1 Introduction

In the last decades, the use of nitrogen fertilizers has increased exponentially at a global scale (Vitousek 1997; Galloway 2008) with detrimental consequences on freshwater and marine ecosystems (Jorgensen and Richardson 1996; Ongley 1996), on nitrous oxide emissions and greenhouse effect (Turner 1991; Gruber and Galloway 2008). Groundwater (Nolan et al. 2002; Foster 2000) and drinking water contamination (Höring and Chapman 2004) are causing serious health problems worldwide, such as methahemoglobinemia or “Blue Child Syndrome” (Cynthia et al. 2002; Fan and Steinberg 1996), congenital malformations (Dorsch et al. 1984) and different forms of cancer (Hill et al. 1973; Ward et al. 1996).

The European Environment Agency showed that high NO −3 concentrations in groundwater across the EU (EEA 1999) is a widespread and serious problem. For example in England, over 70% of surface and groundwater NO −3 originated from agricultural activities (Kraft et al. 2008). A significant fraction of drinking water is exposed to these risks, as supported by the long history of European law amendments for water protection, starting with the Nitrate Directive (OJEC 1991) to the Water Framework Directive (OJEC 2000).

The Po River valley is the largest Italian alluvial plain and is heavily affected by agricultural pollution, especially NO −3 (Onorati et al. 2006; Cinnirella et al. 2005; Giuliano 1995). In addition to synthetic nitrogen fertilizers, intense livestock farming and a high population density produce an unsustainable nitrogen load. Mass balance calculations have indicated that, in some districts, theoretical nitrogen load from all cited sources accounts for more than 1,000 kg N ha−1 year−1 (Palmeri et al. 2005) which is a much larger amount than any maximum crop nitrogen uptake.

Once manure or synthetic urea are distributed, several physical and biological processes regulate the fate of hydrolyzed ammonia. Ammonia may adsorb onto soil particles, be assimilated by crops, volatilized to the atmosphere, or be nitrified to nitrate. Nitrate may be assimilated by crops as well or lost in runoff water and, more likely, be leached to the water table and be partly denitrified along the way to N2O or to N2, with an irreversible loss of nitrogen to the atmosphere (Shomar et al. 2008; Bernot and Dodds 2005; Burt et al. 1999; Martin et al. 1999). In fact, when water percolates in the subsurface, oxygen (O2) may quickly decline and NO −3 can be employed by denitrifying bacteria as electron acceptor (Appelo and Postma 2005). In certain soils, the degree of water saturation is the most important variable, regulating oxygen mobility and consequently, oxidative conditions at microscales, where denitrification takes place (Fenchel et al. 1998). Organic matter availability is the other regulator of denitrification (Hofstra and Bouwman 2005) and if labile substrates are not abundant, denitrification occurs at small rates and NO −3 can travel long distances within aquifers (Bölke and Denver 1995).

Following this rationale, a small-scale field site within a shallow unconfined aquifer was monitored to quantify the time required to completely remove NO −3 . In 2004, the field site was converted into a recreational area and no more N fertilizers were applied on it and upstream of it. Thus, this experiment represents an example of integral NO −3 protection area of particular interest since the field site lies on a river paleochannel, a preferential geological structure for contaminants migration (Spalding et al. 1993; Meinardus et al. 2002). The main focus of this study was to determine the mechanism of NO −3 attenuation or persistence in groundwater. For this purpose, the data set was used to implement scenario modeling to evaluate the role of recharge and aquifer flow transient conditions in NO −3 natural attenuation.

2 Materials and Methods

2.1 Field Site

The study site, consisting of a monitoring grid of 40 × 40 m, is located in a rural area (44°51′32 N, 11°39′15 E) between the town of Ferrara and the Po River main course, at about 100 km from the outlets into the Adriatic Sea, NE Italy (Fig. 1). The test site has been investigated with a combination of core logs and electrical resistivity tomography (ERT) techniques, from which a conceptual and numerical flow model was established (Mastrocicco et al. 2009). Briefly, the hydrogeological units present in the test site are the unconfined aquifer composed of recent fluvial sandy deposits with clay and silt lenses (from 0 to around 5 m b.g.l.) and the underlying aquiclude constituted of clay and silt sediments (from 5 to almost 14 m b.g.l.). The site was cultivated from 1986 to 2004 under a rotation of cereals, mainly maize and wheat, using urea as nitrogen fertilizer at an average rate of 300 kg/ha/year. From September 2004, the field was converted into a park with grass cover. Since then, no fertilizers and other agrochemicals were applied. Potential nitrate sources from upstream were not detected from existing shallow boreholes located immediate upstream of the field site (see Fig. 1), which never showed nitrate ammonium and nitrite concentrations above 3 mg/l. At the end of May 2007, piezometers were installed to monitor groundwater flow in the unconfined aquifer of the paleochannel. Piezometric variations during the year were attributed to variable recharge rate and water level changes in canals surrounding the study area. Evapotranspiration played an important role in determining recharge rate but it was not directly controlling the heads fluctuation since maximum root depth of grass was 0.3 m b.g.l., while groundwater table usually lied 2 m b.g.l and it was unlikely that the root system was fed by capillary rise.

Fig. 1
figure 1

Location of the study area and potential nitrate point sources, i.e., isolated houses not connected to the sewage network (red rectangle) and recent buildings connected to the sewage network (green dashed areas); principal canals (light blue lines) and their direction (light blue arrows); contour lines of groundwater elevation in m above sea level (dark blue lines), principal groundwater directions (dark blue arrows) and location of shallow groundwater wells (red dots) used for groundwater elevation monitoring and for chemical analysis (red crosses)

Full size image

2.2 Analytical and Field Methods

LTC M10 Levelogger Solinst® dataloggers were placed in piezometers A3, C1, C5, and E3 (Fig. 2) to monitor hourly groundwater level, electrical conductivity, and temperature. The other piezometers were monitored for groundwater level and sampled after low-flow purging for major ions determination during eight surveys. Vertical aquifer discretization was accomplished by straddle removable packers sampling, each piezometer was sampled every meter via low flow purging and sampling method.

Fig. 2
figure 2

Plan view of the site, in gray are the paleochannel boundaries, cross-circles are the monitoring wells, half circles P1 and P2 are the piezometers used for heads boundaries and S1, S2, S3, S4 are the core logging sites used to characterize the unsaturated zone

Full size image

A meteorological station recording daily rainfall, wind speed, temperature, and humidity was located on the field site while solar radiation data were recorded in another station located 4.6 km from the site. Data are available on-line from the meteorological regional service (www.dexter.it).

The unsaturated zone was sampled at the beginning of this study by means of auger coring. Four samples were taken at depths from 0 to 2 m b.g.l. (Fig. 2), and sediments were analyzed for major anions and cations. Unsaturated zone sediment analysis consisted of a batch with a sediment/water ratio of 1:10, using 10 g of air dried sample dispersed in 100 ml of Milly-Q water (Millipore, USA). A biological inhibitor (1 g/l phenylmercuric acetate) was added to prevent microbial activity and the solution was stirred for 1 h and then allowed to stand for 1 day. The insoluble residue was removed by filtration and analyzed for major cations and anions.

In-well parameters were determined with the HANNA Multi 340i instrument which includes a HIcell-31 pH combined electrode with a built-in temperature sensor for pH measurements, a CellOx 325 galvanic oxygen sensor for DO measurements, a combined AgCl-Pt electrode for Eh measurement and a HIcell-21 electrode conductivity cell for EC measurements. Samples were filtered through 0.22 μm Dionex vial caps. The major cations, anions and oxianions (acetate and formate) were determined with isocratic dual pump ion chromatography ICS-1000 Dionex, equipped with an AS9-HC 4 × 250 mm high capacity column and an ASRS-ULTRA 4 mm self-suppressor for anions, and a CS12A 4 × 250 mm high capacity column and a CSRS-ULTRA 4 mm self-suppressor for cations. An AS-40 Dionex auto-sampler was employed to run the analyses, quality control (QC) samples were run every ten samples. The standard deviation for all QC samples run was better than 4%. Charge balance errors in all analyses were less than 5% and predominantly less than 3%. Organic carbon was determined with a carbon analyzer (Carbon Analyzer Shimadzu TOC-V-CSM) after acidification with one drop of 2 M HCl to remove dissolved carbonate.

2.3 Modeling

Recharge was assumed to occur homogeneously over the study area; the recharge rate was calculated with the finite element model HYDRUS-1D (Šimunek et al. 2008). The HYDRUS-1D numerical grid was subdivided in 200 nodes of 0.01 m each, to form a regular 2 m long grid. The grid was subdivided into two regions representing the upper (from top soil to 0.8 m b.g.l.) and the lower soil horizons (0.8–2 m b.g.l.). Initial water content conditions of collected soil cores (every 0.25 m) at each site were measured with the gravimetric method and interpolated linearly along the vertical axis.

At the soil surface, variable flux and head conditions were selected to represent the atmospheric boundary. Daily reference evapotranspiration was calculated with the FAO-56 recommended Penman–Monteith equation (Allen et al. 1998). It was assumed that the grass cover at the experimental site resembles the reference surface as defined by Allen et al. (1998). Potential transpiration and evaporation were split using an estimated surface cover fraction of 0.8 (Ritchie 1972). Root water uptake was simulated using the grass dimensionless water stress response function available in the HYDRUS-1D database (Feddes et al. 1978). Run-off was assumed negligible due to the flat topography and based on other studies in this area (Antonellini et al. 2008). Soil hydraulic parameters were assigned using default values included in the soil catalog of HYDRUS-1D for silt-loam (upper layer) and sand (lower layer). Free drainage was selected as lower boundary condition and the cumulative bottom flux allowed the estimation of groundwater recharge for different periods.

A fully transient 3D flow model using MODFLOW-2005 was implemented and calibrated versus heads monitored in 13 observation wells recorded during eight piezometric campaigns, and versus continuous data recorded with the A3, C1, C5, and E3 piezometers. The piezometric campaigns were performed on the following dates: 19/06/07, 18/07/07, 18/08/07, 29/09/07, 17/01/08, 4/08/08, 10/02/09, and 16/06/09. The numerical grid of MODFLOW-2005 was discretized by a regularly spaced grid of 0.5 × 0.5 m. Vertically, between 4.3 and −1.7 m a.s.l., the model domain was discretized into six layers each of 1 m thickness to represent vertical variations in hydraulic conductivity. The permeability distribution was derived by a combination of multi level slug tests and ERT techniques (Mastrocicco et al. 2009). The total simulation time of 730 days (from 18 June 2007 to 17 June 2009) was subdivided into 13 different stress periods to reproduce groundwater fluctuations recorded with dataloggers. The number of stress periods was selected after preliminary modeling using an initial number of 110 stress periods, then decreasing the number of stress periods until the correlation coefficient (R 2) of the heads residuals (difference between calculated and measured values) started decreasing compared to the initial value. This exercise allowed obtaining the minimum number of stress periods needed to achieve an accurate simulation of the heads temporal variation. The time-variant specified-head package (Harbaugh et al. 2000) was used to represent the northern and southern boundaries. No flow was assumed along the eastern and western sides, since they consist of impermeable silty clay sediments (Mastrocicco et al. 2009). Recharge rate was assigned by summing up the daily recharge calculated with HYDRUS-1D in each stress period, and it was assumed uniform across the model domain.

The transport and reaction processes were simulated using MT3DMS code (Zheng and Wang 1999). Transport parameters like longitudinal, transversal, and vertical dispersivity were initially assumed to be 2, 0.2, and 0.02 m, the molecular diffusion coefficient was assumed to be 1e−9 m2/s (Gelhar 1993), while the effective porosity was assumed to be 0.3 for sandy sediments and 0.4 for silty sediments. The advection dispersion term was solved using the total variation diminishing scheme with a Courant number of 0.1, since it is the most robust numerical solution method and is virtually free of numerical dispersion. Dispersivity and effective porosity parameters were adjusted via trial and error calibration (Zheng and Bennett 2002) to fit the Cl− concentrations observed in eight sampling campaigns performed at the same time as the piezometric campaigns. After calibration of the conservative transport model fitted well Cl− observed concentrations (where only dilution processes were accounted for), the denitrification rate was determined by tuning it at incremental steps (via trial and error calibration) to fit the NO −3 observed concentrations.

3 Results and Discussion

3.1 Unsaturated Zone Characterization and Recharge Estimation

Low concentrations of mineral nitrogen species were found in the unsaturated zone (from 0 to 2 m b.g.l.) from auger coring at the beginning of the experiment. Average NO −3 and Cl− concentrations were respectively 2.5 ± 0.4 and 4.7 ± 2.6 mg/kg of dried sediments, while nitrite and ammonium were found occasionally, always at concentrations close to the limit of detection (0.10 ppm for nitrite and 0.30 ppm for ammonium). The very low concentration of mineral nitrogen found in the soil resulted from leaching after the cessation of fertilizer use in April 2004. At the beginning of this study, the highest groundwater concentrations of NO −3 and Cl− (used as conservative tracer) were 102.0 and 201.1 mg/l, respectively, while average concentration were 34.2 and 87.6 mg/l. Very low concentrations of organic carbon were present in groundwater, with an average value of 0.83 mg/l and a standard deviation of 0.56 mg/l. This suggests a carbon limiting condition for denitrification (Rivett et al. 2008).

From Fig. 3, it is evident that actual evapotranspiration (ETa) takes place throughout the year, although at slower rates during winter time because the evaporative demand is less pronounced. The precipitation distribution pattern is bimodal with maximum peaks in autumn and spring, and sparse storm events during summer periods. This is typical of sub-coastal temperate climates characterized by cold winters and warm summers (Pavan et al. 2008). The cumulative rainfall in the monitoring period was 1,586 mm, while estimated recharge was 599 mm. The 2007/08 season was much drier than 2008/2009, and the calculated recharge was 147 mm in 2007/08 and 452 mm in 2008/09. From Fig. 3, it is also evident that recharge took place only during late winter–early spring.

Fig. 3
figure 3

Calculated cumulative actual ETa plotted with a dashed line, calculated cumulative recharge (upper plot) plotted with a continuous line, and daily rainfall (lower plot) recorded with the onsite meteorological station

Full size image

3.2 Transient Groundwater Flow Model

The transient flow model was calibrated without changing the parameters of the original model (Mastrocicco et al. 2009); only boundary conditions were updated for the additional monitoring period with actual measured heads. The absolute residual mean (ARM) error was 0.058 m with a normalized root mean square (NRMS) error of 5.79% and a R 2 of 0.97. Figure 4 shows the overall good match between observed and calculated heads and the time series highlight a cyclical variation of the piezometric heads with minimum values at the end of the summer season and peaks during spring. The groundwater fluctuation was mainly driven by the canal located up-gradient as shown in Fig. 4, while recharge during the late winter and spring seasons prevented the groundwater to suddenly drawdown when the canal level decreased.

Fig. 4
figure 4

Scatter diagram of the calibrated transient flow model comparing observed and computed heads in all observation wells (left) and time series graph comparing observed and computed heads of two piezometers at the opposite sides of the monitoring grid, A3 and E3 (right)

Full size image

The groundwater flow was always in the direction from A3 to E3, since the piezometric level of the latter was always lower than the former; but the hydraulic gradient varied during the year, as the head difference between A3 and E3 was not constant. Variable flow therefore occurred during the year. In the high recharge period (during winter 2008/09), the head difference became very small and a flat groundwater table established.

3.3 Transient Solute Transport Model

The initial concentrations of Cl− and NO −3 , observed in the first survey on 19/06/07, displayed a similar spatial distribution pattern (Fig. 5). This pattern was captured by multi-level sampling, with increasing concentrations from the model inflow (lower side) to the model outflow (upper side).

Fig. 5
figure 5

Three-dimensional box plot contour maps of initial observed Cl− and NO −3 concentrations and location of monitoring wells

Full size image

The concentration distribution, with increasing values from the inflow to the outflow, was due to gradual replacement of NO −3 polluted groundwater with NO −3 free (≈1 mg/l) groundwater coming from upstream and through recharge.

Trial and error calibration of dispersivity and porosity produced satisfactory results with little variation from the initial values; in fact, only longitudinal dispersivity was adjusted, while its ratio with transversal and vertical dispersivity was left unchanged. Final values were 0.53, 0.053, and 0.0053 m for longitudinal, transversal, and vertical dispersivity, respectively. These low values suggest a relatively homogeneous transport of solutes dominated by advection. Effective porosity also needed a small refinement of initial values with final values of 0.31 for sandy sediments and 0.42 for silty sediments. The calibrated conservative transport model showed a good match between calculated and observed concentrations (Fig. 6), with an ARM error of 5.43 mg/l, a NRMS error of 4.54%, and a R 2 of 0.977.

Fig. 6
figure 6

Scatter diagram comparing observed and computed Cl− concentrations in all observation wells (left) and time series graph comparing observed and computed Cl− concentrations in piezometers A3, C3 and E3 (right)

Full size image

A decrease in Cl− concentration was recorded from the beginning of the monitoring period in all observation wells, approaching the pristine Cl− concentration (from 28 to 41 mg/l) at the end of the monitoring period. The initial difference in Cl− concentration in the wells shown in Fig. 6 was due to dilution processes with pristine groundwater and through recharge. Probably, a cyclical Cl− input through agricultural amendments (that was stopped in 2004) provided elevated Cl− in the whole area, but since monitoring started in 2007, the dilution effect was more evident in the wells located upstream with respect to those located downstream. A complete clean up from Cl− was reached within the monitoring grid at the end of summer 2008, although it is difficult to provide an exact time when Cl− concentrations became stable.

The behavior of NO −3 was similar to Cl−, since a decrease in concentration was recorded from the beginning of the monitoring period. NO −3 reached concentrations below 10 mg/l in January 2009 (Fig. 7). Simulated NO −3 concentrations fitted reasonably well observed data using a first order degradation rate of 2.5e−3 mg-NO −3 /l/d, but the scatter plot of Fig. 7 shows a less accurate fit compared to Cl−.

Fig. 7
figure 7

Scatter diagram comparing observed and computed NO −3 concentrations in all observation wells (left) and time series graph comparing observed and computed NO −3 concentrations in piezometers A3, C3 and E3 (right). The shaded area shows results for E3 without degradation (upper area border) and with a doubled degradation rate (lower area border)

Full size image

However, an ARM error of 4.11 mg/l, a NRMS error of 5.71%, and an R 2 of 0.960, indicated that the reproduction of observed concentrations was still robust. In addition, the calibrated numerical model was compared to a conservative (no degradation) NO −3 dilution scenario and to a higher degradation rate (5e−3 mg-NO −3 /l/d) scenario. The NO −3 conservative scenario showed that (at least after the first year of monitoring) the observed concentrations cannot be matched with calculated ones, as observed NO −3 were always lower than those calculated (Fig. 7). A 100% increase of the degradation rate produced unrealistically low NO −3 concentrations. Consequently, model results were found to be sensitive to changes in the degradation rate input value. To quantify this aspect the normalized sensitivity coefficient (Zheng and Bennett 2002) was calculated for a 100% perturbation of the NO −3 degradation rate with respect to the total computed mass of NO −3 during the simulation, and found to be 0.26.

3.4 Comparison with Steady State Results

The results obtained from transient flow and transport models were judged satisfactory. Despite this, numerical models often assume steady state flow conditions and this may not be representative of field conditions. To clarify this aspect, a series of numerical scenarios were run. The transient flow model boundary conditions were time variant heads and variable recharge rate; both these terms were varied by means of scenario modeling. In general, the first term contributes significantly to change the average flow velocity field, while the second term contributes to dilution processes through recharge. Different scenarios were built up and compared to the transient calibrated model (Table 1). Starting from the most complex transient model, the temporal discretization was decreased toward steady state conditions, by decreasing the number of stress periods, Sp (Table 1). Flow model results in each scenario were compared to piezometric heads measured only at corresponding stress periods. Consequently, the steady state scenario was compared against average piezometric heads.

Table 1 Scenario modeling with decreasing number of stress periods (Sp.) from the fully transient calibrated model to a simple steady state model.
Full size table

Results of the flow simulations are summarized in Fig. 8. A deterioration of R 2 is appreciable when a decreasing number of stress periods are employed. The steady state scenario displayed poorer statistics in comparison with the calibrated model (AMR from 0.058 m to 0.045 m, NRMS from 5.79% to 8.49%); this is due to the averaging of all observed and calculated heads. The small value of AMR error in the steady state simulation is not due to a close fit between observed and calculated heads, but it occurred because the average heads used to calibrate the steady state model had a variation range of just a few centimeters within the piezometers grid, while the transient model was calibrated versus transient head data which span over 1.2 m (see Fig. 4), and consequently the head residuals were higher in the transient model. The comparison between the calibrated transient model and the other scenarios was performed also for the conservative and reactive transport simulations (Fig. 8). In this case, a deterioration of simulation results is appreciable only for the conservative tracer simulation. This was probably due to a major dilution effect most evident for the initial elevated Cl− concentrations (not shown), while this effect was smoothed out for NO −3 . The same effect was present in all scenarios because the flow field was changed and subsequently the conservative transport of Cl− was principally affected by this change. A closer inspection of Fig. 8 reveal that variable recharge is affecting the reactive and non-reactive transport more than the flow field variations. In fact, the NO −3 and Cl− best-fit scenarios were obtained using a variable recharge (transient scenario 1, 4, and 7).

Fig. 8
figure 8

Plot of R 2 for flow simulation, conservative transport (Cl−) and reactive transport (NO −3 ) of each scenario described in Table 1, the scenarios with variable recharge rate are indicated with an arrow and the acronym “VR”

Full size image

4 Conclusions

A small field-site experiment was established to evaluate the clean-up time of a shallow unconfined aquifer contaminated by NO −3 . The site represents a case of vulnerable area (river paleochannel) where agricultural practices were stopped. NO −3 persistence in the field confirmed the high vulnerability of unconfined aquifers to this common pollutant and the long bioremediation time required. NO −3 removal within the monitoring grid appeared to be mainly driven by dilution processes. The three-dimensional flow and transport modeling provided a useful tool to asses NO −3 fate, discriminating the higher effect of dilution compared to the lower effect of degradation by denitrification. The value of degradation rate determined in this study is comparable to those found recently for unconfined aquifers under agricultural fields (Green et al. 2008). Scenario modeling provided evidence that in a flat area, like the one of this study, recharge is one of the key parameters in NO −3 remediation. The flow field may not significantly affect NO −3 clean up time due to the small head gradients in these areas (usually 1‰), which are not sufficient to induce large seasonal heads and flow variations. Thus, an accurate estimation of recharge rate should be taken into account when evaluating clean up times especially in intensively cultivated areas, where a yearly input of nitrogen from fertilizer may not easily allow a noticeable effect on NO −3 removal.