Introduction

Losses of anthropogenic nitrogen (N) and phosphorus (P) from landscapes to aquatic systems represent major concerns to agricultural sustainability and water quality. In agricultural landscapes, excess N and P delivered to streams and rivers has been linked mainly to losses of excess chemical fertilizers and manure from fields, with additional contributions from wastewater, lawn fertilizers, and pet waste in urban/suburban areas (Goolsby and Battaglin 2001; Robertson and Saad 2011, 2013; Turner et al. 2003). However, not all nutrients that enter streams are transported to river mouths (Howarth et al. 1996; Seitzinger et al. 2002). Efforts to explain the “missing” nutrients have emphasized the role of biogeochemical processes including denitrification in streams (e.g., Hedin et al. 1998; Mulholland et al. 2008) and wetlands (e.g., Hanson et al. 1994; Jordan et al. 2011; Kadlec and Wallace 2009) as well as P retention and burial in natural lakes (Dillon and Evans 1993). More recently, landscape modeling studies at basin- and global-levels have indicated substantial retention of N and P by reservoirs in aggregate (Brown et al. 2011; Harrison et al. 2009; Wollheim et al. 2008), suggesting these ecosystems could be more influential to downstream nutrient delivery than previously thought.

In contrast to natural lakes, whose formation and geographic extent are strongly restricted by historic glaciation, reservoirs have been constructed in a wider array of landscapes including the most productive agricultural areas of the globe. Agricultural landscapes often contain a remarkable array of reservoirs (Dendy 1968; Downing et al. 2006, 2008; Renwick et al. 2005; Smith et al. 2002). In the US, smaller farm ponds are thought to number in the hundreds of thousands (Downing et al. 2006; McDonald et al. 2012; Renwick et al. 2005), and many of the 80,000 registered, mostly larger dams (USACE 2013) also occur in agricultural areas. As a result, reservoirs now clearly play a substantial role in global elemental cycles of N (Harrison et al. 2009; Wollheim et al. 2008) and P (Carpenter and Bennett 2011). Still, our understanding about agricultural reservoirs and their effects on nutrient transport has mostly arrived through case studies of individual sites, or coarser studies of large basins, rather than through comparative studies of reservoirs that might reveal cross-site variation. For example, in the Midwest US, a detailed study of one large agricultural reservoir by David et al. (2006) indicated more than half of the N inputs were lost via denitrification, while transport models of large basins have shown aggregate annual retention of agricultural P by multiple reservoirs (Bosch and Allan 2008; Brown et al. 2011; Robertson and Saad 2011; Smith et al. 1997), presumably due to P burial. Consequently, much heterogeneity in reservoir effects on N and P transport remains to be refined across sites and regions.

Reservoirs constructed within river networks are often highly open ecosystems, that is, major transfers of water and nutrients occur across ecosystem boundaries and promote turnover and export of the nutrient pool. Indeed, all ecosystems are open to some degree (Fisher and Likens 1973), and the openness of reservoirs is attributed to features such as elongate morphometry, longitudinal currents, and discrete inlets and outlets (Thornton et al. 1990). Though a basic tenet within stream ecology, the open ecosystem view has not been embraced in many ecological studies involving lakes (but see Bartrons et al. 2013; Cole et al. 2007). Rather, much research on eutrophication, nutrient cycling, and nutrient retention in lakes (e.g., Brett and Benjamin 2008; Carpenter 2005; e.g., Dillon et al. 1990; Schindler 1980) has proceeded from a view of lakes as closed systems, or the endpoints of watershed flows, not as conduits that control the nutrient supply to downstream ecosystems. The persistence of the closed system view may be partly explained by hydrology, because natural lakes are often surface manifestations of the groundwater system, whereas reservoirs are more strongly driven by surface water flows (Thornton et al. 1990) and are often more directly connected to other ecosystems within flow networks. Appreciation of reservoirs as open ecosystems is particularly crucial within agricultural landscapes, where excess nutrient loading can overwhelm their capacity for retention and transformation, thereby affecting nutrient supply to downstream water bodies, as has been shown in streams (Bernot and Dodds 2005; Dodds et al. 2002; O’Brien and Dodds 2010). We therefore emphasize a need to understand how reservoirs in agricultural landscapes function, and in turn influence downstream nutrient supply by virtue of their openness and connectedness to the river network. Further, reservoir biogeochemical processes are of interest to both water and land managers, because reservoir processes can modify and delay river water quality responses to land management change (Hamilton 2012; Powers et al. 2014; Sharpley et al. 2013).

Biogeochemical processes in reservoirs are affected by several interacting internal and external factors that ultimately influence nutrient transport through river networks. For instance, proportional nutrient loss/retention is strongly influenced by water residence time and hydraulic loading (e.g., Essington and Carpenter 2000; Koiv et al. 2011; Persson et al. 1999; Seitzinger et al. 2002) as well as the strength of hydrologic connections between sources and sinks (Powers et al. 2012) but the efficiency of retention is expected to decrease with nutrient loading (Koiv et al. 2011; e.g., Shostell and Bukaveckas 2004). In addition, while reservoirs are generally thought to be sinks for nutrients, they can become net sources to downstream waters when internal nutrient stores are mobilized and exported (Nowlin et al. 2005; Teodoru and Wehrli 2005; Wurtsbaugh et al. 2005), perhaps especially along the recovery from eutrophication. Yet, there remain unanswered questions about the long-term functioning and cross-site variation of reservoirs apart from natural lakes, and differences in their effects on N versus P transport, especially in highly agricultural landscapes. That is because most prior studies of nutrient retention by reservoirs have been generally limited by at least one of the following issues: absence of heavy agriculture, emphasis on one element only (either N or P), emphasis on internal functioning irrespective of transport through the outlets, limited time frame, limited number of sites, or inability to partition reservoir influences from those of natural lakes. To further illustrate the last point, a landscape-level analysis by Powers et al. (2014) showed aggregate losses of N and P in lakes and reservoirs, but the approach did not allow for reservoir processes to be partitioned from those of natural lakes in the uniquely lake-rich landscape of Wisconsin. Similarly, a global analysis by Wollheim et al. (2008) reported N removal by lakes and small reservoirs as a combined term, and for Mississippi River Basin these losses represented 21 % of total network N inputs. Thus, there remains a need to better resolve the effects of reservoirs on the transport of agricultural nutrients, apart from effects of natural lakes, across the range of hydrological and agricultural conditions that occur in heterogeneous regions.

In this paper, we examine the influence of reservoirs on nutrient transport to downstream waters in agricultural landscapes. To address this interest, we quantified the effects of reservoirs on annual river nutrient transport in the Midwestern US, considering potentially different effects on N and P, as well as variation among sites. We hypothesize that reservoirs have broad control on N and P transport in agricultural landscapes, thus influencing nutrient supply to downstream water bodies. Our analytical approach capitalizes on the placement of key long-term river monitoring sites near reservoir inlets and outlets. Overall, we incorporated water quality records spanning an 18 year period (1990–2007) at 869 river sites within the Upper Mississippi River (MISS), Great Lakes (GRL), and Ohio River (OHIO) Basins, where the majority of sites were dominated by agricultural land use. We predicted that reservoirs would mainly function as nutrient sinks, reducing the export of agricultural N and P through nutrient burial as well as denitrification in the case of N (David et al. 2006), which would corroborate prior case studies of reservoirs and modeling studies of large basins. However, long-term nutrient enrichment can potentially reduce the ability of reservoirs to function as sinks, especially in the case of P, which unlike N has no major pathway of gaseous removal to the atmosphere. So we also predicted that reservoirs subjected to long-term P enrichment could show net P export, though to our knowledge this has not yet been shown for reservoirs in agricultural landscapes.

Study area and methods

In the US, many long-term river water quality monitoring sites are positioned in close proximity to reservoirs. We used two approaches to examine the influence of reservoirs on annual river transport of total nitrogen (TN) and total phosphorus (TP), focusing on agricultural landscapes of the Midwestern US. In the first approach, we used watershed comparisons to quantify the difference in nutrient yields between rivers that lacked nearby upstream reservoirs along the mainstem (hereafter termed unimpeded sites) and rivers that were reservoir-fed (hereafter termed reservoir outflow sites), or in other words were positioned directly downstream of reservoir outlets. Among all river sites in the analysis (N = 869), there were 100 reservoir outflow sites, including 95 outflow sites having TP data and 82 having TN data, compared to 697 unimpeded sites for TP and 563 unimpeded sites for TN. In the second approach, annual mass balances (net fluxes) were compiled and examined for the subset of reservoirs that had paired monitoring of both nutrient inflows and outflows. There were 16 reservoirs in the mass balance analysis for TP, and 17 for TN. In both the watershed comparison and mass balance approaches, we used sites with sufficient water quality records for annual load estimation between water years 1990 and 2007, and sites were selected using screening criteria based on Saad et al. (2011) that we describe further below. Our study area included lands drained by tributaries to: the Laurentian Great Lakes (GLR); the Mississippi R. (MISS), including eastern tributaries upstream of the Ohio River confluence, and western tributaries upstream of the Missouri R. confluence; the northern bank of the Ohio R. (OHIO), including the Allegheny River. To help understand cross-site variation, we quantified the effects of reservoirs within multiple subbasins (based on HUC4, four-digit hydrologic unit code; Table 1; Fig. 1).

Table 1 Subbasins of the analysis
Full size table
Fig. 1
figure 1

River monitoring sites used in the analysis (N = 869). Reservoir outflow sites are shown in gray (n = 99 for TP data, n = 87 for TN data). Unimpeded river sites are shown as smaller black dots. Triangles are reservoirs with paired monitoring of both inflows and outflows, and these were used to calculate net fluxes

Full size image

Catchment boundaries for river monitoring sites were delineated in ARCmap 10 using the National Hydrography Dataset (NHD) and 90-m National Elevation Dataset. The percent of catchment area occupied by row crop agriculture (P crop , or cropland density) and urban land (P urb ) were calculated from the National Land Cover Dataset (in year 2001; Homer et al. 2007). Drainage areas (A) ranged from 8.0 to 96,000 km2, with 72 % of the drainage areas occurring in the range of 50–5000 km2, including 37 % <1000 km2.

Identification of reservoir outflow, reservoir inflow, and unimpeded river sites

We used flowlines and surface water bodies of the NHD (NHD-PLUS) to identify river sites located downstream of nearby reservoirs. The site was classified as a reservoir outflow site if there was an upstream dam positioned along the same stretch (NHD levelpath) of the river mainstem, or if the monitoring site was directly downstream of such a levelpath. This procedure selected for rivers dominated by flows that had recently passed through a nearby reservoir, and excludes rivers that receive reservoir water only through minor tributaries or distant reservoirs. We then confirmed that reservoir outflow sites were indeed positioned downstream of reservoirs, and that the reach between each reservoir and river site lacked major tributaries, by manually inspecting aerial imagery of the local landscape using Google Earth, in which water bodies and flow directions of rivers were discernable. To prevent double-counting of reservoirs, we removed a small number of reservoir outflow sites from the analysis that received water from the same reservoir, discarding all but the nearest reservoir outflow site. After exclusions, a total of N = 869 monitoring sites remained for analysis, including 100 reservoir outflow sites, the majority of which were reservoirs created by registered dams (USACE 2013). Typically, reservoir outflow sites were located <10 km downstream of reservoir outlets, and in many cases they were positioned immediately below the outlet. To illustrate, the contributing area downstream of the outlet, approximated as the square of the river distance from outlet to sampling station, was usually small relative to the total contributing area, accounting for <7 % of total contributing area at 75 % of sites. We also excluded 18 natural lake-fed sites from our watershed comparative approach so as not to interfere with reservoir comparisons. Remaining river sites that were neither reservoir-fed nor lake-fed were classified as unimpeded sites.

Following identification of reservoir outflow sites, we then identified corresponding inflow sites for use in the mass balance approach. To do so, we manually inspected the contributing area of each reservoir outflow site for candidate inflow sites using Google Earth. For the majority of reservoirs, no inlet site was available in our data set, necessitating the watershed comparative approach. Of the remaining 18 reservoirs that had inlet sites, usually only one inlet was monitored, but in a few exceptions where multiple inlets were monitored, we chose the inlet site with the largest contributing area (i.e., most downstream site). To ensure the inflow sites accounted for the majority of nutrient inputs, we required that the contributing area of each inflow site encompassed >50 % of the contributing area of the reservoir outflow site.

Estimation of nutrient loads and yields

For each river monitoring site, we estimated annual nutrient loads using statistical relationships between daily concentration, daily discharge, and day of year (Saad et al. 2011; Schwarz et al. 2006). This procedure uses a water-quality model (Eq. 1) that relates the logarithm of concentration at time t (c t , mg L−1) to the logarithm of daily flow (q t , L day−1), a decimal time term to represent trend (T t ), sine and cosine functions of decimal time to account for seasonal variation, and a model residual (e t ),

$$c_{t} = b_{0} + b_{q} q_{t} + b_{T} T_{t} + b_{s} \sin \left( {2\pi T_{t} } \right) + b_{c} \cos \left( {2\pi T_{t} } \right) + e_{t} ,$$
(1)

where b 0 , b q , b T , b s , and b c are fixed coefficients for each site that are estimated by the ordinary least squares method, or if some of the c t measurements are censored, by the adjusted maximum likelihood method, and e t is assumed to be independent and normally distributed with mean 0 and variance \(\sigma_{e}^{2} .\) Daily loads (l t , M T−1) for each site were in turn estimated from c t  × q t , and summed for each water year to estimate annual loads (L). All sites in our analysis met the following water quality/discharge screening criteria, which are based on those of Saad et al. (2011). First, ≥20 non-censored water samples and ≥2 years of water quality data were required at each site, although most sites had >50 non-censored water samples and >5 years of water quality data. Second, if the period of record for a given site was short (<5 years), the site was required to have water quality data collected for at least 1 year between 1996 and 2004; if the period of record was longer, the site was required to have water quality data collected for at least 1 year between 1995 and 2007. Third, the discharge record was required to have ≥2 years of overlap with the water quality record.

We took additional steps to ensure our load estimations were as accurate as possible. Stenback et al. (2011) reported substantial bias in load estimates of nitrate, and indicated the rating curve/regression procedure can result in potential bias to annual load estimates after the estimated loads in log space are transformed back into real space, even after statistical transformation bias adjustments. In contrast, Saad et al. (2011) found that TN and TP load estimates from this procedure are reasonably unbiased. To help ensure unbiased load estimates, we omitted sites where the regression relation performed poorly (i.e., where standard error, SE, was >50 %). More specifically, for the days when concentration was measured, the total of the loads predicted by the regression model was compared to the total of the observed loads, and we omitted all sites with a ratio of the total estimated load to the total observed load (O/E) greater than 1.5 and less than 0.667; we chose these values in order to be consistent with the 50 % SE criterion for omitting sites with poor regression models. Overall, there was a lack of systematic bias in our data set between reservoir outflow sites and unimpeded sites. Further, paired t-tests showed no significant differences in O/E between reservoir outflow sites and unimpeded sites within each basin. For TP across sites, median O/E was 1.05 (i.e., biased slightly low), and was 0.98 for TN (i.e., biased slightly high). Also, O/E fell between 0.75 and 1.25 for 84 % of TP sites, and 91 % of TN sites, indicating that estimated daily loads deviated from observed daily loads by less than 25 % at the majority of sites.

To compare load and discharge estimates over a common time frame, and to best deal with asynchronous precipitation as well as a minority of sites that showed long-term trends, we linearly interpolated each annual loading and discharge time series to a common water year, 2000 (i.e., 1 October 1999–30 September 2000, giving L and Q). In the absence of a significant long-term linear trend, L is equivalent to the simple long-term average of loads across years; indeed, most sites lacked long-term linear trends. We then calculated nutrient yields (Y = L/A, kg km−2 year−1, where A is drainage area). Yields were log-normally distributed. Coefficients of variation (CV) for the annual yields were calculated at each site. For use in the mass balance approach, we identified the highest and lowest flow years occurring between 1990 and 2007 at each site, and used the corresponding load estimates from those years to represent wet year loads (L wet ) and dry year loads (L dry ). For reference purposes, we report flow-weighted mean nutrient concentration (C = L/Q, mg L−1) here. The interquartile range for nutrient concentration, across all sites, was 0.088–0.29 mg L−1 for TP and 1.7–6.5 mg L−1 for TN, illustrating that these waters were typically nutrient-rich. To facilitate interpretation of cross-site patterns in our yield data, we add that when all yield and concentration data were pooled, an increase in TP or TN yield of 1000 kg km−2 year−1 corresponded to about a 3.0 mg L−1 increase in concentration.

Watershed comparison approach

To examine effects of reservoirs, we first used watershed comparisons. The statistical approach below identifies whether yields at reservoir outflow sites are significantly different from those at unimpeded sites, and how large the difference, providing a substitute for the difference between reservoir inputs and outputs. We also incorporated relationships between yield and percent cropland (if significant). Of central interest in this analysis is the b res term, the coefficient that expresses the mean difference between reservoir outflow yields and unimpeded yields on a log scale. This potential difference between reservoir-fed and unimpeded sites was examined at three different levels of organization: overall (region-wide), basin-specific (GRL, OHIO, and MISS), and subbasin-specific (e.g., 512-WAB, Wabash River). Expressing the above in model terms, we first considered a simple statistical model of the yields at reservoir-fed sites (model res )

$$\log ( Y) = b_{res} *RES_{1,0} + b_{0} + err,$$
(2)

where b res is the coefficient for the reservoir effect, b 0 is the intercept, RES 1,0 is a binary variable indicating which river sites are reservoir outflow sites and thus shall contribute to the adjustment from b 0 owing to b res (1 = reservoir outflow site, 0 = unimpeded site), and err is a vector of normally distributed residuals. Recognizing the likely increase in river yields as a function of cropland density, we also considered a combined model of cropland and reservoir effects (model res+crop ) and the statistical model is of the form

$$\log (Y) = b_{crop} *P_{crop} + b_{res} *RES_{1,0} + b_{0} + err,$$
(3)

where P crop is the percentage of cropland present in the basin, b crop is the coefficient describing the crop effect, and remaining terms are as defined in Eq. 2. The individual cropland model was also considered (model crop ). We then examined whether inclusion of the b res and b crop terms was justified by comparing model res , model crop , and model res+crop using Akaike information criteria (1974). Following the overall analysis, reservoir effects on nutrient transport were quantified using model res and model res+crop within each basin and subbasin, choosing model res+crop only when the relationship between P crop and log(Y) was significant within the group. The significance of each b res coefficient was determined using Bonferroni corrected p-values (α = 0.05). The proportion of variance explained (r 2), 95 % confidence intervals (CIs), and SEs were calculated. For the overall data, reservoir effects on interannual variability were also examined, by substituting the yield CV in place of Y in Eqs. 2 and 3, then repeating the above procedures.

Using the fitted model coefficients (b res , b crop ), we calculated the mean difference between reservoir outflow sites and unimpeded sites in units of yield (kg km−2 year−1) as well as the percent difference from the unimpeded value [100 × (Outflow − Unimpeded)/Unimpeded]. In both cases, negative values indicate lower yields at reservoir outflow sites than unimpeded sites. These calculations were carried out at each level of analysis using the selected model (Eq. 2 or 3); when b crop was significant, we used the mean value of P crop from the data to estimate the unimpeded yield and provide an approximation of the group-wide pattern. Using paired t-tests, we also examined differences in land use (P crop , P urb ) and position within the river network (represented by log A) at the subbasin and basin levels, to determine if factors other than reservoir effects could explain yield differences.

Mass balance approach (reservoir net fluxes)

In a second approach, we used reservoirs with paired monitoring of reservoir inflows and outflows to estimate net fluxes of TP and TN as a percentage of the inputs [F net  = 100 × (Outputs − Inputs)/Inputs]. Negative values for net flux indicate net nutrient retention/loss and positive values indicate net nutrient export. At each reservoir, separate estimates of net flux were calculated for typical, wet, and dry years using the corresponding load measures (L, L wet , and L dry ). Estimation of these net fluxes required the incorporation of additional mass inputs not delivered through inflow monitoring sites. We estimated total input loads (Inputs) to each reservoir as L inflow  + L point,in  + L trib where L inflow is the monitored load at the inflow station, L point,in is loading from point sources delivered directly to the reservoir or inlet areas external to the inflow site, and L trib is additional loading from unmonitored areas that are tributary to the reservoir. In most cases, L inflow accounted for the majority of Inputs. L point,in was estimated as the sum of all 1997 point sources located within approximately 1 km of the reservoir shoreline, as indicated by the USEPA permit compliance system database. L trib was estimated as Y inflow  × (A outflow  − A inflow ), where Y inflow is the yield at the inflow site, which is based on the assumption that unmonitored areas share an identical nutrient yield with the monitored inflow site, in other words that loading from unmonitored areas is directly proportional to the unmonitored catchment area. We estimated output loads (Outputs) from each reservoir as L outflow  − L point,out , where L outflow is the monitored load at the outflow station and L point,out is loading from point sources entering between the reservoir outlet and the outflow monitoring station. Tributary inputs entering between the reservoir and outflow monitoring station were assumed to be zero within these generally short subreaches, where major tributaries were also lacking. Nutrient losses via water withdrawal were also assumed to be zero. The decision on whether to route point sources through the reservoir (L point,in ) or downstream of the reservoir outlet (L point,out ) was made based on manual inspection of each individual point source location and surrounding area using aerial imagery in Google Earth. Point nutrient sources (L point,in ) represented <10 % of Inputs in these calculations, with the exception of Cecil M. Harden Lake (30 % of TP input). Potential errors in Inputs associated with assumptions about unmonitored areas and point sources were protected against by several factors. These factors, generally stated, include: the smaller size of unmonitored catchment areas relative to that of monitored tributaries; similarities in land use among sub-catchments of each reservoir, such that yields from monitored tributaries may reasonably represent yields from the nearby unmonitored areas; the generally small value of L point,in relative to overall Inputs. Group-level (basin, and overall) means and standard deviations (SDs) of the net fluxes were calculated.

We then compiled data from the National Inventory of Dams database (USACE 2013) for features that were potentially associated with biogeochemical processes, including reservoir surface area, normal water storage capacity, as well as latitude. We used normal water storage capacity, when available, to estimate mean depth (i.e., volume/surface area) and mean water residence time (i.e., volume/annual discharge). Hydraulic loading (H L , m year−1) was calculated as Q/surface area, and uptake velocity (v f , m year−1) as −1 × H L  × ln(1 − R) (Wollheim et al. 2008), also known as settling velocity or piston velocity, where R is the proportional removal, calculated as −1 × F net /100. We also calculated TN:TP (i.e., ratio of input loads) for reservoirs that had data for both nutrients. Relationships between reservoir net fluxes and the above metrics were examined using linear regression.

Results

Watershed comparison approach: region-wide effects of reservoirs on river nutrient transport

Nutrient yields varied considerably among sites, and many sites were dominated by agricultural land use (>50 % P crop , Table 2). Inclusion of both the b res and b crop terms (model res+crop ) was necessary to account for combined influences of reservoirs and cropland on yield magnitude, with model res+crop showing a decrease by >8 AIC units over simpler model res and model crop (Table 3), although the improvement in explanatory power was rather low (r 2 increased by 0.02–0.03). Model crop explained much more variation in TN yield (r 2 = 0.55) than TP yield (r 2 = 0.12).

Table 2 Summary of all river monitoring data
Full size table
Table 3 Model comparisons involving the effects of cropland density and reservoirs on river nutrient export, determined from the watershed comparison approach
Full size table

Generally, for a given cropland density, reservoir outflow sites had lower nutrient yields compared to unimpeded sites (Fig. 2), suggesting reservoir nutrient retention/loss. Based on model res+crop , which accounts for the effect of cropland, reservoir outflow sites had on average 25 % lower TP yield (95 % CI = 10–38 %, Table 4) and 23 % lower TN yield (95 % CI = 10–33 %). The absolute difference in TN yield between reservoir outflow sites and unimpeded sites increased with cropland density due to a strong positive exponential relationship between TN yield and P crop . In contrast, the absolute difference in TP yield was more uniform across the P crop gradient. Interannual variability in TP yields was also 11 % lower at reservoir outflow sites (Table 3, CV = 0.38 at mean P crop ) compared to unimpeded rivers (CV = 0.43 at mean P crop ) according to the best model (model res+crop ; CV outflow  − CV unimpeded  = −0.048 ± 0.021 SE; b res significant, p = 0.024). However, there was no significant difference in the interannual variability of TN yields between reservoir outflows sites and unimpeded rivers (CV = 0.38 at mean P crop ).

Fig. 2
figure 2

Lower total phosphorus (TP) and total nitrogen (TN) at reservoir outflow sites (gray circles) compared to unimpeded rivers (smaller black dots) along a gradient of cropland density (% of catchment area), determined using the watershed comparison approach. Lines are fitted values for reservoir outflow sites and unimpeded sites based on a statistical model (Eq. 3)

Full size image
Table 4 Differences in nutrient yields between reservoir outflow sites and unimpeded rivers determined from the watershed comparison approach
Full size table

Watershed comparison approach: differences in reservoir effects among basins and subbasins

When the overall reservoir analysis above was broken down to the basin level, reservoir outflow sites in OHIO and GRL had lower TP and TN yields compared to unimpeded rivers according to model res+crop (Fig. 3). In the OHIO basin, yields from reservoir outflow sites were 60 % lower for TP and 39 % lower for TN. Significant retention (on a percent basis) in OHIO basin combined with high annual loading (unimpeded site TP and TN yields of 77 and 1700 kg km−2 year−1 (Table 4)) indicated that a substantial total nutrient mass (kg) was retained by monitored reservoirs in this basin. Similar to OHIO basin, yields from reservoir outflow sites in GRL basin were 49 % lower for TP and 31 % lower for TN than unimpeded sites. Unimpeded sites of GRL basin had lower yields than those of OHIO and MISS, with median TP yield = 33 kg km−2 year−1 and median TN yield = 710 kg km−2 year−1, suggesting lower nutrient loading to GRL reservoirs. GRL basin also had lower nutrient concentrations across sites, with median TP = 0.091 mg L−1 and TN = 2.2 mg L−1, respectively, compared to 0.20 and 5.7 mg L−1 in MISS basin and 0.25 and 4.6 mg L−1 in OHIO basin. For MISS basin, yields at reservoir outflow sites were not significantly different from unimpeded sites, and median nutrient yields were 62 kg km−2 year−1 for TP and 1400 kg km−2 year−1 for TN (Table 2).

Fig. 3
figure 3

Differences between nutrients in reservoir outflow sites and unimpeded rivers determined using the watershed comparison approach. Results are plotted as the difference in yield magnitude (left, kg km−2 year−1) and percent difference [right, 100 * (Outflow − unimpeded)/Unimpeded]. Significant differences from zero are shown as black squares with star (*) alongside right axis labels. Sample sizes are shown at right (number of reservoir outflow sites in parentheses). Error bars extend beyond the plotted axis for a small number of subbasins that had few reservoir outflow sites

Full size image

While cropland density was generally high, there was also substantial variation represented within each basin (P crop inter-quartile range of 29–77 % in MISS, 4.9–44 % in GRL, and 45–75 % in OHIO) as well as variation in other landscape factors (P urb , log A). In general, P crop , P urb , and log A were not significantly different between reservoir outflow and unimpeded sites, indicating differences in (non-reservoir) landscape factors were not strongly influencing differences associated with b res . The few exceptions did not explain differences associated with reservoirs, and we explain these below. For TP sites in MISS basin only, reservoir outflow sites had 10 % lower P crop at reservoir outflow sites compared to unimpeded sites (p = 0.02), a relatively small difference compared to basin-level P crop range that is accounted for in the model used to estimate the b res term (model res+crop ). MISS basin TP sites also had higher 10 % higher P urb at outflow sites compared to unimpeded sites (p < 0.01). Reservoir outflow sites did not have significantly higher drainage areas than unimpeded sites within any basin. Rather, for OHIO basin only, reservoir outflow sites had smaller drainage areas than unimpeded sites (p = 0.05 for TP sites, p = 0.02 for TN sites).

Basin-level effects of reservoirs on rivers were supported by underlying patterns at the subbasin level. After Bonferroni correction, three subbasins showed significantly lower TP yields at reservoir outflow sites compared to unimpeded sites (Table 4, b res <0 in 409-SCDE, 506-SCIO, and 512-WAB; Fig. 3), whereas five subbasins showed significantly lower TN at reservoir outflow sites (b res <0 in 409-SCDE, 506-SCIO, 508-GRMI, 512-WAB, 702-MIN). In these subbasins, yields were lower at reservoir outflow sites by 34–81 %, suggesting reservoirs cause a substantial decrease in both TP and TN delivered to downstream ecosystems. Underlying log-transformed yields for each subbasin indicated that individual reservoir outflow sites often had lower yields, and also these sites usually fell within a similar range for P crop as the unimpeded sites (TP in Fig. 4, TN in Fig. 5). The ability to detect significant reservoir effects within subbasins was sometimes limited by the small number of reservoir outflow sites. For instance, subbasins 706-MIUK, 708-MISI, and 710-DESM each had <4 reservoir outflow sites, and b res was not significant within these. However, when the data were pooled across these three subbasins, b res was significant, showing 37 % lower TP yields at reservoir outflow sites (p < 0.002) and 21 % lower TN yields (p = 0.03). Additionally, after Bonferroni, there were no basins or subbasins in which reservoir outflow sites had significantly higher nutrient yields than unimpeded sites according to the watershed comparison approach (i.e., no basins or subbasins had significant b res > 0).

Fig. 4
figure 4

Subbasin breakdowns for total phosphorus at reservoir outflow sites compared to unimpeded rivers along the gradient of cropland density, determined using the watershed comparison approach. Reservoir outflow sites are shown as enlarged gray circles, unimpeded rivers as smaller black dots

Full size image
Fig. 5
figure 5

Subbasin breakdowns for total nitrogen at reservoir outflow sites compared to unimpeded rivers along the gradient of cropland density, determined using the watershed comparison approach. Reservoir outflow sites are shown as enlarged gray circles, unimpeded rivers as smaller black dots

Full size image

Mass balance approach (reservoir net fluxes)

Using available data from paired inflows and outflows, we estimated net nutrient fluxes for 17 reservoirs. Figure 6 shows that reservoirs were often net nutrient sinks, including during wet or dry years. Net retention/loss of TP and TN in reservoirs commonly accounted for >20 % of inputs (Table 5). Net fluxes (mean ± SD) were −26 ± 31 % for TP and −26 ± 23 % for TN, and several sites showed net retention accounting for >40 % of the inputs (Table 5). Compared to TN, net fluxes of TP were more variable, including four reservoirs that had positive net fluxes. When we used a simple bias adjustment for Outputs and Inputs provided by the daily observed:estimated loads (Inputs × O/E, Inputs × O/E), the sign of the net flux (source/sink) was upheld, with one exception for TP (Wisconsin River Flowage) and one exception for TN (Big Eau Pleine Reservoir; Table 5). The above net fluxes corresponded to an average uptake velocity (v f ) of 20.4 m year−1 for TP (interquartile range 0.082–38) and 26.4 m year−1 for TN (interquartile range 7.4–34). Within each individual reservoir, net fluxes in wet/dry years usually showed only modest deviations from general trends (Fig. 6; wettest year on record depicted as “x” and driest year on record depicted as “+”), although more than half of the reservoirs showed decreases in net TN retention (as % input) during wet years. A minority of reservoirs showed increases in TN retention during dry years. Wet/dry year differences in TP net fluxes varied more widely, especially among reservoirs showing net export.

Fig. 6
figure 6

Net fluxes of total phosphorus (TP) and total nitrogen (TN) for reservoirs with monitored outflows and inflows, determined using the mass balance approach. Each bar represents the flux for one reservoir. Overall and basin-level means (±SD) are shown at top, with subbasins indicated at left. Flux calculations for wet/dry years are indicated with symbols (crosses flux during wettest year of record, plus flux during driest year of record). Each bar is a single reservoir

Full size image
Table 5 Net fluxes and associated features for reservoirs determined from the mass balance approach
Full size table

Nutrient retention in relation to reservoir structure

For TP, as expected, reservoir retention/loss increased with the log of water residence time (negative relation when expressed as net flux; p = 0.026, r 2 = 0.38, slope = −30 ± 11 SE; Fig. 7). The corresponding relationship between net flux and hydraulic loading (H L ) was also significant (p = 0.034, r 2 = 0.27, slope = −34 ± 14 SE). Although TN net fluxes were not significantly related to water residence time, like TP there was a significant relationship between TN fluxes and H L (p = 0.025, r 2 = 0.28, slope = −26 ± 10 SE). Across reservoirs included in the mass balance approach, there were no statistically significant linear relationships between net fluxes and input loads, or TN:TP ratios. There were also no significant linear relationships between v f and nutrient loading for either TP or TN. Three reservoirs had particularly high TN retention (>50 % of inputs), all of which were located within the Wabash River subbasin (512-WAB). These three Wabash reservoirs shared a suite of features including: (a) very high TN loading, with input yields >3000 kg km−2 year−1 (Table 5) and annual flow-weighted input concentrations > 8 mg L−1, (b) shallow-to-intermediate mean depth, in the range of 4–8 m, (c) intermediate-to-high water residence times (40–210 days, Table 5), and (d) location at a lower latitude (<41° N) than most other sites, and thus potentially warmer water temperatures.

Fig. 7
figure 7

Reservoir net nutrient fluxes in relation to water residence time and hydraulic loading, determined using the mass balance approach. G Great Lakes, O Ohio River, M Mississippi River. Fitted lines are significant (TP, log10 water residence time, p = 0.026, r 2 = 0.38; TP, log10 hydraulic load, p = 0.034, r 2 = 0.27; TN, log10 hydraulic load, p = 0.025, r 2 = 0.28)

Full size image

Discussion

Our results provide substantial evidence that reservoirs broadly influence the transport of N and P at time frames >1 year, across multiple hydrologic settings and glacial histories of our large Midwestern US study area, including N- and P-rich agricultural landscapes. These reservoir influences translate as substantial modifications to the downstream nutrient supply, and there were important distinctions between reservoir effects on N compared to P. For N, net losses in reservoirs commonly exceeded 20 % of the inputs, suggesting that denitrification or nutrient burial reduce downstream N loading. For P, mass balances at 7 out of 14 reservoirs also showed net losses exceeding 20 % of the inputs (i.e., negative net flux), although four reservoirs showed net P export (i.e., positive net flux). Thus, we provide new evidence that some reservoirs are now net P sources within river networks, likely due to mobilization of legacy P from internal stores or exceedance of internal P storage capacity. At the current time, it remains unclear how long-term accumulation of nutrients and sediment, or changes in reservoir management, may alter nutrient storage or denitrification in the future. Nonetheless, given multiple considerations of reservoirs in this work, such as their placement within nutrient-rich agricultural river networks, and their open nature, future shifts in reservoir nutrient retention/loss may alter nutrient supply to downstream water bodies, perhaps dramatically.

Previous studies using models of large basins and empirical analyses of individual reservoirs have demonstrated net reservoir retention of N (Alexander et al. 2000; David et al. 2006; Harrison et al. 2009; Seitzinger et al. 2002) and P (Brown et al. 2011), yet there have been few comparisons of the effects of agricultural reservoirs on N versus P transport examined over multiple years. Additionally, there remain unanswered questions regarding cross-site variation, distinct effects of reservoirs apart from naturally-formed lakes, as well as the transferability of both local- and regional-level findings to the intermediate spatial scales most relevant to land management (e.g., hydrologic units). We also found evidence that reservoirs can buffer interannual variability in river TP yields, as reported by Powers et al. (2014), perhaps due to high P sedimentation during high loading periods or stable P outflows owing to consistent P recycling or entrainment (Soranno et al. 1997). Basic justifications for our interest in agricultural reservoirs are evident in Fig. 2, which shows approximately half of the reservoirs of our analysis are positioned in cropland-dominated areas (>50 % of contributing area as cropland), and also outflow sites below natural lakes (excluded from analysis due to insufficient sample size) were outnumbered by reservoir outflow sites by more than 5:1. Overall, our results emphasize the importance of understanding the role of reservoirs in N and P transport through agricultural river networks.

Effects of reservoirs inferred from watershed comparison and mass balance

Results from the watershed comparison approach (Figs. 2, 3) suggest reservoirs are sinks for N and P across a large extent of the Midwest, thus reducing nutrient export to downstream ecosystems. Additionally, results from the mass balance approach (Fig. 6) provide more direct evidence, from the available subset of sites, that reservoirs are generally N and P sinks and that differences in nutrient export from the watershed comparison approach are indeed associated with reservoir processes. There was also reasonable agreement between the two approaches in terms of the large magnitude of reservoir effects, with net reservoir nutrient losses commonly >20 % of inputs, and in certain cases >40 % of inputs. This agreement confirms the utility of the watershed comparison approach and yields (load per area per time) as a proxy for the reservoir mass balances, as loads scale strongly with drainage area, and thus the yield from any one tributary provides a reasonable measure of total input load once multiplied by the drainage ratio (total area:monitored site area). Other studies have reported similarly large reservoir N and P losses as a percentage of inputs, but not with such a focus on many agricultural reservoirs that receive high N and P loads and may disproportionately influence nutrient transport to large rivers and the coastal zone. For example, reservoir P retention represented 34 % of the inputs to reservoirs along large rivers across the entire US (Smith et al. 1997) including non-agricultural areas, while reservoir N retention represented 0–32 % of the inputs to 17 temperate reservoirs from five studies compiled by Harrison et al. (2009) and 58 % of the inputs to one large agricultural reservoir in Illinois (David et al. 2006). We add that despite generally high TN inputs, the average TN uptake velocity for reservoirs of our analysis, 26.4 m year−1 (n = 15), exceeds the Harrison et al. (2009) averages of 9.35 m year−1 for TN in temperate reservoirs (n = 17) and 17.7 m year−1 for TN in tropical reservoirs (n = 18).

Compared to P, a larger number of reservoirs were net sinks for N. Nonetheless, reservoir net sinks (and sources) indicated by the mass balance approach were not universally found in watershed comparisons (i.e., we found no significant differences between reservoir outflow and unimpeded sites in several subbasins). This apparent discrepancy between analyses indicates that effects of reservoirs can sometimes be masked at larger spatial scales, including the watershed level and regional level, especially when there is substantial variation in nutrient loading or site-specific reservoir processes. Consequently, although we found no statistically significant difference in nutrient export between reservoir outflow sites and unimpeded sites of MISS basin (Table 2; Fig. 3), this could be explained in part by underlying spatial heterogeneity in reservoir functioning within the basin. In addition, underlying log-transformed yields for each subbasin (TP in Fig. 4, TN in Fig. 5) showed that some subbasins with non-significant b res (i.e., no significant influence of reservoirs/lakes after Bonferroni) were limited by low sample sizes (e.g., subbasin 710-DESM), and also this statistical approach is somewhat conservative, so we cannot preclude reservoir influences in these subbasins. Overall, there was a predominant pattern of net nutrient loss within reservoirs, but also moderate variation in flux magnitude among landscapes and sites, especially for P.

Nutrient loading and reservoir design with respect to landscape context

We found that reservoirs consistently retained 20–50 % of N inputs, which is striking given the high inputs of agricultural N. When alternatively viewed on the basis of total mass (Outputs–Inputs), there was more cross-site variation in N retention across reservoirs, reflecting cross-site differences in N loading. Thus, reservoir functioning should not be viewed as independent of landscape context. For example, when compared to GRL basin, MISS and OHIO basins both contain higher average N loading associated with very high rowcrop (largely corn/soybean) production (Monfreda et al. 2008), which under uniform percent loss produces higher total mass loss (per year) in reservoirs. While GRL basin also contains high rowcrop production in the southern areas, a substantial proportion of the basin is mixed cropland, forest, grassland, and wetland, producing generally lower cropland density and TN yields (interquartile range of 380–1200 kg km−2 year−1). Similarly, MISS basin has abundant agricultural production, but is more heavily forested in the north (Homer et al. 2007), producing a wider gradient for cropland density, and subsequent N loading, than OHIO and GRL basins (TN yield interquartile range of 510–2600 kg km−2 year−1 in MISS; Table 2). Consequently, reservoirs of the GRL basin, despite being important on the basis of % input, may be modulating a smaller total flux of nutrients (both P and N) than reservoirs in the MISS and OHIO basins. Overall, P showed more residual variation than N along the increasing pattern with cropland (Fig. 2), perhaps in part due to point P sources, but the general trends of increasing N yields with cropland, and reduced N yields at reservoir outflow sites, were paralleled for P.

For reservoirs in the mass balance approach, there were no simple declines in TN retention (as % of input) or uptake velocity as a function of TN loading. This potentially suggests the capacity of the agricultural reservoirs to remove landscape N loads has not saturated, and that N-rich agricultural reservoirs have a robust denitrification capacity, although interactions with other controlling variables are possible.

Reservoir size and depth range widely in the study area, and consideration of these variations in reservoir structure may facilitate interpretation about N and P functioning. Several large reservoirs occur in the MISS and OHIO basins and were included in our study, especially the southernmost subbasins (e.g., 506-SCIO, 512-WAB, 708-MISI, 710-DESM). In contrast, reservoir outflow sites of the GRL basin were often below smaller reservoirs, mill ponds, and in-town dams, as indicated by the lower median surface area (0.26 km2, Table 2) compared to OHIO (3.9 km2) and MISS (0.39 km2). Meanwhile, moderate heterogeneity in topography also occurs within the study area, and was clearly linked to heterogeneity in reservoir structure. For instance, many reservoirs in the OHIO basin are positioned in hillier terrain (e.g., subbasin 508-GRMI) which likely relates to the higher mean depth of OHIO reservoirs (Table 5) than the monitored reservoirs in low-gradient landscapes within the MISS and GRL basins. Also, in the case of one subbasin with especially low topography, 410-LER (Lake Erie) which includes Maumee River basin, we were unable to detect reservoir effects on P, likely due to a scarcity of major reservoirs in this area. Recognizing that shallow reservoirs and lakes can be prone to nutrient resuspension and recycling (Wetzel 2001), it is possible that deeper reservoirs such as those occurring within certain OHIO tributaries could be disproportionately important sites of long-term nutrient storage. In contrast, the prospect of continued nutrient burial is questionable for systems that have already been subjected to extensive eutrophication or sedimentation, although small shallow systems might remain persistent hot spots for denitrification.

Hydraulic and physico-chemical controls on reservoir N and P functioning

For reservoirs in the mass balance approach, there were no simple declines in TN retention (as % of input) or uptake velocity as a function of TN loading. This potentially suggests the capacity of the agricultural reservoirs to remove landscape N loads has not saturated, and that N-rich agricultural reservoirs have a robust denitrification capacity, although interactions with other controlling variables are possible. There was also evidence that reservoirs with shorter water residence time and higher hydraulic loading had a diminished capacity to retain P and N (Fig. 7), consistent with many prior mass balance studies of aquatic systems (e.g., Koiv et al. 2011; Persson et al. 1999; Seitzinger et al. 2002). Still, there was substantial residual variation in these retention–hydraulics relationships, perhaps due to differences in the mobility and fate of internal nutrient pools, and thus P and N residence time residence time, or differences in reservoir management. Compared to TP, TN losses were more consistent across sites (CV of 0.90 for TN compared to 1.20 for TP), perhaps with the exception of three reservoirs in the Wabash River subbasin that had especially high TN losses (>50 % of inputs).

Recently, Finlay et al. (2013) implicated successful land management and wastewater P removal as explanations for declining P inputs to large natural lakes. We add that long-term declines in P inputs to large natural lakes, as well as reservoirs, may also have been made possible by retention of P in upstream reservoirs. At the current time, it remains unclear how strongly past or future trends in nutrient loading to aquatic systems may be linked to changes in retention/loss of upstream reservoirs. It is also questionable whether water quality trends of larger natural lakes should foretell those of agricultural reservoirs.

In the mass balance approach, reservoirs in Wabash subbasin (512-WAB) with shallow-to-intermediate depth had disproportionately high N losses, and these results warrant further consideration. This may indicate that shallow littoral sediments have a high denitrification rate due to the co-occurrence of anoxia, nitrate supply, and carbon supply (Bruesewitz et al. 2012). Further, in order for denitrification to be maximized at the ecosystem-level over the annual time frame, zones or patches of sediment denitrifier activity must occupy a sufficiently large spatial or temporal extent, and N sources originating in surface water (oxic) must be transferred across the sediment–water interface into sedimentary (anoxic) N sinks. In shallow eutrophic systems like the Wabash reservoirs, such favorable conditions for denitrification may develop through: (a) greater susceptibility to hydrodynamic events that deliver pulses of water N (and/or C) to sediments (Cook et al. 2006), such as through mixing events (Grantz et al. 2012), wave actions (Risgaard-Petersen et al. 1999), or water level fluctuations, (b) higher sediment/water contact (Alexander et al. 2000) during otherwise stable hydrodynamic conditions, which could again result in larger or more consistent N supply to the benthos, (c) more widespread or rapid oxygen depletion of sediments (David et al. 2006; Wall et al. 2005), resulting in a larger area or depth of denitrifying zones, or their more frequent establishment. In addition, we point to a possible (d) thermal effect on denitrifier activity that could accelerate denitrification in warmer, shallow waters (Saunders and Kalff 2001). The apparently high denitrification of Wabash reservoirs combined with their position in lower, warmer latitudes is consistent with d (thermal effect), but could also indicate a mixture of a (unique hydrodynamics), b (higher sediment/water contact), and c (expanded oxygen depletion of N-rich sediments) due to the strong linkage between thermal regime and mixing regime in lakes and reservoirs (Straskraba et al. 1993; Wetzel 2001). We caution these are not the only explanations, and meanwhile further study of short-term and long-term N fate in agricultural reservoirs is needed to isolate denitrification apart from other processes such as N burial. In addition, sedimentation is a major concern for reservoirs in the Wabash River subbasin, as well as many others (Palmieri et al. 2001), because reservoir functioning is often compromised by infilling sediment, with potential effects on N fate that currently are not well understood.

Reservoir N and P fluxes in wet versus dry years

In general, the direction (sign) of the averaged annual net flux at each individual reservoir was maintained during both wet years and dry years, although in some cases, net fluxes were exaggerated in wet or dry years. For example, a subset of reservoirs had higher N retention (as % of inputs) during dry years (Fig. 6), perhaps explained by either higher denitrification in dry years (David et al. 2006), or higher water residence time and associated higher biological N uptake or algal-N settlement. Additionally, in three of four reservoirs that had net P export on average, net export (as a %) was higher in wet years, perhaps caused by flushing and resuspension of P from within the reservoir, or by a decrease in sedimentation rates. Nonetheless, three other MISS reservoirs actually had higher P retention during wet years, so the above trends were not universal and could be mediated by site-specific details such as water level management. More generally, the lack of simple linear relationships between net fluxes and mass inputs, or TN:TP ratios, likely indicates the importance of interactions among multiple controlling variables across the diverse reservoirs represented in our data set. Interactions among reservoir features such hydraulic loading (Seitzinger et al. 2002), uptake velocity (Harrison et al. 2009), nutrient stoichiometry (Grantz et al. 2014), sedimentation (Graf et al. 2010), timing and form of inputs/outputs, and of course reservoir management, all remain areas of needed research.

Influence of monitoring station placement and drainage size

It is important to recognize that most rivers included in our study had intermediate-to-large drainage areas (72 % of sites had drainage areas of 50–5000 km2, including 37 % <1000 km2), as these have been the focus of major water quality monitoring effort (Robertson and Saad 2011; Smith et al. 2003). The tendency for larger reservoirs to be positioned lower in the landscape, and thus within larger catchments, could mean that a relatively small number of systems in our analysis have a disproportionate influence on nutrient export to large rivers and coastal zones. This reflects that reservoirs in larger drainages receive a substantial proportion of landscape P and N losses, and also are closer to river mouths than many other aquatic systems. While smaller headwater stream networks contain reservoirs and ponds that can also influence river export (Downing et al. 2008; Powers et al. 2013), the collective effects of these systems on nutrient transport is not yet sufficiently understood and probably cannot be quantified using monitoring information from intermediate-to-large catchments alone. Further, it is difficult to make inferences about reservoir processes for certain landscapes in our analysis. For example, in suburban areas near Chicago (subbasin 712-ILU), reservoir outflow sites were mostly positioned below very small reservoirs and ponds that did not show significant effects in our watershed comparison approach. This lack of significance could indicate suburban reservoirs and ponds do not affect river N or P, but we are not confident in that assertion because the unimpeded yields of subbasin 712-ILU were highly variable, likely due to heterogeneous nutrient sources. Similarly, several large reservoirs in Iowa and southern Illinois (710-DESM, 713-LIL, 714-MIUK) are surrounded by intense agriculture and may be substantial sinks for N or P, but we were unable to examine these directly with our mass balance approach, due to insufficient monitoring of either reservoir inflows or outflows. In addition, while cropland is known to contribute the majority of N to rivers in the study region, leading to especially high N:P ratios, animal agriculture may also influence reservoir functioning, especially for P. Overall, the above issues emphasize a need for coordinated monitoring of reservoir inputs and outputs in drainages of multiple types and sizes.

Conclusions

Our analysis suggests that reservoirs in agricultural landscapes commonly function as net nutrient sinks, but there was variation in the magnitude of nutrient losses among sites, subbasins, and basins, and differences between N and P. In general, there was more cross-site variation in reservoir P retention compared to N retention, partly reflecting that reservoirs were consistent net N sinks, but sometimes net P sources. Because many reservoirs are positioned between the source of nutrients (e.g., fertilizer application on farmland) and the site of long-term water quality monitoring (i.e., larger rivers), internal processes of reservoirs that influence river N and P export can mask or delay river responses to land management (Hamilton 2012; Magnuson 1990; Powers et al. 2014; Sharpley et al. 2013). Consequently, investigations of the efficacy of management practices in reducing watershed N and P delivery may benefit from the incorporation of reservoir effects. We also emphasize that we are not advocating new construction of large dams as a means of water quality improvement. Small dams and ponds, however, often occur in landscapes where natural wetlands have been lost to agriculture, and these can potentially be adapatively managed for nutrient retention and other ecosystem services.

Reservoirs may buffer downstream aquatic ecosystems from nutrient pollution, yet reservoir functions could be vulnerable to future change, in part because nutrients cannot accumulate within bottom sediments indefinitely. Currently there is a need for a more complete understanding about how linked physical and biogeochemical processes in reservoirs change over the long-term (decades–centuries), especially in relation to sedimentation, eutrophication, infrastructure aging, and reservoir management issues including timing of outflows, water level control, dredging, and dam removal. Overall, our results show widespread though not universal effects of reservoirs on the transport of nutrients lost from agricultural land, which begs a question: how will future reservoir management influence water quality beyond the boundaries of these ecosystems? In the meantime, reservoir operations and management may be influencing N and P export to downstream waters, but are not well coordinated around this issue. Unknowns about reservoir nutrient pools and fluxes, in particular their vulnerability to future change, provide new motivation for improved land management and source reduction in the contributing agricultural catchments.