Importance of long-term monitoring for detecting environmental change: lessons from a lowland river in south east England

Abstract. Rising nitrate concentrations in rivers and groundwater in regions with intensive agriculture have been of concern for several decades. Long records of nitrate concentration are rare; in the UK, few pre-date 1974. Records from the River Stour from 1937 are analysed using moving-windows to explore the effects of window width on trend detection and process inference. Nitrate concentrations rose sharply from the late 1950s in response to widespread ploughing of grassland, but have fallen since the early 1980s, suggesting that widespread nitrogen control measures, which were introduced after the start of the downward trend, may not be the principal driver for recent change. Short windows (5 years) are too noisy; longer windows (10, 15 and 30 years) identify signals (i.e. trends) more reliably. Since system response times can be of the order of 20–30 years, management decisions made on shorter data sets (even up to 15 years) could be misleading.


Introduction
High nitrate concentrations in rivers and groundwater have been a matter of concern 15 throughout the developed world for several decades (Burt et al., 1993). Initially, the main concern was about human health, reflected in the European Drinking Water Directive (EU 80/778). More recently, legislation, viz the European Nitrates Directive, has included reference to nitrate's role in eutrophication (including coastal waters where primary productivity is likely to be nitrogen-limited). Since the 1970s, a clear link has been 20 established between agricultural activity and nitrate transfer from land to surface and ground water (Royal Society, 1983). Of the nitrogen which enters surface waters in England and Wales, 61% estimated to originate from agricultural land and 32% from sewage effluent (Hunt et al., 2004), although these proportions will vary in different catchments. In rural catchments, nitrate losses from farmland predominate (>90%: Introduction Interactive Discussion tive demonstrated that protective action was considered to be a feasible alternative to water treatment as a remedy. However, at the catchment scale, solute transfers are often characterised by considerable delays in response times (Owens and Edwards, 1992;Kirchner et al., 2000) so detection of water quality improvements resulting from NVZ designation is usually impossible in the short term.

5
Studies of temporal variation in nitrate concentration in relation to changing climate and land use have focused on single river basins (Casey and Clarke, 1979;José, 1989;Worrall and Burt, 1999) or combined records from many catchments to make regional comparisons (Roberts and Marsh, 1987;Betton et al., 1991;Stuart et al., 2007). For example, an upward trend in nitrate concentrations was reported for the period 1971-10 1985 at Slapton Wood, a small catchment of mixed land use in Devon, UK (Burt et al., 1988). Agricultural intensification within the catchment (including ploughing of grassland, increased ploughing frequency, higher fertiliser application rates, the introduction of new crops and higher stocking densities) is most likely to account for the upward trend (Burt and Johnes, 1997). After accounting for the trend, nitrate concentration 15 was shown to be controlled by antecedent rather than concurrent conditions. A "memory effect" was evident with dry years followed by higher concentrations than expected in subsequent years and vice versa. Recent analysis  showed that the upward trend had not continued and that the "memory effect" had reversed in sign, suggesting change from supply-to transport-limited controls on nitrate leaching (Burt 20 and Worrall, 2007). There have been many studies of nitrate trends, but unless analysis spans many years, short-term fluctuations caused by climatic variations may be misinterpreted as trends. For example, a recent report that 77% of c. 6000 surface water sites in England had a declining nitrate concentration (ADAS, 2006) must be viewed with caution since only five years' data were employed and very few "trends" were sta-BGD 5,[2369][2370][2371][2372][2373][2374][2375][2376][2377][2378][2379]2008 Detecting long-term environmental change low-lying (7-128 m a.s.l.) and dominated by arable land (76% arable, 13% grassland, 4% urban: National Rivers Flow Archive). Annual mean nitrate-nitrogen concentrations (mg NO 3 -N l −1 ) and rainfall totals were calculated for water years beginning 1 October. Fig. 2 presents partial correlations between annual mean nitrate concentration and independent variables: year number (Anno Domini); annual rainfall total in the current 10 year (rf); annual rainfall total lagged by one year (rf-1); and, annual rainfall total lagged by two years (rf-2). A partial correlation coefficient indicates correlation between a dependent variable and one independent variable whist holding constant effects of other specified independent variables (Johnston, 1987). Here, we consider the effect of year, independent of rainfall, and vice versa. Correlations are repeated using a 15 "moving window" analysis (Worrall et al., 2003) with the result plotted against the final year of the series. Consistent trends can be identified from patterns of consecutive significant correlations. Correlations are deemed significant if the probability of chance occurrence is less than 5%. Moving windows of 5, 10, 15 and 30 years are used to demonstrate the effect of window width on trend identification and the role of climatic 20 controls at varying time scales.

Results
A seasonal cycle is evident in the raw data with maxima in late winter and minima in late summer (Fig. 1) (Doornkamp et al., 1978). The annual series shows that lower concentrations are often associated with drier years in the 1990s. The higher concentrations that follow occur in wetter years, probably due to the flushing of excess nitrate is flushed from the system.

5
There is great inconsistency between successive 5-year windows (Fig. 2a). Depending on when a period starts, it is possible to get very different trends. There are some statistically significant trends, positive and negative, but overall the record is noisy with no clear signal. Rainfall is highly variable in short records; correlations with rf are mainly positive as expected, with many non-significant results. Correlations with rf-1 10 are mainly negative when significant, implying exhaustion effects as described above, but there are three significant positive results in the dry period of the early 1990s. There are only six significant correlations with rf-2, three positive and three negative, but these do not suggest any consistent "memory effect".
Ten-year windows show longer periods of significant upward trend but there are sig-15 nificant downward trends ending in the early 1990s (Fig. 2b); the influence of high concentrations in the mid-1980s is crucial. Correlations with rf are all positive including many significant results; correlations are weaker in the early 1970s, which was a very dry period. Correlations with rf-1 shift from mainly negative in windows ending in the mid 1980s to positive in the late 1990s. This mirrors findings elsewhere (Betton et al.,20 1991) and reflects high rainfall in this period. For rf-2, two correlations are significantly positive in the early 1960s and gradually become negative, but not significant. Results from 15-year windows immediately indicate coherent patterns of change (Fig. 2c). Long-term trends are positive and mostly significant through to the mid 1980s. Trends then switch (over just five windows) to become significantly negative through Interactive Discussion what was seen using 10-year windows. There is no pattern with correlations for rf-2: all are non-significant.
Thirty-year windows show a very smooth pattern of change (Fig. 2d). Even so, strong positive trends identified by windows ending in the late 1980s are followed by a relatively rapid change with trends for the last 6 windows being significantly negative. High 5 concentrations in the period 1977-1986 exert key influence over observed trends. Correlations with rf, which are nearly all significantly positive, imply clear process links between rainfall, nitrate leaching and river water concentrations. Correlations with rf-1 are significantly negative for windows ending around 1990, a period when annual rainfall totals were low. Once again, there are no significant correlations for rf-2.

Conclusions
We draw the following conclusions: 1. The pattern of long-term change revealed for the River Stour reflects the complex interactions between nitrogen availability in the catchment, hydrological controls over nitrate leaching and hydrogeologically controlled delays in solute break-15 through. The decrease in nitrate concentrations appears to have started in the early 1980s, before widespread nitrogen control measures were introduced. This suggests that much of the rising trend in the post-war period was caused by nitrogen mineralisation in soil resulting from ploughing of old grassland (Whitmore et al., 1992) and augmented by increased nitrogen inputs. 20 2. Climatic fluctuations induce high-frequency variations around the long-term trend. Rainfall in the current year is significantly and positively correlated with nitrate concentration, as would be expected; this is seen for 10-year windows and longer.
The "memory effect" is seen to shift from negative to positive in the 10-year window. This could either relate to a shift from drier to wetter years, or indi-Introduction have moved from supply-limited conditions (in which the overall trend appears to have been influenced predominantly by nitrogen availability in the system) to a transport-limited state (in which nitrogen availability is relatively constant and temporal variations in concentrations are controlled mainly by hydrological conditions). Both the timing and the shift mirror results found elsewhere (Burt and Worrall, 2007). Only the negative memory effect is significant when longer windows are considered. Here, no significant lags are identified for rf-2.
3. Use of moving window analysis has shown that statistically significant trends may be identified for almost any time period. However, this study suggests that the response times for nitrate in this system can be of the order of 20-30 years (par-10 ticularly in the case of re-equilibration of soil organic matter levels in ploughed-out pasture: Whitmore et al., 1992). Management decisions made on shorter data sets (even up to 15 years) could, therefore, be misleading: interventions could be judged unsuccessful, even if they have had significant positive impact in the long term. 4. There are three implications for scientists and government agencies involved in environmental policy and management. First, in order to know if change is happening, monitoring systems must be adequately maintained over long periodsrecords less than thirty years may well be of little use (Burt, 1994). Secondly, when results are analysed, they must be viewed within an appropriate historical 20 context (i.e. a sufficiently long record) in order to be able to distinguish the impact of policy-driven change from shorter-term variability. Thirdly, present nitrate management strategies have not taken into account the evidence that the impact of such strategies will always have a limited effect within the context of major land-use change (i.e. widespread ploughing of grassland in the post-war period). year, rainfall in current water year (rf), rainfall in previous water year (rf-1), and rainfall two years ago (rf-2). Solid and open symbols represent significant (p<0.05) and non-significant correlations respectively.