites, where (what types of environments) they cover, sampling interval (frequency) as it relates to future use of the data for trend analysis and load calculations. A key question we addressed was the extent to which the major freshwater resources (rivers, lakes and groundwater) were represented in the SoE programmes. As part of this, for each region, we assessed the extent to which each network of sites represented the environmental variation in freshwater resources. An allied but less easily answered question concerns the overall adequacy of the network. A single definitive test of the adequacy of networks is hard to justify because all the potential uses of SoE data cannot be known. However, a reasonable test is whether the distribution of monitoring sites across a region's water bodies is sufficient to establish general patterns in both state and trends in statistically robust way. We responded to this question for each regions rivers, for which we were analysing state and trends, by grouping both state and trend data by environmentally defined groups. The details of these tests are discussed in section 2.2.5 below.
2. Water quality variables. We considered the measured variables and analysis methods. This included review of methods used for sample collection, preservation and stabilisation (for samples analysed in a laboratory at a later date) and analysis. We assessed detection limits (sensitivity of the analytical methods) and significant figures in reported data. Detection limits can be particularly important when monitoring water bodies that have high water quality status and there is an expectation that the time series will be used to perform trend analysis.
3. Flow data. When monitoring river water quality, it is important to have flow measurements accompanying each water quality measurement as many water quality variables are subject to either dilution (decreasing concentration with increasing flow, e.g., conductivity) or land runoff (increasing concentration with increasing flow, e.g., total phosphorus). Data can be flow adjusted before trend analysis, to remove the effects of changes in flow on water quality variable concentrations. Because changes in flow are tied to natural changes in precipitation and evapotranspiration, flow adjustment of water quality variable concentrations allows trends caused by other, largely anthropogenic, changes to be more directly assessed. Without a proper consideration of flow-dependency (of any given variable) it is difficult to decide if concentration increases are a result of more rainfall or increased land loadings. Furthermore, flow data enables accurate calculation of loads and specific yields to characterise land use change in a given catchment.
4. Microbial variables. We reviewed the microbial variables that are included in the SoE programmes. Microbial variables provide measures of the risk of infection from waterborne pathogens and may include the faecal indicator bacterium Escherichia coli (E. coli), or the whole Faecal coliforms (FC) group (that includes E. coli , but also other "coliforms" such as those of the genus Klebsiella) that are found in the gut of warm-blooded animals.
5. Biological monitoring. We reviewed the biological variables (viz. macroinvertebrates, cyanobacteria and periphyton cover) that are included in the SoE programmes. Surveys of biological variables complement water quality monitoring by providing measures of ecosystem health and habitat condition (invertebrates), nuisance growths of plants (periphyton) or potential health risks (cyanobacteria), and by integrating water condition over time.
6. QA/QC methods. We considered the quality assurance and control (QA/QC) methods and data storage and QA/QC procedures.

2.2: Question 2: Analysis of state and trends in river water quality

We analysed state and trends in river water quality data because state of environment monitoring of freshwaters is most comprehensively and consistently carried out on this type of water body (in terms of the time period for which monitoring has been carried out, sample frequency, variables analysed and intensity of sampling). This study did not analyse state and trends in lake or groundwater quality data. Lake data is collected in a less consistent manner across the regions due to differences in the distribution of lakes (e.g., Taranaki and Horizons have few iconic lakes) and because of differences in the intensity of lake management issues. There are also differences in how groundwater is monitored across the regional councils reflecting differing regional focuses of the groundwater programmes. Most groundwater monitoring programmes indicate stable water chemistry other than for nitrate, which is usually monitored in separate (non-SoE) programmes. There have been recent national studies of the state and trends of lakes (Verburg et al. 2010) across New Zealand for the ten year period up to and including 2009 and groundwater (Daughney et al. 2009) up to and including 2008. In addition, we did not analyse state and trends in biological variables (e.g., macro-invertebrates and periphyton data). Again, this was because of differences in the time period for which biological monitoring has been carried out, sampling frequencies and range of variables analysed by the four regions. These differences reflect differing regional focuses for their biological monitoring programmes.

2.2.1: Obtaining and formatting river water quality data

All New Zealand regional councils maintain extensive water quality databases, which are frequently used by MfE and other agencies (including NIWA) for specific research projects (e.g., Ballantine et al., 2010). When discussing data requirements for the current project with OAG, it was decided to compile all state-of-environment water quality monitoring data collected by each of the four regional councils for the 10 year period up to the end of 2009. To this we also added the National River Water Quality Network (NRWQN) water quality monitoring data for sites in the four regions. The NRWQN is a national network of river water quality monitoring sites that is operated by NIWA. However, these data are often used by councils to augment their own SoE data and reports.

The data sets used for this study provided records of commonly measured water quality variables (Table 1) at a range of sites over time, but varied widely in reporting formats, reporting conventions, variable names, units of measurement, and sampling frequency. For example, reporting formats ranged from a single Excel sheet with all variables for all sites stored in a single column, to multiple workbooks for individual sites with data for each site distributed over multiple worksheets with each variable stored in a separate column. Electrical conductivity was provided as a field measurement (labelled "Conductivity" or some near equivalent), as a laboratory measurement (typically labelled EC25, i.e., conductivity at 25°C), and sometimes as both within a single region. Units of measurement (most notably for conductivity) varied between regions, and (less commonly) for a single variable within a region. To consolidate these data into a uniform structure and minimise the potential for error associated with manually copying data between worksheets, we used a modified version of a MS-Access database developed for a previous MfE water quality review (Ballantine, et al., 2010). When retrieving data for subsequent analyses, we adopted the following conventions:

1. field conductivity (COND) was used where available, otherwise EC25 (which was highly correlated (r2 = 0.85) with COND for sites where both variables were reported) was used as a surrogate.
2. variables marked as below a specified detection limit were recoded as half the detection limit. For variables marked as above a specified level (e.g., E. coli > 20 000), we used the numerical value as given.
3. total nitrogen (TN) for regions which did not specifically report this variable was calculated (where possible) as the sum of Nitrate+Nitrite Nitrogen (NNN) plus Total Kjeldahl Nitrogen (TKN).
4. Sites in estuarine waters were flagged so as to avoid skewing data for variables (such as conductivity) which are likely to be highly elevated in such environments.

Data associated with each site included:

1. site name
2. location and regional council identifier (if available)
3. NZMS260 grid reference (converted from NZTM as appropriate)
4. reach number (NZ Reach) as defined in the River Environment Classification (REC; see Section 2.2.2) scheme (Snelder and Biggs, 2002).

All sites were then assigned a unique identifier based on the corresponding regional council name and site identifier. All analyses were derived from queries of this database, which produced water quality data for the 11 variables described in Table 1 in consistent units.

Table 1:

Water quality variables included in this study

Within the regions, over the duration of the sampling, water quality analytical methods have changed. One example of this is field conductivity and lab conductivity at 25oC. Some regional councils previously used one method but, during the sampling period, changed to another method. In such cases, we combined the data that was analysed using different methods to provide a continuous record. In the case of field conductivity and lab conductivity, this was justifiable because the two methods produce data that are strongly correlated (r2 = 0.85).

The resulting data set contained some gaps in temporal coverage corresponding to missed sampling occasions, mixed (quarterly and monthly) sampling by individual councils and the discontinuation or commencement of sites during the period. Trend analysis is only robust if calculated using a data set with few missing values and must be data collected consistently on either a quarterly (i.e. seasonal) or monthly basis. Not all data sets provided by the regional councils were sufficiently complete to provide robust trend analyses for the 10-year period of our trend analysis. For example some sampling occasions (either months or seasons) were missed for many sites (Figure 1). In addition, Horizons historically employed a system of "rolling SoE sites" whereby sites were monitored discontinuously, for example once every three years a years worth of monthly sampling may be undertaken (Figure 1). This strategy increases spatial coverage, but means that data cannot be robustly analysed for trends. To ensure our trend analysis was robust, we limited our analysis to data sets for which at least 80% of sample occasions had data. Thus, for sites that were monitored quarterly, we included sites that had data for 32 quarters of 40 possible quarters. For sites that were monitored monthly we included sites that had data for 96 of 120 possible months.

2.3: Water Quality State

We used the median concentration of all observations and for each water quality variable over the entire time period to describe the water quality state of each site that met our criteria for trend analysis. To place these values in context they have been compared with guidelines and 'trigger values' (Table 2). The median nutrient concentrations have been compared with the New Zealand trigger values for the protection of aquatic ecosystems from the Australian and New Zealand Environment Conservation Council (ANZECC) guidelines (ANZECC, 2000). The trigger values are not national standards but rather, have been devised to assess the levels of physical and chemical stressors which might have ecological or biological effects. Rather than implying that there will be ecological and biological effects caused by increased levels of physical and chemical stressors, exceedances of trigger levels indicate cause for further investigation of water quality issues. Conversely, where trigger levels are not exceeded we can have reasonable confidence that water quality is sufficient to support the ecological values. We compared the median clarity measurements to the MFE (1994) water quality guidelines for clarity.

Figure 1:

Typical sample calendars (years on the horizontal axis and months on the vertical axis) showing when data was present for specific variables at three sites. Gaps in temporal coverage are white and sample occasions with data are grey. The upper calendar shows months when E.coli data was available for an Environment Waikato site that has been sampled quarterly (seasonally). The middle calendar shows months when NH4-N data is available for a Horizons Regional Council site that has been repeatedly discontinued and re-commenced (a rolling site). The lower calendar shows months when DRP data is available for a Horizons Regional Council site that is sampled monthly but for which some dates have been missed.

We compared the 95th percentile values for E. coli with the microbiological water quality guidelines for recreational use (MfE and MoH, 2003), which are based on the 95th percentile value for E. coli. Finally we nominated a guideline of 148/100 mL for Faecal Coliforms (FC) based on the ANZECC (2000) guideline for E.coli of 126 /100ml. Because E. coli makes up approximately 85% of all faecal coliforms the guideline represents a FC guideline of 148 /100ml.

Table 2:

ANZECC trigger values for nutrients (based on median values), MfE guideline for clarity (based on median values) and MfE/MoH guideline value (95th percentile) for Escherichia coli and modified ANZECC 2000 guidelines for Faecal Coliforms.

To facilitate comparisons, and to provide an insight into the spatial patterns of water quality and the environmental and human factors that determine these, we compared the median values (95th percentile for E. coli) of selected variables for sites for which at least 80% of sample occasions had data, grouped by REC Topography and Land-cover categories.

2.4: Trend analysis

2.4.1: Statistical analysis

The trend assessment was carried out on data for a ten year time period (2000–2009). Trends in water quality variables can be evident when the data are viewed graphically. For example Figure 2 shows time series for TN, TP and DRP collected over the 10-year period at a site in the Southland region. Trends at all sites and variable combinations that met our criteria were formally assessed using the non-parametric Seasonal Kendall Sen Slope Estimator (SKSE, Sen 1968). The SKSE is used to quantify the magnitude and direction of trends in data that are subject to appreciable seasonality such as water quality data. Regional councils commonly use the Time Trends software (http://www.niwa.co.nz/our-science/freshwater/tools/analysis) to estimate SKSE values. We used the same method that that is provided by Time Trends within alternative (bespoke) software because of the number of sites considered which would make trend analysis onerous.

It is important to have flow measurements accompanying each water quality measurement because many water quality analytes are subject to either dilution (decreasing concentration with increasing flow, e.g., conductivity) or wash-off (increasing concentration with increasing flow, e.g., total phosphorus). Data can be flow adjusted before trend analysis, to remove the effects of variation in river flow on water quality analyte concentrations. Because changes in river flow are tied to natural changes in precipitation and evapotranspiration, flow adjustment of water quality analyte concentrations allows trends caused by other, largely anthropogenic, changes to be more directly assessed. Trend analysis was carried out on raw data and on flow adjusted data but only flow adjusted trends are discussed in this report since these are usually the most useful basis for inferring change in water quality.

The flow adjustment procedure was performed using LOWESS⁷ (Locally WEighted Scatterplot Smoothing) with a 30 per cent span. Every data point in the record was adjusted depending on the value of flow as outlined by Smith et al. (1996): adjusted value = raw value – smoothed value + median value (where the "smoothed value" is that predicted from the flow at time of sampling using LOWESS). For cases where flow data were provided for at least 80% of water quality sampling occasions, we used these flow data to flow adjust each variable. In cases where flow data were provided for less than 80% of water quality sampling occasions we used a flow estimation method (Ballantine et al 2010) to estimate flows and therefore perform flow adjustment.

Values of the SKSE were normalised by dividing by the median to give the relative SKSE (RSKSE), allowing for direct comparison between sites measured as per cent change per year. The RSKSE may be thought of as an index of the relative rate of change. A positive RSKSE value indicates an overall increasing trend, while a negative RSKSE value indicates an overall decreasing trend. The SKSE calculations were accompanied by a Seasonal Kendall test (Helsel and Frans, 2006) of the null hypothesis that there is no monotonic trend. If the associated P-value is "small" (i.e. P

Figure 2:

Scatter-plots of Total Nitrogen, Total Phosphorus and Dissolved Reactive Phosphorus data collected over the 10-year period at a site in the Southland region. A smoothed line has been fitted to the data to illustrate a temporally averaged concentration that indicates the longer term trend. When formal trend analyses were performed on these data the variable in the upper plot (Total Nitrogen) had a significant increasing trend, the variable in the middle plot (Total Phosphorus) had middle plot was stable and variable in the lower plot (Dissolved Reactive Phosphorus) had a significant decreasing trend.

< 0.05), the null hypothesis can be rejected (i.e. the observed trend or any larger trend, either upwards or downwards, is most unlikely to have arisen by chance).

2.4.2: Flow estimation methods

Many regional council water quality sampling sites either did not have flow recording stations or did not provide flow measurements for the sampling occasions. Therefore, we used a method for estimating flows that interpolates data from gauging stations in the New Zealand Hydrometric Network (Ballantine et al. 2010). Only flow gauges with five or more years of data and that were free from flow modification due to abstractions and dams were used (n = 264). For each water quality site and each date when water quality had been measured we identified the most appropriate flow gauging station. This gauging station was defined as the geographically closest gauging station (Euclidean distance) that shared the same REC Climate and Topography category as the monitoring site (see Section 2.5.1) and that also had a record of flow at the time of sampling. The flow recorded at the closest flow gauging station was standardised by dividing by mean flow for the entire flow monitoring period. Standardised flows were then converted to units of m³ s^-1 multiplied by the national estimate of mean flow (Woods et al. 2006) associated with each water quality monitoring site.

In a previous study Ballantine et al. (2010) showed that we can have a reasonable level of confidence in the overall findings of water quality trend analyses derived using flows estimated using this method. However, trends for some individual sites have large errors due to uncertainties associated with the flow estimation. Uncertainties associated with these flow estimates reduce the robustness of our trend analysis in comparison to having flow measurements associated with all water quality observations. The implication of this for this study is that the trends for individual sites need to be treated with caution. However, Ballantine et al. (2010) showed that we can be confidant concerning our findings for overall trends (that is trends at the regional scale or by environmentally defined groupings within regions, see Section 2.6).

2.4.3: Categorisation of trends

To provide an interpretation of the trends we categorised them according to their direction and magnitude. Scarsbrook (2006) recognised that statistical significance of a trend does not necessarily imply a 'meaningful' trend, i.e., one that is likely to be relevant in a management context. We followed Scarsbrook (2006) in denoting a 'meaningful' trend as one for which the (statistically significant) RSKSE has an absolute magnitude > 1 per cent year^-1. Scarsbrook (2006) recognised that the choice of 1 per cent year^-1 as the 'meaningful' threshold is arbitrary, but this has the advantage that it corresponds to a magnitude that people are likely to detect within a human lifetime. Therefore, trends were categorised as follows:

1. no significant trend – the null hypothesis for the Seasonal Kendall test was not rejected (i.e., P > 0.05)
2. significant trend – the null hypothesis for the Seasonal Kendall test was rejected (i.e., P < 0.05) but the magnitude of the trend (SKSE) was less than one per cent per annum of the raw data median (i.e., the RSKSE value was less than 1 per cent year^-1). Note that the trend at some sites may be 'significant but not meaningful'
3. 'meaningful' trend – the null hypothesis for the Seasonal Kendall test was rejected (i.e., P < 0.05) and the magnitude of the trend (SKSE) was greater than one per cent per annum of the raw data median (i.e., the RSKSE value was greater than 1 per cent year^-1 or about 10% per decade)

2.5: Ranking of sites within regions

To help identify locations or catchments of management concern within each region or risk we ranked the sites based on an index derived from the state and trends analysis. The ranking for a site is made by summing scores that are assigned to state and trends for each variable. Variables that fail guidelines were assigned a score of 1 and that pass a score of 0, as follows:.

• meaningful degrading trend was assigned -2,
• significant degrading trend was assigned -1,
• insignificant or stable trend was assigned a score of 0,
• significant and meaningful improving trend was assigned 1,and
• significant and meaningful improving trend was assigned 2.

High values of the index indicate sites that fail several guidelines and for which several trends are degrading and low values represent the reverse. The sites were ordered in tables from highest risk (i.e. those with the largest index) to lowest concern. We stress that this is a subjectively defined ranking and the actual level of management concern must also include consideration of the values that are affected and their significance, which we have not considered. We also urge caution in using state and trends at specific sites as a basis for making conclusions about management because water quality conditions can be affected by very localized activities and be associated with legacies. We therefore consider that an overview of the region's water quality is more robustly made by grouping sites by River Environment Classification (REC) categories (see section 2.6.3).

2.6: Determination of overall state, trends and assessment of the monitoring network

We assessed overall state, trends in each region and each council's river monitoring network by grouping sites according to River Environment Classification categories (REC; see Section 2.5.4 below). The REC groups rivers that share similar environmental characteristics and which therefore tend to have similar physical and biological characteristics (Snelder and Biggs, 2002). REC Topography and Land-cover categories classify rivers according to the dominant topography and land cover of their catchments. Such groupings are commonly used to provide insights into the causes of spatial patterns of water quality state and trends in relation to environmental and human factors and to describe how well a network of sites represents the overall environmental variation within a region (e.g., Ballantine et al. 2010).

2.6.1: Representativeness of council's river monitoring network

We performed an analysis to assess how well each council's river monitoring network represented the environmental variability of the region's rivers by first evaluating the number of SoE sites in all combinations of REC Topography and Land-cover categories. We then compared the number of sites in each combination of REC Topography and Land-cover categories with the total length of rivers belonging to this category. The implicit assumption here was that river length is an appropriate weighting of 'representativeness'. This assessment, therefore, provides an indication of how representative the regional councils monitoring networks are in relation to river length in various categories. We acknowledge there are other physically and ecologically meaningful weightings that could be applied also (e.g., flow or riverbed area). We also acknowledge that regional councils have specific issues that influence the exact layout of their networks. The assessment provided here is just one way of assessing the 'representativeness' of the monitoring. Other criteria against which a monitoring network might be assessed could consider the overall perceived importance of water bodies or focus on areas subject to the greatest impact.

We made two sets of representativeness analyses. First, we counted the sites that that met our criteria for trend analysis (see Section 2.4). This provides an assessment of the representativeness of the historic network (i.e., how representative the council's network was – over the 10-year period 1999-2009). Second, we counted sites in the SoE network as of 2009 to analyse how representative the council's network is now. Ongoing changes in the number and location of network sites will mean that, in future, the existing network may be more or less representative than it was over the 10-year period.

2.6.2: Determination of overall state, trends and assessment of the statistical power monitoring network

Previous studies have shown that water quality state and trends vary strongly between sites within regions. In addition, studies have shown that within sites, there can be strong variation in water quality state and trends between variables. Sites can meet guidelines for some variables and not for others (e.g., Ballantine et al. 2009). There can also be conflicting trends at sites for different variables. For example, Figure 2 shows three quite different trends at the same site; a significant increasing trend for TN, a stable trend for TP and decreasing trend for DRP.

To provide regional summaries of water quality state and trends we grouped water quality sites into REC Topography and Land-cover categories and provided an overview of the category. For state we use boxplots to show the central tendency (i.e. the median) and dispersion (5th, 25th, 75th and 95th percentile values) of the median values of the individual sites in each group for each variable. We compared the median of the grouped values to the guidelines to show whether the categories "overall" tended to be within or exceed guideline values. We also tested whether there were statistically significant differences in the median of site median values when grouped by REC Topography and Land-cover categories using the Kruskal–Wallis one-way analysis of variance. A significant Kruskal–Wallis statistic indicates that there are differences in the group medians. We consider that the Kruskal–Wallis is one of many possible measures of the adequacy of the number of sites in each region's monitoring network. A significant test indicates sufficient statistical power (numbers of sites relative to the variability of the site medians) to detect large scale patterns (as defined by REC categories) in water quality state. Insignificant Kruskal–Wallis statistics would suggest that more sites are needed.

We used the binomial test⁸ to determine "overall trends" for sites grouped by REC Topography and Land-cover categories in each region and for each variable. We deemed that there was an overall trend in a certain direction for a grouping if the number of sites that exhibited that trend were greater than could be expected if increasing and decreasing trends were equally likely. The binomial test determined whether there are more trends in a group of sites than could be expected by chance. To perform a binomial test we first counted the number of positive RSKSE values (increasing trends). Note that all RSKSE values were included regardless of their p values. We then performed a "two-tailed" binomial test based on expectation that sites have a 50 per cent probability of having an increasing trend. If the resulting p-value was less then 0.05 we rejected the null hypothesis, i.e. we concluded that there were more trends in a group than could be expected by chance and that the group exhibited an "overall" trend. We then determined the overall trend direction as positive if the proportion of positive trends was greater than 50 per cent and negative if the reverse were true. A complication arises because RSKSE values can take the value zero for several reasons, some of which are related to data quality. In particular, RSKSE can be zero if there are many non-detect values in the time-series or if there are many identical values (ties), which occurs if the precision of the test or recorded concentrations are low. We added half of the number of sites with RSKSE values equal to zero to the number of increasing trends and performed the test based on this number. Note that the reported values are the number of sites with RSKSE values equal to zero regardless of their p-values and should not be confused with stable trends (i.e. RSKSE values equal to zero and p < 0.05).

We used the binomial tests as another measure of the adequacy of the monitoring network. A significant test indicates sufficient statistical power (numbers of sites relative to the variability of the trends among sites) to detect large scale patterns (as defined by REC categories) in water quality trends. We consider that insignificant binomial tests suggest large scale patterns in a region's water quality trends cannot be inferred and that more sites would be needed to detect such large scale patterns in trends.

2.6.3: River Environment Classification

The REC is based on a digital representation of the New Zealand river network comprising segments with a mean segment length of ~700 m. Each segment is associated with its unique upstream catchment. The catchment of each segment is described by various environmental variables (i.e. catchment characteristics) and these are categorised to define REC categories. REC Topography and Land-cover categories have previously been shown to distinguish significant differences in many river characteristics including water quality and hydrology (e.g., Snelder et al., 2005). We used the geographic coordinates and site names to locate all sites in the database on a single NZ Reach⁹ in the REC river network. Once linked with the river network, all sites were able to be associated with their REC categories and other data (e.g., site elevation) that were subsequently used in our analyses.

Table 3:

REC categories for the Topography and Land-cover groups of categories and the category criteria (see Snelder and Biggs, 2002 for details)

5: Above 150 metres a.s.l.

6: The action threshold for E. coli is 550 /100 ml. This guideline is for recreational water quality and applies to the "summer season" (1 November to 31 March).

7: LOWESS (locally weighted least squares) is a data analysis technique for producing a "smooth" function that describes a "noisy" relationship between two variables (Cleveland, 1979).

8: The binomial test is used for discrete dichotomous data, where each sampling event can result in one of only two outcomes.

9: The NZ reach is a unique valley segment, defined by the upstream and downstream tributaries, which is represented by the digital river network on which the REC is based.