David K. Mueller, Hydrologist

U.S. Geological Survey

Phone: (303) 236-2101, x235; Fax: (303) 236-4912; E-mail: mueller@usgs.gov

Jeffrey D. Stoner, Hydrologist

U.S. Geological Survey

Agricultural practices have been linked to water contamination by nutrients, including nitrate, in the United States. High concentrations of nitrate can be a health hazard in drinking water and can cause eutrophic conditions to develop in estuaries. Nitrate discharging from the Mississippi River has contributed to the depletion of oxygen in a large area of the Gulf of Mexico. Water-quality data collected during 1993-95 for the U.S. Geological Survey’s National Water-Quality Assessment (NAWQA) Program can be used to determine areas in the United States that might be susceptible to nitrate contamination. Analysis of data collected prior to 1990 identified significantly higher nitrate concentrations in water from agricultural areas compared to urban areas or undeveloped areas. For this paper, data describing nitrate concentrations in samples collected from streams under consistent protocols were compiled for 72 sites downstream from agricultural basins in NAWQA study areas. Mean annual concentrations at these sites, along with computed annual nitrate yields and loads, were compared relative to geographic factors, including land use, soil characteristics, and nitrogen inputs in the upstream basins. The most significantly correlated factors were used in multivariate statistical models to help identify the potential for nitrate contamination associated with various environmental settings within the United States. The best model results were for relations between nitrate yield or load and measures of streamflow and upstream nitrogen sources. These results indicate a potential for using spatial data to estimate nitrate contamination in streams over broad areas of the Nation.

According to the U.S. Environmental Protection Agency’s 1994 evaluation of the state of the Nation’s waters, 37 percent of the assessed rivers and streams were impaired by excessive nutrient levels (USEPA, 1994). An important negative effect of excessive nutrient concentrations is accelerated eutrophication of streams and receiving waters. Although the growth rate in most fresh-water ecosystems is limited by phosphorus, nitrogen is usually more important in salt water. In recent years, the Mississippi River has discharged as much as one million megagrams (Mg) of dissolved nitrate-nitrogen annually into the Gulf of Mexico (Goolsby and Battaglin 1995). Nitrate and other nutrients are suspected of being responsible for a large zone of hypoxia (seasonally low dissolved oxygen concentrations) along the Louisiana-Texas coast (Justic et al 1993). Agriculture, specifically the use of nitrogen and phosphorus fertilizer, is a suspected cause of the Gulf of Mexico zone of hypoxia (Rabalais et al 1996).

A portion of the nitrate in the Nation’s streams and ground water comes from distributed nonpoint agricultural activities, such as the application of inorganic fertilizer and animal manure. Nitrogen not used by plants or returned to the atmosphere is readily converted to nitrate in the soil. Nitrate is soluble and persistent in water and therefore can be leached to the water table or delivered to nearby streams. Environmental factors other than land use, such as available water, soil drainage, water residence time, and available carbon, can have important effects on the presence and concentration of nitrate in streams and ground water (Mueller et al 1995; Nolan et al 1997).

Knowledge of the quality of the Nation’s streams and aquifers is important because of the implications to human and aquatic health and because of the significant costs associated with decisions involving land and water management, conservation, and regulation. In 1991, the U.S. Congress appropriated funds for the U.S. Geological Survey (USGS) to begin the National Water-Quality Assessment (NAWQA) Program to help meet the continuing need for sound, scientific information on the spatial extent of the water-quality problems. Objectives of the NAWQA Program include understanding how these problems are changing with time, and what effects human actions and natural factors have on water quality conditions. Understanding regional patterns and environmental factors affecting nitrate concentrations in streams and shallow ground water is essential for effectively developing programs to manage and protect these water resources.

This paper describes methods used to characterize nitrate, in terms of concentration, load, and yield, at stream sites sampled for the NAWQA Program. The distribution of nitrate at these sites is compared to measures of agricultural land use upstream. This analysis uses data collected by consistent methods during 1993-95 within the 20 NAWQA study units shown in figure 1. A study unit is a major hydrologic system of the United States in which the NAWQA studies are focused (Gilliom et al 1995).

Each NAWQA study-unit investigation team screened and reviewed available data on nutrients, pesticides, and associated environmental data for streams and ground water. These analyses were used to develop an environmental framework for selecting sampling locations to answer questions about agricultural and urban land-use effects on water quality. A study-unit stream-monitoring network typically included 10 to 12 sites representing watersheds of various scales. About half of the watersheds are small, typically 50-500 square kilometers (km²), and have relatively homogeneous land use. The remaining sites are located at outlets of large complex watersheds, which commonly contain multiple land uses and include a substantial percentage of study-unit area (10-100 percent). More specific information about study-unit sampling design is provided by Gilliom et al (1995).

Consistent methods for the collection and handling of water-quality and ancillary data are critical for national assessments where trends are analyzed in space and over time. Rigorous protocols were established for stream sampling procedures (Shelton 1994). Samples were collected using a depth-integrating sampler at multiple vertical locations in the stream cross section. Nutrient samples were chilled and maintained at 4 degrees C until analyzed at the laboratory. Samples for analysis of dissolved constituents were filtered in the field, within two hours of collection, through a nitrocellulose filter or a polyether-sulfone medium with a pore size of 0.45 micrometers. All analyses were performed at the USGS National Water-Quality Laboratory according to methods described by Fishman (1993). Determinations were made for a suite of nutrients that included dissolved nitrate plus nitrite as nitrogen, hereinafter referred to as nitrate, at a detection limit of 0.05 milligrams per liter (mg/L).

Sampling at stream sites in the 20 NAWQA study units generally started by March 1993 and continued through the 1994 and 1995 water years (October 1993 through September 1995). Routine sampling included field blanks and replicates to assess measurement bias and variability (Mueller et al 1997a).

Spatial geographic data for the drainage basins upstream from sampling sites were compiled from a variety of sources. These data included land use and land cover classified by Anderson et al (1976) and soil-series characteristics reported by the U.S. Soil Conservation Service (1993). Estimated nitrogen inputs included commercial fertilizer application (Battaglin and Goolsby 1995), manure application (Smith et al 1997), and deposition of atmospheric nitrogen (National Atmospheric Deposition Program 1997). Only the atmospheric-deposition data were available for the entire period of NAWQA sampling (1993-95). Fertilizer data were available through 1994, so the period 1992-94 was used to obtain an average that was likely to have been applied in the NAWQA basins prior to and during the time of sampling. This was justifiable because fertilizer application rates have varied little over the past decade. Manure application data are based on the Census of Agriculture, which is made on five-year intervals. The latest available data were for 1992.

Samples were collected at about a monthly frequency from 227 sites in the 20 NAWQA study units. Three of these sites were affected by substantial upstream diversions; therefore, drainage areas could not be defined. Two sites in the San Joaquin Valley of California were on sloughs draining the same area and were considered a single site in subsequent analyses. Thus, ancillary data could be determined for 223 drainage basins. Of these, 92 were classified as small to moderately sized agricultural basins, according to the following criteria:

1. More than 25 percent of the area is cropland or more than 50 percent is cropland plus pasture,
2. The site is not significantly affected by urban sources, and
3. The drainage basin is less than 20,000 km².

Additional criteria were applied to select sites suitable for computing annual nitrate loads. These criteria were:

4. Availability of daily streamflow data for at least one of water years 1994 and 1995.
5. Significance of a statistical model relating measured nitrate concentrations to streamflow at the site.

Streamflow records were adequate at 88 of the 92 sites downstream from agricultural basins, but 16 sites were excluded because of problems identified in fitting the statistical model.

The resultant data set included 72 agricultural sites, distributed among the 20 NAWQA study units, as shown in figure 1. This distribution of sites was affected by the intensity of agriculture in the study units and the local interest in agriculture-related surface-water quality issues. Sites were concentrated in the Mid-Atlantic area, the upper Midwest, and the Northwest. No sites were selected from study units in Nevada or the Rio Grande valley; however, intensive agriculture is limited in these areas.

The basic sampling frequency at NAWQA stream sites was monthly, with several additional high flow samples collected to characterize that part of the hydrograph where concentrations might be more variable. A subset of sites was selected for more intensive sampling during time periods when concentrations of certain constituents, primarily pesticides, were expected to be high. Intensive sampling frequencies varied from biweekly to as often as every other day.

Two difficulties arise from this variability of sampling frequency:

• Summary statistics, such as mean concentration, used to characterize a site might be biased because of variations in sampling frequency during the period of record, and
• Comparisons among sites might be biased because of different sampling frequencies and numbers of samples at different sites.

These difficulties might be avoided by using flow-weighted mean concentrations, but only if sampling was distributed evenly over the hydrograph. Intensive sampling during high flow or low flow, such as was common at some NAWQA sites, might overemphasize one end of the hydrograph and yield a biased result.

Another method for decreasing bias in site characterization and comparison is to estimate a concentration value for each day of a common period of record, such as a particular water year, and compute a flow-weighted mean of these estimates. There is a long history of statistical models that have been proposed to make such estimates. The model selected for the present analysis is a modification of a multivariate regression equation used by Cohn et al (1992) to estimate nutrient transport to Chesapeake Bay. This model relates nitrate concentration (C) to streamflow (Q) and time (T), measured in years:

where ln( ) denotes the natural logarithm. The sin and cos terms are included to consider seasonality that is independent of streamflow; the remaining time term corresponds to long-term trends. Retransformation of the estimated ln(C) values produces a bias that is corrected in the present analysis by using a method proposed by Ferguson (1986):

where C_DV is the nitrate concentration estimated for particular daily values of Q and T, and s² is the estimated variance of the residuals from regression of sample data at a particular site.

The regression model (equation 1) was fit for each site that met the first 4 criteria listed above. All measured nitrate concentrations in NAWQA samples from a site were used to fit the model. Models for each site were evaluated by assessing the overall significance and analyzing the distribution of residuals. This constituted the fifth site selection criteria. Sites were retained for subsequent analysis only if the model fit was significant (p<0.10) and the distribution of residuals was reasonably normal and homoscedastic.

Adequate models were obtained for 72 sampling sites downstream from agricultural areas. For each of these sites, flow-weighted mean nitrate concentration was computed from the daily streamflow and estimated daily concentration values. The period used for this computation was water years 1994 and 1995. If daily values were not available for a complete water year, that year was not used in the computation. The distribution of these flow-weighted means of daily concentration estimates is shown in figure 2 in relation to the means and flow-weighted means computed from the sample data from the 72 agricultural stream sites. The sample means are generally lower than the flow-weighted means for estimated daily concentrations, probably because high concentrations are not adequately sampled at many sites. The flow-weighted sample mean might be the most appropriate summary statistic for individual sites, but the variability among sites is smaller than for the other statistics, and differences among sites might be underestimated. The flow-weighted mean of daily estimates preserves the variability among sites without the low bias for individual sites. Therefore, these mean values from nitrate concentrations estimated by the regression models for each site seem to provide the best representation of nitrate at the site and also the best data for comparison among sites.

Another advantage of using daily estimates of nitrate concentration is that they can be combined with daily streamflow and converted to mass loads (also referred to as flux, discharge, or export). The daily loads can then be summed to provide an estimate of annual load. The annual estimates are generally more accurate than the daily values because positive and negative errors in the daily values tend to cancel out in the annual summation. For the 72 NAWQA sites downstream from agricultural areas, mean annual nitrate loads were computed using the estimated daily nitrate concentrations for complete water years 1994 and 1995. The ratio of these loads to the upstream drainage area for each site was used to estimate mean annual nitrate yield, in mass per unit area.

Initially, nitrate at the 72 stream sites was compared to selected geographic characteristics of the upstream basins using bivariate correlation. Nitrate at the sites was represented by:

1. Flow-weighted mean concentration (of sample data or daily-value estimates), in milligrams per liter (mg L^-1),
2. Mean annual load, in megagrams (Mg), and
3. Mean annual yield, in megagrams per square kilometer (Mg km^-2).

The selected basin characteristics were:

1. Contributing drainage area of the basin, in km²,
2. Mean streamflow during water years 1993-95, computed from the daily-values data, in cubic meters per second (m³ s^-1),
3. Mean annual runoff, computed from mean streamflow and basin area, in centimeters per square kilometer (cm km^-2),
4. Mean annual nitrogen (N) input, computed as the sum of applied commercial fertilizer, applied manure, and atmospheric deposition, in Mg,
5. Mean N input rate, in Mg km^-2,
6. Cropland area, in percent of the total basin,
7. Population density, in number of people km^-2, and
8. Soil drainage category.

These characteristics represent some of the expected influences on nitrate mass and movement in a basin. The influence of the first 6 characteristics is obvious; all are related to basin size, water availability, or agricultural sources of nitrate. In addition, runoff provides a regional gradient because the amount of water per unit area that flows from a basin is large in the humid East and small in the arid West. Population density is an indication of nitrate sources from wastewater treatment plants or septic systems. These sources are expected to have less influence than agricultural sources of nitrate in rural areas, but are included in the analysis for completeness. Soil drainage has been shown to influence nitrate concentrations in ground water (Nolan et al 1997) and in streams during the late spring and early summer (Mueller et al 1997b). In this analysis, drainage in a mapped soil polygon was numerically categorized based on soil hydrologic group. Poorly drained soils (group D) were assigned a value of 1 and well-drained soils (group A) were assigned a value of 4. The soil drainage number for each basin was the area-weighted mean for all soil polygons in the basin.

Results of the correlation analysis are listed in table 1. Two coefficients are shown for each bivariate correlation: the Pearson coefficient indicates the degree of linear correlation and the Spearman coefficient indicates the degree of monotonic, but not necessarily linear, correlation. The highest correlation coefficients (at least 0.4) are shown in bold. The strongest correlations were between mean annual load and either mean annual N input or mean streamflow. This is certainly reasonable because the outflow load is a function of the available nitrate mass and the available transport medium (water). Both types of flow-weighted mean concentration were most strongly correlated to the N input rate and somewhat less strongly correlated to cropland area. Both of these basin characteristics indicate intensity of agriculture. Nitrate yield was moderately correlated to several basin characteristics, including runoff, N input rate, and population density. These three characteristics, as well as nitrate yield, are evaluated on a per area basis, so it seems logical that they would be correlated. N input rate and population density represent the primary nitrate sources in the basin, runoff is the transport medium, and nitrate yield is the downstream result.

The moderately strong negative correlation between nitrate yield and N input is unexpected, but might be numerically related to the nearly equivalent correlation between yield and basin size. Agricultural intensity is generally low in large basins compared to what it can be in small basins. Thus, even though the nitrate load downstream from a large basin might be high, the per-unit-area yield can be relatively low.

Soil drainage was not significantly correlated with any measure of nitrate at these NAWQA stream-sampling sites. However, all the measures of nitrate were on an annual basis. If nitrate had been evaluated on a seasonal basis, there might have been stronger correlations with soil drainage. For example, high nitrate concentrations during spring and summer might result primarily from storm events, which would be more likely to occur in areas conducive to surface runoff. Such conditions are common in agricultural areas with poorly drained soils, where tile drains or pipes are installed to enhance drainage. Conversely, high nitrate concentrations during periods of baseflow might be more likely in areas where nitrate concentrations are high in ground water. This condition is common in agricultural areas with well-drained soils on reasonably flat slopes.

Based on the correlation results, several multiple regression models were evaluated to determine the potential for predicting nitrate concentration, load, or yield using characteristics of the upstream drainage basin. Initially, the characteristics that were most strongly correlated with each measure of nitrate were included in the model. In some cases, characteristics were excluded to maintain independence of the explanatory variables. For example, basin area was excluded from the nitrate-load model because it was cross-correlated with streamflow, which was considered a better explanatory variable. After the initial calibration, explanatory variables that did not contribute significantly to a model’s fit were removed from that model.

Separate models of nitrate concentration were calibrated using the flow-weighted sample mean and the flow-weighted mean of daily estimates. N input rate and cropland area were significant explanatory variables in both models, and they fit the nitrate data about equally well. Results for the model fit using flow-weighted means of daily estimates are shown in figure 3. The 1:1 line represents a perfect fit of model estimates to ‘observed’ data. The standard error of the detransformed estimates is 2.1 mg/L, or 76 percent of the observed mean. This model tends to overestimate for sites with low mean nitrate concentration and underestimate for sites with high concentrations. This pattern is even more pronounced for the model fit using flow-weighted sample means.

The model of annual nitrate load (figure 4) included streamflow and mean N input as explanatory variables. The standard error is 0.56 Mg, or 80 percent of the mean ‘observed’ load. This error is similar to the concentration model, but the load model appears to fit the observed data better than the concentration model. However, the fit of the load model is strongly controlled by a few sites that had large loads. The four sites with the largest loads were all in Midwestern study units, so regional differences affecting sites with lower loads might not be apparent.

The model of annual nitrate yield (figure 5) appears to fit about as well as the load model. The standard error of the detransformed estimates is 0.71 Mg km^-2, or 75 percent of the mean ‘observed’ yield. It also has a cluster of highly influential sites with high yields, but these are located in several regions of the Nation. Therefore, this model might be more appropriate than the load model for national-scale application. The explanatory variables in the yield model are runoff and N input rate. Population density did not contribute significantly to the fit when these other variables were in the model.

Analysis of the nitrate and geographic data from streams draining agricultural basins suggests the following conclusions:

• NAWQA data can be used to estimate daily nitrate concentrations using periodic measurements of nitrate concentration and daily measurements of streamflow. A regression model including seasonality and time trend seems appropriate for many sites in a variety of locations around the Nation.
• Estimates from these models can be used to compute flow-weighted mean nitrate concentrations, nitrate loads, and nitrate yields. Correlations between these measures of nitrate and the geographic characteristics of upstream drainage basins are reasonably strong and can be explained logically.
• Development of regression models for predicting nitrate in streams as a function of upstream basin characteristics seems promising. The models presented in this paper were calibrated using data from only 72 sites in 18 of the 20 NAWQA study units. Model fit was adequate, particularly for prediction of nitrate yield from runoff and N input rate, but errors for some individual predictions were large.

These results are an initial attempt at synthesizing the NAWQA nutrient data on a national scale, and represent only a preliminary analysis. Using the daily estimates of nitrate concentration, it is possible to determine not only annual, but also seasonal nitrate loads and yields. Analysis of seasonal data might be useful in the assessment of management practices to limit nitrate impacts. When data are available from the second set of study units, expected by the spring of 1999, it might be possible to refine the models. In particular, data from additional study units will provide a better representation of the Nation and might be more adequate for determining regional differences in nitrate outflow in response to basin characteristics.

Future efforts of the NAWQA nutrient synthesis project will be to expand this analysis to include other nitrogen and phosphorus species. Results are not expected to be as good for these other species. Concentrations are generally lower than for nitrate and many measurements are less than detection, both of which can cause problems in fitting regression models. The analysis might also be used to evaluate nitrate at nonagricultural sites or at sites downstream from lower-intensity agriculture. This might be useful in determining nitrate concentrations that could be expected in the absence of major human influences.

[Coefficients greater than or equal to 0.4 are shown in bold type.]