Decadal surface water quality trends under variable climate, land use, and hydrogeochemical setting in Iowa, USA

Authors


Abstract

Understanding how nitrogen fluxes respond to changes in agriculture and climate is important for improving water quality. In the midwestern United States, expansion of corn cropping for ethanol production led to increasing N application rates in the 2000s during a period of extreme variability of annual precipitation. To examine the effects of these changes, surface water quality was analyzed in 10 major Iowa Rivers. Several decades of concentration and flow data were analyzed with a statistical method that provides internally consistent estimates of the concentration history and reveals flow-normalized trends that are independent of year-to-year streamflow variations. Flow-normalized concentrations of nitrate+nitrite-N decreased from 2000 to 2012 in all basins. To evaluate effects of annual discharge and N loading on these trends, multiple conceptual models were developed and calibrated to flow-weighted annual concentrations. The recent declining concentration trends can be attributed to both very high and very low discharge in the 2000s and to the long (e.g., 8 year) subsurface residence times in some basins. Dilution of N and depletion of stored N occurs in years with high discharge. Reduced N transport and increased N storage occurs in low-discharge years. Central Iowa basins showed the greatest reduction in flow-normalized concentrations, likely because of smaller storage volumes and shorter residence times. Effects of land-use changes on the water quality of major Iowa Rivers may not be noticeable for years or decades in peripheral basins of Iowa, and may be obscured in the central basins where extreme flows strongly affect annual concentration trends.

1 Introduction

Abundant nitrogen (N) in surface waters can have harmful effects on human and environmental health. Agriculture is a primary source of excess N. In the United States (US), the intensification of agriculture has occurred in recent years as a result of food demands and promotion of biofuels crops to meet energy needs [US Congress, 2007]. Sustainable growth of agriculture requires evaluation of long-term effects of practices and environmental factors such as precipitation on N contamination. In this study, we examine long-term and recent trends in N and relate those changes to hydrogeochemical features of basins and conceptual models of responses to changing discharges and N-fertilization.

In the last decade, corn production has increased in the midwestern United States to satisfy demands for production of ethanol, now the most readily available alternative fuel in the US. Ethanol production in Iowa has increased over 700% from about 1.5 billion liters in 2003 to 14 billion liters in 2011 [Hart et al., 2012]. As corn prices increase, some of the most productive land has been continuously planted in corn in contrast to a corn-soybean rotation [Secchi et al., 2011], and millions of acres of uncultivated lands have been brought into production [Cox and Rundquist, 2013]. A larger percentage of row-crop agriculture in a watershed has been shown to correlate with greater N concentrations in rivers and streams [Schilling and Libra, 2000; Shilling and Spooner, 2006]. Recent studies indicate that increased corn ethanol production can increase export of N to ocean ecosystems and cause eutrophication of surface waters [Donner and Kucharic, 2008; Yang et al., 2012]. Whether specific conservation practices or environmental factors will modify the response of surface waters to increased corn planting is a critical question for midwestern states and other areas affected by agricultural intensification.

The state of Iowa is an important study area for the effects of agricultural intensification because of its long agricultural history, influence on regional water quality, and diverse hydrogeochemical landscape. Starting in the 1960s and 1970s, N concentrations and fluxes, primarily in the form of nitrate ( math formula) increased substantially in major Iowa basins such as the Raccoon River [Hatfield et al., 2009], Des Moines River [McIsaac and Libra, 2003], and Iowa River [IADNR, 2001]. The increasing concentrations coincided with increases in N-fertilizer applications. In a comprehensive study of fluxes and sources of nutrients in the Mississippi-Atchafalaya River Basin, Goolsby et al. [1999] estimated math formula yields from Iowa to be among the highest in the nation. More recently, Alexander et al. [2007] estimated that total flux of math formula from Iowa is the second highest of the 24 states contributing math formula to the Gulf of Mexico via the Mississippi and Atchafalaya River Basins. Iowa has a diverse hydrogeochemical landscape, with relatively young and reactive sediments as well as older, less reactive, and more highly drained sediments, both of which are representative of large portions of farmland in the United States and elsewhere. The region has also seen considerable variability of precipitation and runoff in recent years, which can contribute to short-term and long-term trends in labile nutrients [Mengistu et al., 2013]. Better understanding of the response of these Iowa basins to intensification of agriculture and climatic variations will have broad implications for global agricultural production and water quality.

Whether the increase in corn production during the past decade has affected the water quality of Iowa's major rivers is an open question. A study of trends in N concentration and flux in Mississippi River drainages between 1980 and 2008 indicated mixed results, with decreasing trends of N concentration and flux in the Iowa River at Wapello and increasing trends in the Missouri River at Hermman, MO and the Mississippi River at Clinton [Sprague et al., 2011]. A time series analysis of 1998–2012 data from 46 river monitoring sites in Iowa found that most rivers did not have statistically significant trends of math formula concentrations, but found increasing trends in some western basins and a slight increasing trend (0.05 mg-N L−1 yr−1) in the average across all rivers [Li et al., 2013]. A study of the Raccoon River reported no trend in math formula concentrations from 1992 to 2008 [Jayasinghe et al., 2012]. The M.P.C.A. [2013] study of math formula in Minnesota streams from 1976 to 2000 found increasing and decreasing trends of flow-adjusted concentrations in rivers near the northern border of Iowa. Differences of water quality trends can result from variable time lags between land use changes and surface water responses [Tesoriero et al., 2013]. Schilling and Spooner [2006] showed that the water quality in approximately 5000 ha watersheds responded to land use changes after about 3 years. Lag times can potentially be much greater because most math formula is exported via base flow [Schilling, 2002] with residence times on the order of decades [Tomer and Burkart, 2003; Schilling and Spooner, 2006; Schilling et al., 2012]. In contrast to the conceptual model of land-use-driven changes, Basu et al. [2010] concluded that annual nutrient exports from agricultural basins are controled mainly by discharge. They suggested that the legacy of accumulated nutrient sources is so large that the system is transport limited rather than supply limited. As a result, variations in flows and fertilizer inputs no longer affect the annual flow-weighted concentration, defined as the annual load divided by the annual discharge [Basu et al., 2010]. Similar invariance of annual concentrations was reported for a small (<500 km2) tile-drained watershed [Guan et al., 2011]. The applicability of this conceptual model to Iowa subbasins spanning a variable hydrogeochemical landscape has not been explored in detail.

The objective of this study is to examine the effects on water quality of recent agricultural and hydrologic changes in 10 of Iowa's major rivers over the period of 1970–2012, with emphasis on the last decade of increasing corn production. Annual nitrate plus nitrite (N) concentrations were estimated using the Weighted Regression on Time, Discharge, and Season Model (WRTDS) [Hirsch et al., 2010]. This model allows regression coefficients to change with time. Trends were analyzed by (1) removing effects of random variations of discharge in WRTDS, and (2) using a multimodel analysis with 10 conceptual models of effects of annual flow and annual N loads on flow-weighted mean annual N concentrations. Model probabilities were calculated to determine the factors that most likely explain the historical and recent trends in N concentrations. Results were compared with the physical and chemical attributes of the basins to determine how the factors of discharge and N-loading interact with the hydrogeochemical landscape to produce the observed trends in N.

2 Study Area

The study area covers the watersheds of 10 major rivers in Iowa comprising 63% of the area of the state (Figure 1). The sizes of the study basins range from a low of 2252 km2 for the Boyer River to a high of 32,365 km2 for the Iowa River. Water drains from Iowa in rivers that flow to the Missouri River in the west and to the Mississippi River in the central and eastern parts of the State. Drainage ditches and subsurface tiles were constructed in the early 1900s to remove water from north-central Iowa. Installation of additional tile in this and other areas of the State have continued to the present.

Figure 1.

Watersheds of major Iowa tributaries to the Mississippi and Missouri rivers included in this study.

Two of the rivers used in this study drain north-central Iowa. The Des Moines and the Raccoon River upstream of Des Moines drain the youngest and generally flattest area of Iowa. These two rivers drain about 23,000 km2 in the area known as the Des Moines Lobe that was glaciated about 11,000–14,000 years before present (Figure 1). The glacier left an undulating surface of till material with poor surficial drainage. The average slope of these two watersheds is <2.0 cm m−1 (Table 1). Unweathered till in the Des Moines Lobe tends to have higher concentrations of organic and inorganic electronic donors, lower O2, and higher rates of denitrification than surrounding sediments [Rodvang and Simpkins, 2001], as well as a greater intensity of tile drainage.

Table 1. Characteristics of River Drainage Basins in Iowaa
Site NumberSite NameArea (km2)Mean Annual Precipb (m)Mean Annual Runoffc (m)Farm Area (%)Mean Soil Kd (cm h−1)Mean Clay Contente (%)Mean Sand Content (%)Mean Slope (cm m−1)
  1. a

    Values are from Falcone et al. [2010].

  2. b

    Precipitation as rain.

  3. c

    Defined as annual discharge divided by basin area.

  4. d

    Hydraulic conductivity.

  5. e

    Percentage by volume.

Mississippi River Drainages
05418500Maquoketa40200.890.29765.824223.3
05422000Wapsipinicon60500.900.32817.723321.6
05465500Iowa River32,4000.880.29795.426231.6
05474000Skunk11,2000.900.28773.030142.1
05481300Des Moines14,1000.790.22864.126300.7
05484500Raccoon88700.830.23873.827271.5
Missouri River Drainages
06607500Little Sioux91200.760.19843.827162.0
06609500Boyer22500.800.19853.42654.1
06810000Nishnabotna72800.860.22893.02963.6
06817700Nodaway39300.900.22852.33393.5

The four major Missouri River tributaries used for the study, Little Sioux, Boyer, Nishnabotna, and Nodaway Rivers, drain a total of about 22,600 km2 of western Iowa (Figure 1). The Little Sioux has headwaters in the Des Moines lobe and has relatively low slope (2.0 cm m−1) as compared to the other three basins (3.5–4.1 cm m−1). These rivers generally flow to the south-southwest through areas covered by a thick layer of loess material. Soils that developed on the loess are composed mainly of silt-sized particles with generally <10% sand content. Erosion of the loess and underlying till material has created a landscape of relatively large topographic relief in the southern basins.

The four major rivers in eastern Iowa used in the analysis were the Maquoketa, Wapsipinicon, Iowa, and Skunk Rivers. These drain about 76,500 km2 of eastern and central Iowa to the Mississippi River (Figure 1). The Iowa and Skunk rivers have headwaters in the Des Moines lobe. The Mississippi River tributaries generally flow to the southeast into areas with a combination of till and thin loess deposits in southern and southeastern Iowa. Soils developed on till material are composed mainly of silt-sized particles but contain substantially more sand-sized material (about 20%–30%) than the loess soils in the Missouri River watersheds. The slopes of the Mississippi River watersheds (1.6–3.3 cm m−1) are greater than those of the Des Moines lobe watersheds and are typically slightly less than the Missouri river water sheds (Table 1).

3 Methods of Estimating Land Use and Water Quality Changes

3.1 Crop and Fertilizer Estimates

In this study, we used N fertilization rate estimates as indicators of N-application intensities in the basins. Fertilizer N estimates were used because farm fertilizer is the largest N source in the Mississippi/Atchafalaya Basin [Goolsby et al., 1999; David et al., 2010; Robertson and Saad, 2013; M.P.C.A., 2013], application rates of N-fertilizer have been shown to correlate with water quality responses [Puckett et al., 2011; Liao et al., 2012], and fertilizer application rates are estimable from established data sources, as described below. In the study area, the majority of N-fertilization is for corn cultivation.

Time series of application rates of N fertilizer were estimated for each basin. Rates of N fertilizer application in each county were estimated for the period 1964–2012 by multiplying county-level estimates of soybeans and corn acreage for each year (http://www.nass.usda.gov/Data_and_Statistics/Quick_Stats/index.asp) with state-level estimates of annual N-fertilization rates for those crops in that year (http://www.ers.usda.gov/Data/FertilizerUse/). Data gaps were filled using linear interpolation. County-level application rates for 1954–1964 were taken from previous estimates based on sales data [Alexander and Smith, 1990]. Annual application rates for each basin were estimated by summing the application rate for each county multiplied by the fraction of the county area located in that basin.

Information on the conversion of conservation reserve acres (CRP) was obtained from USDA Farm Service Agency (http://www.fsa.usda.gov/). Two sources of land use data were compared to determine the best approach for estimating cropping patterns for individual watersheds. The first was the Crop Data Layer (http://www.nass.usda.gov/research/Cropland/SARS1a.htm), which is available beginning in 2000 for Iowa, but not for adjacent states. Corn area in watersheds within Iowa was determined by summing the rasterized pixels of corn land in each watershed. The second source of data was the National Agricultural Statistics Service (http://www.nass.usda.gov/), which has published crop area by county since 1964. Cropping patterns for individual watersheds were calculated based on the proportion of the county within each watershed. The results of the two methods were compared from 2001 to 2010 for the six watersheds lying completely within Iowa to evaluate potential biases and errors. Differences between the two methods averaged <3% for all but the Boyer, for which NASS results averaged 9.6% more than CDL values. The larger difference in the Boyer Basin likely relates to its small size and less uniform cropping pattern. On the basis of these comparisons, the longer time record available, and the availability of data outside Iowa, the NASS data were used to describe land use changes in the studied watersheds.

To evaluate the estimates of N-fertilization, we recalculated N-fertilization rates using county sales data as compiled and analyzed by Ruddy et al. [2006]. The shapes of the sales-based estimates were very similar to those derived from crop acres (R2 ranging from 0.89 to 0.97, average R2 of 0.94). The crop-acreage based estimates were used for the analyses in this study because these estimates are based on fertilizer use for existing crops rather than sold fertilizer that can be applied at a different locations or times.

3.2 Water Quality Data

Nitrate ( math formula) plus nitrite ( math formula) concentration data were selected from sites where consistent sampling techniques were used (Table S1). Because math formula concentrations are typically much greater than math formula in aerobic environments, these data are hereafter referred to as math formula concentrations. Data for the Des Moines River near Stratford and Raccoon River at Van Meter have been collected since the 1970s by Iowa State University under contract with the U.S. Army Corps of Engineers [Lutz, 2011] as part of a program to monitor water quality in rivers flowing into several large flood control reservoirs that may remove substantial amounts of nitrogen from the river [Schoch et al., 2009]. For the other eight sites, early water quality data (1980 to mid-1990s) were mainly collected by the Iowa Department of Natural Resources (IADNR), Iowa Department of Environmental Protection (IADEP), and the U.S. Geological Survey (USGS). Data collected since 2000 were from the IADNR as part of their statewide monitoring network and the USGS as part of the National Water Quality Assessment (NAWQA) and National Stream Quality Accounting Network (NASQAN) programs and as part of a cooperative project with the IADNR. Water quality data were retrieved from the U.S. Environmental Protection Agency (EPA) STORET database, the U.S. Geological Survey NWIS database, and the Des Moines River Water Quality Network database maintained by Iowa State University.

Nitrogen compounds most commonly associated with agricultural crop production were used as indicators of changes in water quality conditions. Dissolved math formula comprises most of the nitrogen in midwestern rivers [Goolsby et al., 2001]. Because math formula is generally well mixed in the water column, both depth and width-integrated samples and dip samples were used for analysis. Data from analysis of unfiltered samples were used when dissolved data were not available. This approach was evaluated by comparing 37 samples of filtered and unfiltered water collected at the same time from the Iowa River. The filtered and unfiltered samples had similar concentrations of math formula (R2 = 0.96, mean absolute difference = 0.2 mg-N L−1).

3.3 Estimates of Annual Concentration and Flux

This study uses estimates of annual mean flow-weighted and flow-normalized concentrations (described later in this section) that cannot be obtained directly from intermittent samples. The records used here had an average of 12.8 math formula samples per year, and the times of sample collections were intentionally biased toward greater frequency during periods of high discharge. Discharge measurements were at daily intervals. Because math formula concentrations can vary rapidly in response to changes in discharge, daily estimates of concentration are needed to estimate of quantities such as the total N load and the annual mean flow-weighted concentration, which is equal to the total N load divided by the total discharge. In this study, daily concentrations and fluxes were estimated from periodic samples and daily streamflow using the weighted regression on time, discharge, and season (WRTDS) model [Hirsch et al., 2010]. WRTDS has been applied previously in studies of the Mississippi River, Lake Champlain, and Chesapeake Bay watersheds [Sprague et al., 2011; Medalie et al., 2012; Zhang et al., 2013]. WRTDS allows parameters to vary as functions of time and discharge (discussed below), which gives flexibility of the modeled response of solute concentrations to discharge and avoids potential biases that can occur in regression methods with time-constant parameters [Stenback et al., 2011; Hirsch, 2014]. The flexibility of parameters in this method as compared to other interpolation tools makes it well suited for application to relatively long-term records (e.g., approximately 20 years) with sampling frequency on the order of 1 or 2 samples per month (which is typical in these data sets).

Concentration is modeled in WRTDS as:

display math(1)

where ln is natural log, c is concentration, βi are fitted coefficients, Q is daily mean streamflow, t is decimal time, and ε is the unexplained variation. Daily fluxes are estimated by multiplying daily concentration estimates by the corresponding daily mean streamflow. A unique set of coefficients can be estimated for any combination of Q and t in the period of record using weighted regression. For a particular (Q, t), the parameters of equation (1) are estimated using a subset of observations that is selected and weighted based on the distance in time, streamflow, and season between the observations and (Q, t). Observation weights vary according to a Tukey tricube weight function [Hirsch et al., 2010], decreasing from 1 at the exact time, streamflow, and season of (Q,t) to zero outside of specified half-window widths. For this study, half-window widths were 10 years for the time, two natural-log cycles for the discharge, and 0.5 years for the season. WRTDS specifies that >100 observations must be available for each estimation of parameters at any (Q,t) and if there are fewer than 100 observations the windows are widened to the point that the number available exceeds 100. Figure S1a shows an example of the relationship between observed math formula concentrations and WRTDS estimated math formula concentrations for the Iowa River. The points cluster around the 1:1 line (mean error = −0.2 mg-N L−1) with most errors <2 mg-N/L (root-mean-square error = 1.6 mg-N L−1, mean absolute error = 1.2 mg-N L−1). The highest concentrations tend to be slightly underestimated by WRTDS in this example, although the differences are generally small relative to the concentrations being measured (mean absolute error = 20% for the upper quartile of math formula concentrations). Figure S1b shows the ratio of observed to predicted math formula plotted against discharge, indicating that errors are independent of, and not biased by, the discharge. The Nash-Sutcliffe efficiency (NSE) for the Iowa WRTDs model was 0.68. The NSE coefficient indicates how closely the plot of observed versus simulated values matches the 1:1 line. The coefficient can vary from −∞ to 1, with a value of 1 indicating a perfect fit and zero indicating that the mean of the observed values performs as well as the model. NSE values for the other models were all greater than zero, ranging from 0.43 for the Des Moines River to 0.75 for the Raccoon River (Table S2).

Because variations in streamflow strongly affect measured concentrations as well as the estimates from equation (1), WRTDS also computes flow-normalized (FN) estimates, which are designed to remove variations in concentration or flux caused by year-to-year streamflow variations from the seasonal norms. The flow normalization method assumes that streamflow on any given day of the record is one sample from the probability distribution of streamflows for that particular day of the year. To compute the FN concentration, first equation (1) is used to estimate n concentrations on that date with the streamflow value, Q, set to each one of the n streamflow values for that day of the year in the entire record. The FN concentration for that date is then calculated as the mean of the n estimated concentration values. Similarly, the FN flux is the mean of the estimated flux values from each of those n weighted regressions. For more detail on the model, see Hirsch et al. [2010] and Sprague et al. [2011]. Daily estimates were summarized into water-year annual means (for concentration) and water-year annual totals (for flux) for a water year (October through September of the following year), which roughly corresponds to a crop year. In addition, annual mean flow-weighted concentrations were calculated by dividing the annual total N flux (from WRTDS) by the annual discharge as done, e.g., by Basu et al. [2010] and M.P.C.A. [2013]. The analysis of annual mean flow-weighted concentrations is presented in section 'Modeling of Water Quality Response to Fertilization and Discharge'. We also report values of runoff, which is equal to the total discharge divided by basin area.

4 Results of Land Use and Water Quality Analyses

4.1 Land Use Changes

Notable land use changes in Iowa from 2000 to 2012 included increasing trends in acreage of corn crops and reports of increases in tile drainage. Corn production in Iowa generally remained in the 4–5 million hectare range from 1920 to 1970. During the decade of the 1970s production increased steadily, peaking at >5.8 million hectares in 1981 (Figure 2a). Programs to encourage farmers to reduce production resulted in a continual decline in corn production from 1981 to 2000. Beginning in 2000 use of corn to produce ethanol increased demand, resulting in a state-wide expansion of production by 9% to a total of approximately 5.4 million hectares by 2010. Because most land is already used for agriculture in Iowa, increases in corn production came mainly at the expense of soybean production (Figure 2a). Additional land was converted from the conservation reserve program (CRP) beginning in 2008 (Figure S2). In the 10 study basins, total corn production in 2010 was 3.8 million hectares comprising 70% of the area planted in corn in Iowa. Overall the increase in these basins from 2000 to 2010 averaged 12%, ranging from a low of 5% in the Des Moines Basin to a high of 17% in the Maquoketa Basin (Figure 2b). Figure 3 shows application rates of N fertilizer in kg per ha of basin area for the period 1970–2010 as 5 year averages. Estimated application rates in the Nodaway basin were relatively low because the acreage of corn, the primary recipient of N-fertilizer, was a smaller fraction of the total area than in other basins. The pattern of changes is similar across all basins. Rates decreased in the 1980s and early 1990s, but have increased since the late 1990s, with particularly steep increases during the last decade.

Figure 2.

(a) Area of Iowa planted in grain crops, 1926–2012 (Data from USDA, 2013, Data and Statistics, U.S. Department of Agriculture National Agricultural Statistics Service, Available at http://www.nass.usda.gov/.). (b) Percent change in corn acreage in the 10 study basins computed using average area 2009 and 2010 minus average area in 2000 and 2001 for each basin.

Figure 3.

Five year running-average fertilizer N application rates for river basins of Iowa.

Tile drainage has been identified as a factor strongly affecting N fluxes in the midwestern United States [David et al., 2010; M.P.C.A., 2013], Changes in land use during the 2000s likely include increased tile drainage, although these changes are difficult to quantify because comprehensive records of tile drainage extents in Iowa do not exist. Expansion of drainage in the mid-20th century was invoked as a possible explanation of altered response of basins to precipitation, causing an increased fraction of flow as base flow [Schilling and Libra, 2003]. In the 2000s, data are available for the numbers of “wetlands determinations” to facilitate drain installations. Requests for these determinations went up in the 2000s as compared to previous decades (D. Carrington, National Resources Conservation Service, U.S. Department of Agriculture, unpublished data, 2013). In addition, industry and news-media reports indicate that tile drainage installations in Iowa basins have been substantial [Swoboda, 2010; Williams, 2012; AP, 2013; Pates, 2013], likely motivated by economic incentives of increased crop yields [Hofstrand, 2012]. As new tile drains are installed to reduce spacing between existing drains [Pates, 2013], increases in tile drainage can occur in areas that are already extensively drained, such as in the Des Moines lobe.

4.2 Water Quality Trends

Figure 4 shows the WRTDS estimates for concentrations of math formula for the 10 rivers during the study period. The plotted points are estimated annual mean values, which can be highly variable due to variations in annual discharges. Precipitation and subsequent stream discharge have a substantial impact on the movement of N from the landscape to rivers, causing changes in concentrations and flux. For example, low discharges around 2000 and high discharges around 2010 tend to create a general trend of increasing annual mean (uncorrected for discharge) concentrations. Similar positive trends were observed by Li et al. [2013] for some basins in Western Iowa based on time series analysis of raw math formula data. In contrast, the smooth curves in Figure 4 are flow-normalized (FN) concentration values computed as described in section 'Estimates of Annual Concentration and Flux'. The FN estimates are designed to correct for the effect of year-to-year variations of daily discharge from the seasonal norm and thus remove effects of the daily discharge on the inferred trends. Any remaining trends after removal of daily discharge effects may reflect changes in longer-term hydrology, such as sustained high or low flows, or changes in watershed characteristics such as crop types, fertilizer application rates, or tile drainage. These factors are discussed in greater detail below and in section 'Modeling of Water Quality Response to Fertilization and Discharge'.

Figure 4.

Trends in estimated annual mean concentration (points) and flow-normalized concentration of math formula (black line) and runoff (blue line; annual discharge divided by basin area) in 10 river basins in Iowa.

Because of the strong link between discharge and concentrations, we analyzed the discharge histories of the rivers. Large amounts of interannual variability of discharge have been the norm for these Iowa Rivers, and variability has increased over the period of this study. Figure 5 shows 5 year running mean values of annual runoff (discharge divided by basin area) for each of the 10 rivers in this study along with the ensemble average for all rivers. The 5 year mean serves as a low-pass filter to remove interannual variability, as used with precipitation time series for the Midwest by White et al. [2008]. In general, runoff continued to increase from 1970 to 2012, beyond the previously observed increases in Iowa Rivers [Schilling and Libra, 2003]. At the same time, the 5 year mean runoff of these Iowa Rivers suggests a quasiperiodic variation with a characteristic time scale of approximately 12 years (Figure 5) during the time frame of this study. The greatest variations of discharge occurred in the 2000s. This result is consistent with a study by Johnson et al. [2009] who found increasing discharge and variability of discharge for the Minnesota River from 1970 to 2003. Variability of precipitation may be affected by tropical Pacific Ocean temperatures [White et al., 2008], although the historical response of discharge to precipitation has changed over the last century, with increasing proportions of base flow in most rivers [Schilling and Libra, 2003]. Implications of this changing climatic factor for water quality are discussed below.

Figure 5.

Running 5 year mean runoff (discharge divided by basin area) for all basins (blue lines) and for the average of the 10 basins (black line).

Annual N fluxes are closely linked to annual discharge. As a result of the increasing variability of discharge in the 2000s, N fluxes also varied widely. During the dry year of 2006, 45.7 Gg of N were transported to the Mississippi river by the Maquoketa, Wapsipinicon, Iowa, and Skunk rivers, and 13.7 Gg were transported to the Missouri by the Little Sioux, Boyer, Nishnabotna, and Nodaway. In contrast, during the wet year of 2010 the same streams transported 176.3 Gg to the Mississippi and 54.3 Gg to the Missouri. Interestingly, although the discharges in the 2000s tended to be greater than in the 1980s, the N loads are comparable for most basins, as a result of slightly lower concentrations in the 2000s (Figure 4). Because of the large variability in annual averages, long-term trends are difficult to discern from those observations. The remainder of this analysis focuses instead on trends of flow-normalized concentrations from WRTDS (below) and comparisons of multiple conceptual models of effects of discharge and N inputs on annual variations (section 5).

Water quality trends were similar to trends of N-fertilization in these watersheds for most of the study period. The decadal trends in flow-normalized N from 1970 to 2000 were similar to changes in land use and fertilizer applications. Figure 6 shows changes of FN concentration during selected periods. Trends were quantified as a rate of change using (ct2 − ct1)/n where n is t2 − t1, ct1 is the annual mean concentration in year t1 and ct2 is the annual mean concentration in year t2. For streams with truncated time series of FN concentrations during a period of interest, the earliest and latest available years were used for t1 and t2, respectively, as indicated in Figure 6. From 1970 to 2000, FN concentrations tended to increase or decrease in conjunction with corn production and fertilizer applications. For example, in the 1970s there was a rapid expansion of corn production (Figure 2) and fertilizer use (Figure 3), and increasing concentrations in streams were also observed (Figure 6). Data are available for five basins to assess trends in N during this decade (Figure 4 and Table S1). There were sharp increases in concentration for the Maquoketa, Wapsipinicon, Des Moines, and Raccoon. The Little Sioux had slight increases over this decade. From 1980 to 2000, concentrations of N changed only slightly in the five basins for which data are available (Figure 6), consistent with relatively small changes of N applications between 1980 and 2000 (Figure 3).

Figure 6.

Trends of flow-normalized (FN) concentration of math formula in the 10 study rivers. For each time period, trends are shown for basins with five or more estimated flow-normalized annual concentrations per decade (having >4 estimates in 1970–1980; having >4 in 1980–1990 and >4 in 1990–2000 in the period of 1980–2000; and having >5 in 2000–2012). Numbers after the name of a river are the last two digits of the starting and ending dates used to estimate the math formula trend with the equation (ct2 − ct1)/n where n is t2 − t1, ct1 is the annual FN concentration in year t1 and ct2 is the annual FN concentration in year t2. The time series of estimated annual FN concentrations are shown in Figure 4.

After 2000, effects of N applications on water quality are less apparent. Between 2000 and 2012, trends of corn acreage and N-fertilizer applications were positive in all basins (Figures 2 and 3), but the changes of flow-normalized N concentrations were negative in all streams (Figure 6). This difference of trends may be affected by transit times between fields and streams. In other words, the declining trends in the 2000s may reflect declining application rates in previous decades. In addition, the 2000s were a period of unusual flow variability, with sustained low discharges from 2000 to 2006 followed by record high discharges from 2007 to 2011, as described above. Sustained high discharges could cause lower than expected concentrations for a given season and daily discharge as a result of dilution of the available pool of N. Also, concentrations of N in 1 year can possibly be affected by discharges in previous years that may tend to deplete (in high flow years) or allow accumulation of (in low flow years) the available pool of N. These potential effects are explored further below and in section 'Modeling of Water Quality Response to Fertilization and Discharge'.

A qualitative analysis shows that discharge-dependent concentration responses are complex, varying among rivers and varying over time in individual rivers. Figure 7 shows annual mean flow-weighted concentrations (annual total N flux from WRTDS divided by annual discharge), color coded by year, along with annual runoff. Potential effects of dilution are apparent in some basins where flow-weighted concentrations decrease at high discharges (e.g., Iowa, Skunk, Des Moines, Raccooon, Nishnabotna, and Nodaway). In other basins, flow-weighted concentrations appear to be relatively unaffected by discharge, with comparable concentrations at high and low discharge (e.g., Boyer) or concentrations that increase with time more-so than with discharge (e.g., Maquoketa and Wapsipinicon). Because of the complexity of responses and variations among rivers, a multimodel approach was implemented, as described below, to quantify the potential effects of hydrologic and anthropogenic factors on math formula concentration trends.

Figure 7.

Annual mean flow-weighted math formula concentration versus runoff for 10 Iowa rivers. Example best fit models are shown for the Raccoon River for model scenarios 2 (gpow(Q)), 3 (gerf(Q)), and 4 (gerf2(Q); Table 2). For Model 2, the exponent of the power-law model (B parameter) is 0.15.

5 Modeling of Water Quality Response to Fertilization and Discharge

5.1 Multimodeling Approach

Multiple conceptual models were developed to test the potential effects of fertilizer-N applications and discharge on annual math formula concentrations over the full period of record for each basin (Table S1). Models that included effects of N-fertilizer applications, current annual discharge, and preceding discharge used the general form

display math(2)

where A is a fitted coefficient, math formula is the annual mean flow-weighted concentration computed as described in section 'Estimates of Annual Concentration and Flux', f(Napp) is a function of the fertilizer N applied in the basin, g(Q) is a function of the annual discharge, and h(Qpre) is a function of the preceding discharge. Effects of fertilizer inputs were estimated with

display math(3)

where Napp is the 5 year running average of N-fertilizer application in a basin (Figure 3), t is the time (year) and τ is the mean transport time between application and arrival in the stream. Values of the transport time, τ, were estimated through model calibration, as described below. For scenarios excluding effects of fertilizer inputs, f(Napp) was set equal to 1.

Three different equations were tested for the effects of current-year discharge on annual concentrations,

display math(4a)
display math(4b)

or

display math(4c)

where B and C are fitted coefficients, C > B, and erf is the error function. The function gpow(Q) (equation (4a)) is a commonly used power-law function to estimate concentrations from discharge [e.g., Godsey et al., 2009; Basu et al., 2010]. Equation (4a) estimates no upper limit to increasing concentration with discharge, except in the case of B = 0, which estimates constant concentration. The function gerf(Q) gives concentration curves that increase approximately linearly with Q for Q/B < 0.5 and then transition to an upper asymptote at Q/B ≈ 2 (equation (4b)). Decreasing concentrations at greater Q are estimated by gerf2(Q) (equation (4c)), which approximates the effects of dilution and depletion of available N at greater discharges, e.g., as observed by Donner and Scavia [2007] for monthly estimates in the Mississippi River and by others for individual storm events [e.g., McDiffett et al., 1989]. Examples of these three models are shown for the Raccoon River in Figure 7.

The h(Qpre) function in equation (2) estimates the effects of previous years' discharge on the current year's concentration. For example, high discharge may deplete the available pool of N stored in a watershed resulting in lower-than-expected concentrations in following years, or conversely the available pool of N may increase in years with low discharge. Such “chemical memory” mechanisms have been suggested, for example, by Ruiz et al. [2002] based on estimates of annual N budgets in small catchments, by Kirchner et al. [2000] based on power spectra of concentrations, and by Acker et al. [2005] based on remote sensing of phytoplankton in the Chesapeake Bay. Similarly, Donner and Scavia [2007] found that a rainfall-hypoxia model for the Gulf of Mexico was improved by including the previous year's N flux as proxy for changes in N storage. Lucey and Goolsby [1993] noted that elevated math formula concentrations and fluxes occurred during a year of moderate precipitation following 2 years of below-normal precipitation in the Raccoon River. The effect of discharge in preceding years was estimated using a weighted average of the preceding years. Various weighting schemes were tested and a scheme was selected that has a linear decrease in weights for the previous 5 years:

display math(5)

where D is a fitted coefficient and Qt1 to Qt5 are the annual discharges for the preceding 5 years. As implemented in equation (2), a positive value of D indicates that the flow-weighted concentration tends to be negatively correlated with the previous-years' discharge.

Ten models were constructed to test individual and combined effects of Q, Qpre, and Napp using equations (2–5) (Table 2). Parameters of each model were calibrated to minimize the sum of square errors between the model estimates of math formula and the “observed” values for each river. The “observed” values of math formula here are the estimates obtained from the WRTDS model based on field measurements as described in section 'Estimates of Annual Concentration and Flux'. The 10 models for each basin were compared on the basis of posterior model probabilities estimated using Akaiki Information Criteria with correction for sample size (AICc). Error variance of observations was estimated to be 1 mg L−1 and this value was used in all calibrations and calculations of AICc. Model probabilities based on AICc give the likelihood that a particular model in a group of models is supported by the data based on the discrepancy between simulated and observed values and the number of parameters. More detailed information on use of AICc to calculate posterior model probabilities is available in Poeter and Anderson [2005].

Table 2. Model Scenarios for Estimating Annual Flow-Weighted Average Concentrations of Nitrate
ScenarioEquationFactorsCalibrated Parameters
1 math formulaf(Napp)A, τ
2 math formulagpow(Q)A, B
3 math formulagerf(Q)A, B
4 math formulagerf2(Q)A, B, C
5 math formulagerf(Q), h(Qpre)A, B, D
6 math formulagerf2(Q), h(Qpre)A, B, C, D
7 math formulagerf(Q), f(Napp)A, B, τ
8 math formulagerf2(Q), f(Napp)A, B, C, τ
9 math formulagerf(Q), h(Qpre), f(Napp)A, B, D, τ
10 math formulagerf2(Q), h(Qpre), f(Napp)A, B, C, D, τ

5.2 Relationships of Concentrations to Discharge and N Applications

Estimates of posterior model probabilities for the 10 models for all basins are shown in Figure 8, with greater probabilities highlighted in red. A greater posterior model probability indicates that the parameterization and structure of a model are better justified by the data than an alternative model with lower probability. The most-probable models included upper limits to concentration at high discharge (error-function equations (4b) and (4c)), effects of preceding years discharge, Qpre, and effects of N-fertilizer inputs. All 10 basins had most-probable models that included error-function-based responses of math formula to Q, with an upper asymptote to math formula (Wapsipinicon, Iowa, Des Moines, Little Sioux, Boyer, and Nishnabotna) or with decreasing math formula at high discharges (Maquoketa, Skunk, Raccoon, and Nodaway). As shown on Figure 7, the error-function-based models were more similar to the typical observations of linearly increasing math formula at low discharges and constant or decreasing math formula at high discharges. The power-law function model, gpow(Q) was not a most probable model for any river and always had probability <0.1. The power-law model tended to overestimate concentrations in years with extreme high and low flows, and to under-estimate concentrations for years with near-median discharge. Model fits of the highest-likelihood models were generally acceptable, with maximum Nash-Sutcliffe efficiency (NSE) coefficients ranging from 0.51 to 0.95 in nine rivers, and with lower values (0.25) for the Boyer where concentrations tended to be relatively flat with respect to discharge (Figure 7 and Table S3).

Figure 8.

Model posterior probabilities for all rivers and scenarios. Darker red shading indicates greater probabilities. Probabilities >0.05 are highlighted in bold.

The effects of Qpre potentially explain apparent decreasing trends in annual mean concentrations (Figure 4), and indicate that the pool of available N is dynamic, can be depleted during years with high discharge, and can accumulate during years with low water flux. As a result of the unusually high discharge variability in the most recent decade of this study (Figure 5) effects of preceding-years' discharge were apparent in many basins. Six basins (Wapsipinicon, Skunk, Raccoon, Little Sioux, Boyer, and Nodaway) had maximum likelihood models that included effects of h(Qpre). All 10 basins had at least one model with probability >0.1 that included effects of Qpre. These results indicate that annual concentrations tended to be lower than expected following a high-discharge year, as compared to a prediction based only on current-discharge conditions. For example, Figure 9 shows “observed” flow-weighted N (estimated by WRTDS) along with best fit estimates from Model 7 (with effects of gerf(Q) and f(N)) and Model 9 (adding effects of h(Qpre) to Model 7, see Table 2) for the Wapsipinicon River. Accounting for current year discharges in Model 7 resulted in a relatively flat concentration time series from 2006 to 2012 that did not match the decreasing flow-weighted concentration trends. Model 9, with effects of current and previous years' discharge, gave a closer match to the pattern of decreasing concentrations following years with high discharge (e.g., 2010) and increasing concentration following years with low discharge (e.g., 2005), and yielded a greater posterior model probability (Figure 8).

Figure 9.

Time series of observed discharge and annual mean flow-weighted math formula from observations (as estimated by the WRTDS model; see section 'Estimates of Annual Concentration and Flux') and from Models 7 and 9 (see Table 2) for the Wapsipinicon River. Model 7 includes effects of current year's discharge gerf(Q) and N-fertilizer applications, f(Napp). Model 9 includes the same factors plus the effects of previous years' discharge, h(Qpre). The plotted values of Qpre are equal to weighted average from equation (5) with D = 1/15. See Table 2 for model equations.

Effects of N inputs were apparent in four rivers, the Maquoketa, Wapsipinicon, Boyer, and Nodaway. In these four basins, the most-probable models included effects of N-fertilizer inputs (Figure 8), consistent with previous reports of strong influence of N-fertilizer on math formula in streams [Goolsby et al., 1999; David et al., 2010; Robertson and Saad, 2013]. Long-term records of math formula from the early-to-mid 1900s also showed strong increases in concentrations in the Iowa, Des Moines, and Raccoon Rivers over the same time span that N-fertilizer inputs burgeoned [IADNR, 2001]. Changes in N-fertilizer after 1970 in some basins may have been too small, however, to substantially change stream water quality relative to the variability caused by discharge. For the Iowa, Skunk, Des Moines, Raccoon, Little Sioux, and Nishnabotna Rivers, the best fit model with N-fertilizer effects had an acceptable Nash-Sutcliffe model efficiency (0.43–0.68, see Table S3), but the model was not most probable, partly because of the AICc penalty for the additional parameter, τ, for mean travel time. The mean travel time is the delay between application and surface water discharge. It represents the mean residence time in the vadose zone and groundwater of each basin. Calibrated values of τ are shown in Table 3 for all models and all rivers and are discussed in section 'Synthesis: Hydrogeochemical Landscapes and Water Quality Responses'.

Table 3. Mean Residence Time Estimates From the Five Models That Included Effects of N Applicationsa
RiverModel 1Model 7Model 8Model 9Model 10
  1. a

    All models include effects of N-fertilizer applications, f(Napp). Additional factors included in model scenarios are gerf(Q) for Model 7, gerf2(Q) for Model 8, gerf(Q) and h(Qpre) for Model 9, and gerf2(Q) and h(Qpre) for Model 10 (see Table 2).

  2. b

    A model with the maximum posterior probability among all models for that river.

  3. c

    The model with the maximum posterior probability among these five models for that river.

Maquoqeta888b88
Wapsipinicon8888b8
Iowa1555c5
Skunk000c00
Des Moines00c000
Raccoon00c000
Little Sioux00c000
Boyer5555b5
Nishnabotna55c555
Nodaway1919191919b

Additional factors were considered that potentially affect N-inputs, including N-fixation, mineralization, and manure. We tested alternative N-input functions, f(Napp), that included inputs from N-fixation over the estimated area of soybeans in each basin and from a constant rate of mineralization per land area, both adjustable by calibrated coefficients. The model fits did not improve for any basin. We also tested alternative f(Napp) functions with effects of manure by adding to each year in each basin a quantity of N proportional to the Iowa hog population, which is the main source of animal manure in the state. Models with manure applications had decreased probabilities. These results are consistent with previous studies indicating that the combined effects of flow and N-fertilizer explain the majority of variability in math formula trends in the midwestern United States [David et al., 2010] and that most N in Iowa streams comes from N-fertilizer [Robertson and Saad, 2013]. Other factors potentially affecting water quality trends were not quantified as a part of this study but were evaluated indirectly as described below.

To test the hypothesis that effects of discharge are sufficient to explain the decline in annual concentrations in the 2000s we compared the predicted slopes from the models to the slopes of “observed” concentrations from 2005 to 2012, when declines tended to be steepest in these basins. Linear best fit regression was used to estimate slopes to minimize effects of annual variability in the annual flow-weighted concentrations. Figure 10 shows the observed and modeled slopes of the concentration trends. Results are shown for Model 3 (with gerf(Q)), Model 4 (adding gerf2(Q) to Model 3), and Model 6 (adding h(Qpre) to Model 4). Figure 10a shows results for the models fitted to the annual flow-weighted concentrations for the entire period of record. To investigate nonstationarity of model parameters during recent years, results are also shown for models fitted to data from 2000 to 2012 in Figure 10b.

Figure 10.

Observed versus modeled trends in 2005–2012 math formula-N concentrations (a) for models fitted to the entire data series of annual mean flow-weighted concentrations and (b) for models fitted to 2000–2012 concentrations.

The recent declines in concentration are consistent in most basins with a single set of model parameters, rather than temporally varying model parameters that might imply changes in the underlying basin response or management practices. Effects of discharge in Models 3, 4, and 6 were sufficient to explain the recent declines in concentrations. Adding effects of decreasing concentration at high flows (going from Model 3 to Model 4) resulted in a closer match between observed and modeled concentration trends (Figures 10a and 10b). Adding effects of Qpre (Model 6) further improved the match between observed and modeled trends. For all basins except the Des Moines, the predicted slopes from 2005 to 2012 were close to the observed values (R2 = 0.83, p = 0.001, excluding the Des Moines) when models were fitted to annual concentrations from the entire period of record (Figure 10a). In other words, the recent declines in concentration do not require a temporal variation of model parameters. For the Des Moines River, the recent declines in concentration were greater than those in any of the other basins and were not well matched by models calibrated to the 1970–2012 data. When calibrated to the 2000–2012 data, the trend estimated by model 6 were similar to the observed value (R2 = 0.95, p = 1 × 10−6) (Figure 10b), indicating that the response of concentrations to discharge from current and previous years may have changed in this basin as a result of N-fluxes not quantified in the models, recent changes in land management practices affecting N processing and transport, or recent modification of the basin hydrologic response.

To summarize, the results of the multimodel analysis indicate that discharge is a primary factor in determining the temporal variability of annual mean flow-weighted concentrations. The current and preceding years' discharge both were important factors in estimating concentration trends that matched observations. For current year discharge, decreasing trends in the data series were affected by decreasing concentrations at very high and very low discharges (e.g., Figures 7 and 8). For previous year's discharge, additional decreases of concentrations occurred in years following very high discharges. These effects of discharge were sufficient to explain the declines in concentrations in the 2000s in most basins. Effects of N-applications were also apparent in some basins, with a delayed response of streams to trends in N-applications, due to residence time in soils or groundwater (e.g., Figure 8 and Table 3).

6 Synthesis: Hydrogeochemical Landscapes and Water Quality Responses

Annual flow-normalized concentration trends varied according to the landscape features of Iowa. The decreasing trends of the flow-normalized concentrations in the 2000s were steepest in basins in central Iowa (the Iowa, Skunk, Des Moines, Raccoon, and Little Sioux in Figure 6) that overlap with the Des Moines Lobe (Figure 1). Trends were less steep in the basins in the east, and southwest part of the State (peripheral basins) with no drainage from the Des Moines Lobe. The peripheral basins include the Maquoketa and the Wapsipinicon basins in the eastern part of Iowa and the Boyer, Nodaway, and Nishnabotna basins in the south-western part of Iowa. The slopes of the flow-normalized concentration trends for all rivers were negatively correlated with the percent of the watershed area in the Des Moines lobe (R2 = 0.64, p < 0.01). The tendency for steeper declines in basins with greater area in the Des Moines Lobe may indicate a greater sensitivity of those areas to the hydrologic and chemical factors that contribute to changes in water quality. Correlations with other landscape features give additional information about controlling factors. For example, the slopes of concentration trends are positively correlated with watershed topographic slope (R2 = 0.60, p < 0.01), meaning that concentrations declined less in areas with greater relief. Increased topographic relief typically corresponds to deeper groundwater and reduced tile drainage. There was no significant correlation of concentration trends with the percent agriculture, annual precipitation, annual runoff, or soil permeability.

These correlations are consistent with the effects of smaller N-storage volumes in central Iowa basins that overlap with the Des Moines Lobe versus peripheral basins that do not. Basins with smaller storage capacities tend to respond more rapidly to variations than basins with greater storage capacities [Kirchner et al., 2000; Mengistu et al, 2013]. Similarly, transport timescales of math formula in agricultural settings are determined by storage capacity, water and solute inputs, and reactivity in the subsurface [Liao et al., 2012]. Central Iowa basins that overlap with the Des Moines Lobe tend to have shallow unsaturated zones, extensive tile drainage, and reducing conditions favorable for denitrification. These factors produce a limited volume available to store N, rapid transport through that volume, and brief mean residence time of math formula (Table 3).

In contrast, basins outside of the Des Moines Lobe tend to have deeper unsaturated zones, less extensive tile drainage, and older, less reactive sediments [Rodvang and Simpkins, 2001] which may lead to greater storage capacity, longer residence times and more subdued, delayed responses to extreme discharge events. Steinheimer et al. [1998] found that in the loess hills in southwest Iowa, the unsaturated zone had a high potential (possibly >1000 kg/ha) for storage of unutilized math formula with minimal denitrification. Tomer and Burkart [2003] found that groundwater may reside in the surficial loess deposits for decades. Estimated mean residence times (Table 3) (obtained from calibration of the τ parameter in the multimodeling analysis) were greater in the peripheral basins (averaging 9 years among these five basins) than in basins overlapping the Des Moines Lobe (averaging 1 year among these five basins). Responses in the peripheral basins to varying discharge, for example in the Maquoketa and Wapsipinicon, tended to be more moderate for a given 5–10 year period than the central basins (Figure 7). The longer term variability in the Maquoketa and Wapsipinicon concentrations may result from long-term changes in N inputs, as shown by apparent effects of the N-fertilizer estimates on water quality trends in those basins (Figure 8).

The effects of N-storage volumes in these basins are further supported by low-flow data that are indicative of groundwater contributions to the streams. Decadal trends of concentrations of math formula during low flows are shown for each river in Figure 11. The figure includes math formula samples that were collected during times when flow was below the first quartile for the period of this study. Four of five rivers outside of the Des Moines Lobe show statistically significant trends of increasing math formula in base flow, consistent with gradual accumulation of groundwater math formula in deeper, less reactive aquifers [Liao et al., 2012]. Sprague et al. [2011] also found increasing concentrations of math formula in low-flow waters for the Missouri River at Hermann, Missouri, and proposed that the accumulation of N in groundwater may explain the increasing trend in flow-normalized concentrations observed in that basin. One basin (Nodaway) has a nonsignificant trend and a relatively brief (10 year) time series of low-flow data. In rivers overlapping with the Des Moines Lobe, three show trends of flat (not significantly different from zero) math formula over the period of this study, consistent with steady state vertical profiles of math formula in shallow, reactive aquifers [Liao et al., 2012]. Statistically significant decreasing trends in the Skunk and Des Moines Rivers may indicate a long-term shift toward lower math formula in base flow as a result of potential factors such as changing management practices, modifications of drainage, or increasing precipitation that may affect the short-residence-time pool of groundwater math formula in these tile-drained areas. The overall geographical pattern in base flow concentration trends corresponds to the relative decreases in flow-normalized math formula concentration from 2000 to 2012 (Figure 6) with the steepest declines in basins central the Des Moines Lobe (Des Moines and Raccoon), with moderate declines in basins partly overlapping the Des Moines Lobe (Iowa, Skunk, and Little Sioux) and with the smallest declines (or increases) in basins outside of the Des Moines Lobe (Maquoketa, Wapsipinicon, Boyer, Nishnabotna, and Nodaway).

Figure 11.

Trends in nitrate concentration in 10 river basins for samples taken during low flow events, below the first quartile of all recorded flows during the time period of the study. Black lines are least squares linear regressions shown in solid for slopes significantly greater than or less than zero (p < 0.05), or dashed for slopes not significantly different from zero. The colors of points indicate the season of collection in winter (“win” = December, January, February), spring (“spr” = March, April, May), summer (“sum” = June, July, August), or fall (“fall”= September, October, November). Plots are shaded yellow for basins that overlap with the Des Moines Lobe.

7 Summary and Conclusions

Ten basins comprising 63% of the total area of Iowa were analyzed for changes in concentrations using a method (flow normalization, FN) that estimates annual concentrations independent of discharge variations. Flow-normalized math formula concentrations followed similar trends as N-fertilizer inputs from 1970 to 2000. From 2000 to 2012, stream math formula concentrations decreased at the same time that corn acreage and N-fertilizer applications were increasing as a result of high corn prices in a strong ethanol biofuel market. Testing of multiple conceptual models of the response of annual mean flow-weighted math formula concentrations (annual total flux divided by annual total discharge) to current-year discharge, previous-years' discharge, and N-applications indicate that recent and historical trends in math formula concentration are affected by the interaction of climatic variability with the hydrogeochemical landscape. Steeper declines occurred in basins overlapping the Des Moines lobe, likely as a result of artificial drainage and more strongly reducing conditions that caused a smaller storage volume, shorter mean residence time of math formula, and greater sensitivity to variations in precipitation and discharge. Trends of math formula in rivers draining deep soils with aerobic aquifers are more strongly affected by the historical application rates and the extended residence time of math formula in the subsurface. Response to climatic variations tends to be more muted in those rivers.

These results have broad implications for forecasting stream water quality and fluxes in agricultural basins. The variable response of basins to climatic and anthropogenic forces points to different strategies for mitigation of agricultural effects on water quality. In the peripheral Iowa basins and in similar settings, there is a short-term insensitivity to fertilizer inputs and a long-term sensitivity. Long-term monitoring in these areas will be essential to evaluate effects of agricultural management practices and other efforts to improve water quality. The central Iowa basins and other tile drained areas with shallow storage of N show greater sensitivity to climatic variability. High frequency sampling, such as with continuous math formula probes, may be necessary to better resolve the daily and annual responses of concentrations and fluxes in these basins, especially if climatic variability continues to increase. Optimum management strategies to improve stream water quality in these basins will also need to account for the variable response across different basin types and the time scales of responses. In peripheral basins, policies focusing on long-term management will be consistent with the relatively long time scales of transport and relatively subdued response to annual changes in discharge. In central basins, shorter-term strategies may be necessary to avoid steeper declines of water quality due to sensitivity to annual climatic effects.

Acknowledgments

Thanks to Donna Lutz from the Iowa State University, the principal investigator of the Des Moines River Water Quality network, who provided sampling and site information and data from sites on the Raccoon and Des Moines River and to Donald Carrington of the USDA Natural Resources Conservation Service, who provided information about wetlands delineation requests in Iowa. We also thank Matt Miller of the USGS, three anonymous reviewers, and the editors of Water Resources Research for their constructive comments that were essential for improving the manuscript. This work was supported by the USGS National Research Program, USGS National Water Quality Assessment Program, and the USGS Energy Resources Program.

Ancillary