• About this site|
• Site index|
• Contact us|
• Jobs

Nationwide and regional state and trends in river water quality 1996-2002

Executive Summary

1. Introduction

2. Methods

3. Results

4. Discussion

5. Acknowledgements

6. References

Appendix 1: Appended Table 1

Appendix 2: Plots of water quality in land-cover classes, grouped by climate classes

2. Methods

2.1. Sources of data and dataset reduction

Water quality data were acquired from the river and stream monitoring programmes of 15 regional, district and city councils, and from NIWA's National River Water Quality

Table 1: REC classification levels, categories, mapping characteristics, and the criteria used to assign river segments to REC classes

View REC classification levels, categories, mapping characteristics, and the criteria used to assign river segments to REC classes (large table)

Network (NRWQN). The original dataset comprised 996 sites across the country, and over 10,000 sampling events. The large amount of data allowed us to apply rigorous rules for the inclusion of monitoring sites in datasets used for state and trend analyses. The intent of the rules for state analyses was to ensure that data for each site was current, and that sampling was conducted year-round, to reduce the influence of seasons or individual sampling dates on average conditions. The intent of the rules for trend analysis was to ensure that the data were contemporaneous, and that sites were sampled with sufficient frequency, and over a sufficient duration to account for seasonal effects. The rules were as follows:

State analyses

• Earliest date after 1 December 1998
• Last date after 1 Dec 2001
• ≥11 months duration
• ≥10 months represented, or regular quarterly or bimonthly sampling
• ≥10 sampling dates

Trend analyses

• Starting date in January 1996
• Final date in December 2002
• 5 to 7 years duration
• Monthly to bimonthly sampling (no gaps > 3 months)

The resulting datasets were inspected for unusable data. Coastal sites at which conductivity levels were routinely greater than 700 μS/cm were considered to be exposed to seawater and were removed (approximately 40 sites). The highest conductivity levels measured in upland streams were < 600 μS/cm. Data collected at high frequencies with automated dataloggers were converted to daily averages. Measurements that were clearly erroneous or estimated (e.g., "clarity ± 6 m") were deleted. No other data were deleted as outliers, due to our lack of familiarity with the potential range of values at individual sites. Nutrient concentrations below detection limits were replaced with values equal to half of the detection limit. There was a two- to three-order of magnitude range in the concentration of each nutrient, so these adjustments at very low levels were unlikely to bias results. The final datasets for the state and trends analyses consisted of 618 and 386 sites, respectively. Figure 2 shows the locations of all monitoring sites in the dataset. Not all monitoring programmes

included the entire suite of water quality parameters considered in this report. Further, one or more parameter values is missing on some sampling dates at many monitoring sites. As a result, the total number of data points varies among parameters.

For state analyses, monitoring sites were used as replicates within REC classes, and data from individual sampling date were treated as subsamples, i.e., parameter values for each site were averaged across dates, and the averages were used as single points in subsequent analyses. For trend analyses, site-specific slopes of parameter values over time were used as replicates.

2.2. REC classification

Monitoring sites were assigned to REC classes in three steps. First, grid references provided by the regional councils were converted into a GIS layer of site locations. Second, this GIS layer was overlain on a digital topographic map (NZMS 260 series) and edited as necessary to ensure that site locations corresponded with stream locations on the map, and details provided by the councils (e.g., bridges, dams). Third, each site was assigned the NZReach number of the stream section where it is located. NZReaches are unique, national-scale stream section identifiers stored in the REC GIS database. Each NZReach number has associated with it attributes which include REC classification. Detailed methods for using the REC with GIS are given in Snelder et al. (2002).

To increase replication within land-cover classes, the following classes were pooled: glacial mountain and mountain source-of-flow classes, alluvial, plutonic and hard sedimentary geology classes, and indigenous forest, scrub, bare and tussock land cover classes. The latter category, referred to herein as "natural" land cover, was used to compare streams in relatively undisturbed catchments with those in catchments developed for agriculture, exotic forestry, and urbanisation. For a subset of the analyses below, the original four natural land cover categories were retained.

In order to characterise the state of New Zealand rivers, it is important that the dataset being employed is representative of the range of conditions present in New Zealand rivers. If the dataset is not representative, there is an unacceptable level of bias that may undermine the validity of conclusions. The 618 monitoring sites used for state analyses were distributed among 88 REC classes (Table 2). If the dataset accurately reflects the distribution of REC classes, the proportion of monitoring sites in a class will be similar to the proportion of total river length in the same class. At the REC land-cover level, the distribution of sites among classes is fairly representative, i.e., most of the abundant classes in terms of river length are present in the dataset, and the proportions of sites in many classes are similar to the proportions of river length.

Table 2: Distribution of monitoring sites among REC classes, percent of sites in each class, and percent of river length in each class

View distribution of monitoring sites among REC classes, percent of sites in each class, and percent of river length in each class (large table)

There are several exceptions, however. Many REC classes are uncommon and are either unmonitored or are monitored too rarely to be included the analyses; of the 265 land-cover level REC classes present in New Zealand, 177 are in this category. The most common REC class in New Zealand, CX/M/HS/N (8.7% of total river length), is underrepresented with 5 monitoring sites (0.8% of sites). The WW/H/HS/P class makes up nearly 6% of the total river length, but there are no monitoring sites in this class. The CW/H/VA/P, CW/L/HS/P, CW/L/VA/P and WW/L/SS/P classes are overrepresented, with proportions of sites > 10 times greater than proportions of river length. There are also mismatches between monitoring sites and river lengths at higher REC levels. Several source-of-flow classes are unmonitored, including CD/M and WD/Lk, both of which comprise 1% of the total river length. The mountain source-of-flow class comprises 17% of the total river length, but only 3% of monitoring sites. In contrast, the CW/H, CW/L and CW/Lk classes are overrepresented with 44% of monitoring sites, and 25% of the total river length.

2.3. Water quality parameters

Five water quality parameters were selected for the state analysis, based on their utility as indicators of environmental degradation, and on the number of councils that include them in monitoring programmes. The parameters were nitrate+nitrite (NO_X), ammonium (NH₄), dissolved reactive phosphorus (DRP), and Escherichia coli concentrations, and water clarity measured by the black disk procedure (Davies-Colley 1988). Trend analyses were conducted using the same five parameters, plus temperature and flow. Stream health assessments frequently consider macroinvertebrates and periphyton, but these parameters were not assessed in the present study, because council data generally did not meet the rules listed above for site inclusion in the national dataset (Stark et al. 2001, Wilson 2001)

NO_X, NH₄ and DRP were selected for analysis because the availability of these inorganic nutrients can affect growth rates of periphyton (algal/microbial assemblages attached to stream substrates) and macrophytes (rooted aquatic plants), not because they pose human health risks. Nitrite (a component of NO_X) and ammonia may both be toxic at high concentrations, but these compounds rarely reach toxic levels in natural surface waters. Many dissolved nutrients are required by periphyton and macrophytes, but nitrogen and phosphorus are particularly important because they are generally scarce relative to physiological demand for them. If either nitrogen or phosphorus is sufficiently scarce, its availability will be the limiting factor for growth. With rare exceptions, other nutrients are present in higher concentrations relative to physiological demand. Increased availability of limiting nutrients causes periphyton and/or macrophyte growth rates to increase, which may result in proliferations of plant material. Such proliferations are undesirable for recreation, and can have negative effects on other aquatic organisms. Simple, inorganic molecules containing nitrogen and phosphorus are used by periphyton and macrophytes; larger organic molecules generally cannot be assimilated. NO_X and NH₄ are the most abundant assimilable forms of nitrogen. DRP consists primarily of orthophosphate, the most abundant assimilable form of phosphorus.

Water clarity was selected as a measure of aesthetic water quality, and potential ecological degradation (as light penetration may limit the growth of aquatic plants). Clarity integrates the effects of suspended particles and dissolved coloured substances in stream water. Related measurements include total suspended solids concentration, light absorption by dissolved substances, and turbidity. Clarity is the optical parameter most frequently measured by regional and district councils. Studies from New Zealand rivers indicate that black-disk water clarity is correlated with turbidity, suspended solids concentration, and colour (Close and Davies-Colley 1990).

E. coli concentration was selected as an indicator of microbiological water quality as it affects contact recreation. Many potentially pathogenic micro-organisms occur in natural waters, and it is not possible to test for each type on each sampling date. Rather, a single type is chosen that is common in mammalian gastric tracts, and is therefore indicative of faecal material in water. E. coli is the current choice of indicator bacteria for most freshwater monitoring programmes, and guidelines recommended guidelines by the New Zealand Ministry for the Environment are based on E. coli concentrations (MfE 2002a, 2002b).

Flow and water temperature were used in the trend analyses as both parameters reflect climate variability (Scarsbrook et al. in review). A consideration of climate variability is required to interpret results and assess the factors that cause interannual trends.

Staff at the councils that provided water quality data also provided details of measurement methods. Because data from the 16 monitoring programmes were pooled for analysis, differences among sample collection methods and among laboratory analysis methods, both between programmes and within programmes over time, were of concern. Most of the laboratories at which water samples are currently analysed are accredited by International Accreditation New Zealand (the national authority for the accreditation of testing laboratories), but some are not. While it is not possible to adjust data for methodological differences, we provide a summary of the methods for comparison (Appendix 1), and make some recommendations for reducing methodological variability in the Discussion.

2.4. Objective 1 - Assessments of water quality state and comparisons among REC classes

The state analysis had two components. The first was an assessment of water quality within REC classes, using several REC levels to vary the spatial scale. These assessments were based on graphic comparisons of the water quality data for each class with recommended guidelines for New Zealand rivers. The second component was a series of comparisons of water quality among REC classes, and was also conducted at several REC levels. These comparisons were made using statistical tests and served in part to validate the assessments in the first component. The REC hierarchy served as a general framework for comparing classes, but simultaneous comparisons across the hierarchy (e.g., by nested analysis of variance) could not be carried out because the dataset was unbalanced; many classes were not represented, and sample sizes varied widely. Instead, comparisons were made among the classes for which there were sufficient replicate sites (≥ 6), when organised at various levels of the REC (Table 3). To reduce the influence of extreme values within REC classes, comparisons were made using nonparametric tests of ranks, adjusted for unequal sample sizes (Zar 1984). For comparisons of two classes, Mann-Whitney tests were used. Comparisons of three or more classes were made in two steps: non-parametric analyses of variance (Kruskal-Wallis tests) were used to determine whether there were significant differences among classes; if the overall test was significant, non-parametric pairwise comparisons were used to assess differences between classes. In all comparisons of classes, we considered "statistically significant" differences to be those in which the probability of obtaining differences as extreme as those observed by chance alone was less than 5% (i.e., P value < 0.05). We used the same nominal probability level was used in trend analyses. Hereafter, the term "significant" is used to mean statistically significant as defined above. It is important to note that statistically significant differences are not always environmentally consequential; small differences may be significant (Discussion, Section 4.1.2).

Table 3: Organisation of water quality state analyses

View organisation of water quality state analyses (large table)

The broadest comparisons were made at the REC climate level. These comparisons encompassed most of the sites in the dataset, but had the poorest resolution, i.e., relatively little variability was partitioned by REC classification. Five of the six climate classes in New Zealand were included; the WX class was represented by only 5 sites, and was not included. The next broadest comparisons were made at the source-of-flow level. These had increased resolution, but some classes were not included (the CD/Lk, WW/H and WX/L classes) because they were represented by fewer than six sites. Further, no comparisons could be made with the WD/L or WW/L classes because there were no contrasting source-of-flow classes within the WD and WW climate classes. No comparisons were made at the REC geology level because most classes (37 out of 50) could not be included, either because of insufficient replication, or because there were no contrasting geology categories within the same climate and source-of-flow class with sufficient replication.

Effects of low replication at the REC land-cover level were more severe than at the geology level. However, the land-cover class was considered to be critical for assessing the state of water quality in New Zealand. Therefore, as many comparisons were made as possible. The first land-cover level comparisons were among fully-classified land-cover classes, i.e., among land-cover classes within each climate, source-of-flow, and geology class (Table 3). Only 10 fully-classified land-cover classes had sufficient replication for analysis. Therefore, two additional sets of land-cover comparisons were added by collapsing part of the REC hierarchy; among land-cover classes within each climate class (pooling all source-of-flow and geology classes), and among land-cover classes (pooling all climate, source-of-flow, and geology classes). Compared with comparisons of fully-classified classes, the latter comparisons had higher within-group replication and greater geographic scope, but lower spatial resolution. Due to insufficient replication in the Urban and Exotic Forest classes, the comparisons described above were limited to Pastoral and Natural classes (Table 3). To examine water quality state within the land-cover classes making up the Natural class, as well as the Urban and Exotic Forest classes, a fourth set of comparisons were made among all of the original land cover classes (Table 3).

The current state of each water quality parameter in each REC climate, source-of-flow and land-cover class is presented graphically. Parameter medians and percentiles are given for classes represented by ≥ 3 monitoring sites; point values or averages are given for classes with 1 or 2 monitoring sites. In each graph, "default trigger values" for water quality parameters for New Zealand rivers are shown. The default trigger values or guidelines are given in Table 3.3.10 of the Australian and New Zealand Guidelines for Fresh and Marine Water Quality (ANZECC and ARMCANZ 2000), and the Ministry for the Environment Freshwater Microbiological Water Quality Guidelines (MfE 2002) and the Water Quality Guidelines for Colour and Clarity (MfE 1994). These guidelines are summarized in Table 4. The guidelines in Table 4 are not legal standards, but are considered "trigger values"; a site-specific investigation or remedial action may be warranted when the guidelines are not met. ANZECC guidelines for upland (> 150 m elevation) and lowland rivers differ slightly. As the sites in the dataset represent a wide elevation range, and these elevations are pooled within REC classes, the ANZECC guidelines for high and low-elevation rivers have been averaged.

Table 4: Guideline water quality values for protection of New Zealand river ecosystems (ANZECC 2000), and human health (MfE 2002)

The New Zealand Ministry for the Environment has recently adopted new guidelines for freshwater microbiological water quality (MfE 2002a). The new guidelines refer to E. coli concentrations in single samples from collections made at daily to weekly intervals over the bathing season. In the dataset we are analysing, samples were generally collected monthly, the values for each site were averaged across dates, and we do not reference individual samples from individual sites. Consequently, we have used the previous guidelines for E. coli, which are based on medians (MfE 2002b). ANZECC guideline values are intended to be compared with the median value from independent samples at a site (ANZECC and ARMCANZ 2000). In our assessment of water quality in REC classes, guideline values are compared to medians of sites within each REC classe, rather than to medians of samples within each site.

2.5. Objective 2 - Recent trends in water quality

2.5.1. Hydrological data and flow-adjustments

The water quality data used for trend analysis was flow-adjusted to reduce the variability associated with fluctuating discharge. Such variability may be caused by dilution, or by hillslope, bank and channel erosion during high flows, and can obscure the monotonic, interannual changes that were of interest in this study. Flow adjustments were made using daily average flows for each sampling site and date. The flow data were acquired by a four-step process:

(1) the locations of monitoring sites were compared to locations of NIWA and regional council flow recorders;

(2) for monitoring sites at the same locations as flow recorders, daily average flows from recorders were acquired from the National Hydrometric database;

(3) for monitoring sites that were not located at flow recorder, the most appropriate flow recorder in the area was selected, based on proximity and characteristics of the monitored stream and its catchment;

(4) daily average flows at the latter monitoring sites were estimated by scaling the flows at the selected recorders.

Scaling factors were specific yield (m3 s-1 km-2) and net precipitation (precipitation - potential evapotranspiration). Mean annual precipitation and evapotranspiration data were provided by Landcare Research. Approximately 75% of flow recorder sites were within 10 km of the corresponding monitoring site, and 50% were within 1 km.

Flow adjustments were made using the procedure of Smith et al. (1996). Briefly, values of each water quality parameter were plotted against discharge, and the LOWESS smoothing procedure applied using visual analysis software (DataDesk 6.1, DataDescription, Inc.). Adjusted residuals were calculated as the differences between observed and smoothed values, plus the median for all data in the plot. Trend analyses were then carried out using the adjusted residuals.

2.5.2. Trend calculations

Methods for conducting water quality trend analyses must account for seasonal variations, and must be robust when used with non-normal datasets. The Seasonal Kendall test meets these requirements and is in widespread use (Griffith et al. 2001). In this procedure, the Seasonal Kendall Sen Slope Estimator (SKSSE) is calculated for values of a parameter through time at individual sites. The SKSSE is the median of all possible combinations of slopes within each month over the period of record. It is this value that provides an estimate of the direction and magnitude of trends in water quality parameters. In the present study, we use the SKSSE as the basis of trend analyses within and between REC classes.

SKSSE calculations were automated using Microsoft Excel macros and scripts written in the batch processing software package Winbatch (Wilson Windowware 2002). The following steps were followed to produce the consistent dataset required for the macros:

1. Data were filtered to remove gross outliers (values that exceeded the next highest values by two or more orders of magnitude). Outliers can create serious errors in trend analyses.
2. 'Greater than' values were rounded up (e.g., E. coli >24200 became 25000). These adjustments do not affect non-parametric analyses.
3. Data were filtered to identify months with two or more data values. The first data value each month was retained, and subsequent values were deleted, except where a later sampling date had a more complete set of parameters. In some situations, there were two sampling occasions in one month (e.g., 2nd January and 30th January), but no sampling in the following month. We changed the date on the second sampling occasion to fill the gap in the succeeding month.
4. A consistent time series was generated. The starting date for each site was during January 1996 and the finishing date was during December 2002. This gave a total of 84 individual months (rows). A macro was used to ensure all sites and parameters conformed to this time series. Rows were inserted for missing months.
5. For each site, values for each parameter were sorted by month and year. We then generated a matrix of all possible combination of slopes for any given month, and the median slope (SKSSE) was calculated from this matrix.

To allow for comparisons of slopes across sites, we divided the SKSSE by the median of data values to produce the relative SKSSE (RSKSSE). RSKSSE values were calculated for the flow data, and for raw and flow-adjusted data for all other parameters. Trend analyses were then made by aggregating RSKSSE values at various levels of the REC (Table 5). Two questions were addressed. First, does the median value of aggregate RSKSSE values comprise a statistically significant trend? We calculated the 95% confidence limits for the median (Zar 1984), and inferred that a statistically significant trend existed if the confidence interval did not include zero. Second, do RSKSSE values vary among REC classes? Comparisons were made using the non-parametric approach described for the state analyses (Section 2.4): Mann-Whitney tests were used in comparisons of two classes, and Kruskal-Wallis tests followed by pairwise comparisons were used for three or more classes. As with the state analyses, a probability of 5% or less that differences between compared trends as large as those observed arose by chance alone was used to denote statistically significant differences.

As with the analyses of patterns in water quality state, we were interested in patterns across several spatial scales. At the broadest scale, national trends in water quality parameters were assessed by pooling all REC levels. Finer-scaled comparisons were made at the source-of-flow and land-cover levels (Table 5).

Table 5: Organisation of water quality trend analyses.

Comparisons among classes were made within the last REC level listed in each column.

2.6. Objective 3 - Water quality state in regions compared to national averages, and inter-region comparisons.

Relationships between water quality and REC classes may vary geographically, and the causes of such variation are likely to change with spatial scale. At fine spatial scales (1 - 10 km²), the types of vegetation that make up some land cover classes differ across geographic regions. For example, the species that dominate indigenous forest and scrub land-cover classes vary with latitude, elevation and distance from the coast. Within the pastoral land-cover class, crop and pasture plant composition varies geographically, as do land-use practices and management activities. Management activities and their effects on water quality are likely to differ among New Zealand's jurisdictional regions, as water and agricultural management policies are created, implemented, and enforced at the regional level. Further, small-scale geographic differences in REC classes maybe propagated to larger scales. For example, geographic differences in water quality within several land-cover classes may collectively cause geographic differences at the geology or source-of-flow level.

To address the possibility that water quality within REC classes varies among jurisdictional regions in New Zealand, water quality state in land-cover and source-of-flow classes within individual regions was compared to the national average for the same classes. Jurisdictional regions were used as geographic units for comparison because differences in land use practices and management that affect water quality were of interest, and because the characteristic spatial scales of New Zealand's regions and unitary authorities (400 - 50,000 km²) match the characteristic spatial scales of source-of-flow classes (100's - 1000's of km²). Regions included in each comparison were selected based on two rules: first, only regions with ≥ 6 sites in the REC class being compared were included, in order to compute confidence intervals. Second, only regions whose sites made up less than half of the sites in the national dataset for a given REC class were included, to prevent bias. For example, if there were 30 CD/L/HS/N sites in the national dataset, and 20 of those sites were in the Canterbury Region, then Canterbury was not compared with the national average.

Inter-region comparisons were made graphically and statistically using REC land-cover classes and source-of-flow classes in the Canterbury, Southland and Waikato Regions. These regions are to be used for case studies in subsequent Ministry for the Environment issue-based reports. There were some differences in data availability and REC classes among the regions: Canterbury data does not include clarity, and Waikato data does not include NO_X concentrations for some REC classes. At the land-cover level, only four classes were common to two of the regions; the CD/H/HS/P, CD/L/HS/P, CW/H/HS/P, AND CW/M/HS/N classes are present in Canterbury and Southland. No land-cover classes were common to all three regions. Shifting to the source-of-flow level increased the number of common classes to seven, and the CW/H and CW/L classes were present in all three regions.

2.7. Objective 4 - Differences in water quality among stream orders.

Stream order is frequently used as a surrogate for stream or river size. It is possible that variability due to changing stream orders obscures some larger-scaled spatial patterns caused by higher-order factors. The REC system treats stream order as a factor that is spatially nested within land-cover classes. Stream order is therefore predicted to be a significant source of variability within land-cover categories, but not within climate, source-of-flow or geology categories. If stream order is a significant source of variability within REC classes, then water quality assessments should either treat stream order as an explanatory variable (which may reduce replication within classes), limit the sites in the assessment to those of similar order (which may reduce the scope of the assessment), or increase the number of sites to cover more orders (which may be unavailable or expensive). To determine whether significant water quality differences among stream orders are primarily at the land-cover level, or are frequent (or rare) at all REC levels, we compared stream orders within each REC level in turn: among stream orders within climate classes, among stream orders within source-of-flow, and so forth (Table 6).

Comparisons were made in each class in Table 6 using stream orders for which there were six or more sites in the dataset. As in Objective 1, comparisons among stream orders were made using nonparametric tests, adjusted for unequal sample sizes.

Table 6: Organisation of comparisons among stream orders.

In each case, comparisons of water quality parameters were made among stream orders, within each class of the higher REC levels shown.