The total inflow of water to the constructed wetland is assumed to consist of the upstream surface water inflow (i.e., generated from the Bells Creek catchment area upgradient of the Upstream location, Figure 2), direct overland flow from the immediate wetland catchment area, and direct groundwater discharge. As has been visually confirmed, all outflow from the wetland occurs via surface water discharge at the Outlet location (Figure 2).

Samples were collected for water quality analyses approximately monthly between May 2023 and July 2024 at the Upstream, Outlet, and Springs locations (Figure 2). Samples from the Springs area were collected using a peristaltic pump and a dedicated drive-point piezometer penetrating ∼0.6 m into the fractured sandstone bedrock below the wetland. All samples were passed through a 0.45 µm filter and were refrigerated until analysis. Samples for cation analysis were acidified using 67%–70% nitric acid. Anions were analyzed using a Metrohm ion chromatograph (761 Compact IC equipped with a Metrosep A Supp 5 column, Switzerland), and cations were analyzed using an inductively coupled plasma-optical emission spectrometer (Varian Vista MPX, Australia) at the University of New Brunswick, Department of Civil Engineering Water Quality Laboratory. Field parameters including pH, temperature, specific conductivity, and dissolved oxygen (DO) were measured using a portable water quality meter (Thermo-Scientific Orion Star A329 Portable Meter). Field alkalinity values were obtained using a HACH digital titrator (Hach, Canada). Charge balance errors were calculated for all samples and a Piper plot of the major ion results was used to assess the hydrochemical facies of the water.

To obtain high-frequency data for NO₃-N, the dominant form of nitrogen in PEI freshwaters (Somers et al., 1999; Danielescu and MacQuarrie, 2011), Submersible Ultra-Violet Nitrate Analyzers (SUNA V2, Seabird Scientific, United States), were installed at the Outlet, Upstream, and in Well #6 (Figure 2). NO₃-N concentrations were recorded at 6-h intervals. The SUNAs were powered using 40 W solar panels and 12 V batteries protected inside large plastic bins (Supplementary Figure S2). The SUNAs at the Upstream and Outlet sites were secured inside stainless-steel cages, which were then fixed within the middle of the channel by driving rebar through the gaps in the cage and into the streambed. Between August 2023-March 2024, the Upstream SUNA was equipped with a copper biofouling guard while awaiting a SUNA equipped with a wiper that would clean the optical window prior to every sampling event. The Upstream SUNA was replaced by a SUNA with a wiper in March 2024. The Outlet SUNA was equipped with a wiper for the entirety of the study period. The sensor performance was verified using SUNA’s internal diagnostics and concentration standards (i.e., 0, 0.5, 1, 2.5, 5, and 10 mg NO₃-N/L) before field deployment and on all field visits where the air temperature was above freezing to avoid damaging the optical lens. SUNA readings were also compared to concentrations measured in discrete NO₃-N samples collected at each of the locations (Supplementary Figure S3). As discussed by Miller et al. (2016), SUNAs do not explicitly account for absorbance by nitrite in the range of 210–220 nm; however, at the study site the concentration of nitrite is known to be less than laboratory detection limits.

To assess whether the NO₃-N concentrations from the Springs location were representative of the concentrations within the aquifer, and to provide groundwater NO₃-N data between the times of manual sampling, concentrations from the Springs location were compared to continuous nitrate concentrations obtained from Well #6, located approximately 400 m upgradient from the Springs location (Figure 2). To account for the lag time arising for groundwater to travel from Well #6 to the Springs sampling location, the Well #6 nitrate concentration time series was shifted forward by 748 days, which was found to provide the best correlation with the discrete samples collected from the Springs. This time shift falls within the range of travel times (519–1,600 days) estimated using the groundwater velocity range of 0.25–0.76 m/d previously estimated for the aquifer underlying Harrington Experimental Farm (Zebarth et al., 2015).

The results for dissolved chloride and NO₃-N from the manual water sampling were used with the corresponding stream discharge values, obtained on the same day as the water samples, to determine discrete (i.e., approximately monthly) loads. Chloride was used in this study as a conservative tracer as it is unlikely to be affected by biogeochemical reactions in either the subsurface or surface water system. The discrete Upstream and Outlet loads for both chloride and NO₃-N were computed using Equation 2. L D = C Q (2) where LD is the load (M/T), C is the concentration (M/L³), and Q is the corresponding discharge (L³/T).

To produce a time series of NO₃-N loads, linear interpolation was used between the SUNA concentration data (i.e., obtained at 6-h intervals) to estimate NO₃-N concentrations at 1-h intervals for all surface water monitoring locations. The NO₃-N load was then calculated at 1-h intervals for the duration of the study period using Equation 2. Equation 3 was used to calculate the cumulative NO₃-N load over a particular time interval (e.g., week, month). L D ∆ t = ∑ i = 1 n C i Q i ∆ t (3) where ∆ t is the time interval used in the concentration and discharge time series [T] and n is the number of observations within the interval.

During the study there were several periods when the SUNA at the Upstream site did not record NO₃-N concentrations (equivalent to 119 days of missing data or ∼33% of the study period), and thus the Upstream NO₃-N loads could not be directly computed. For the first nine months of the study the Upstream SUNA did not have a wiper, and it appears that periods of missing data were related to high sediment loads in Bells Creek. Also, on one occasion, the power cable to the SUNA was damaged by animals. To estimate the missing NO₃-N load values, scikit-learn’s implementation of gradient boosting regression was used (Pedregosa et al., 2011). Based on the results of a manual feature selection process, six features were used as input to the model: Upstream discharge, Upstream water temperature, Outlet NO₃-N load, atmospheric pressure, wetland water temperature, and air temperature. To evaluate the performance of the regressor, a 10-fold cross validation approach was taken whereby 90% of the data were used for training and the other 10% for testing. To explain the output of the model, Shapley Additive exPlantations (SHAP) values (Lundberg and Lee, 2017) were determined, which indicate the impact each model feature had on the predicted Upstream NO₃-N load time series.

The NO₃-N removal efficiency of the wetland was assessed on a weekly basis. A paired two-sample t-test for means was run on weekly subsets of the data for the duration of the study period (52 weeks), allowing for 168 observations per t-test (i.e., 168 h per week). A statistical significance level of p < 0.05 was used.

Discharge results showed good agreement between the measured stream discharge values and the values obtained from the stage-discharge relationships (i.e., rating curves) (Figures 4A–C). The best-fit rating curve for the Upstream stage-discharge relationship was a linear function (R² = 0.86, n = 21), while the best-fit rating curve for the Outlet location was a power function (R² = 0.78, n = 24).

Stream discharge at the Upstream and Outlet surface water sites was seasonally variable (Figures 4A, C; Table 1), with the highest discharge observed in the late winter-early spring and the lowest discharge observed in the early summer months. The computed groundwater discharge to the wetland also exhibited seasonal variability (Figure 4B; Table 1), with the highest discharge coinciding with the two primary groundwater recharge periods (late fall and early spring). The average discharge from the wetland (0.160 m³/s) was consistently higher than the average Upstream (surface water) inflow (0.055 m³/s) due to the direct groundwater discharge to the wetland (0.110 m³/s average). The cumulative direct overland flow simulated from the four subcatchments adjacent to the wetland (Figure 3) was only 1.04% of the cumulative Upstream flow, therefore, for subsequent computation of NO₃-N loads the load due to direct runoff was assumed negligible. Stream discharge at the Upstream and Outlet locations shows good correspondence (e.g., similar response to short-term precipitation events, and long-term seasonal trends) with the Winter River hydrometric station data (Supplementary Figure S4).

The results of the major ion analyses exhibited little variability, with all surface water and groundwater samples being calcium-bicarbonate-type waters (Figure 5). The similarity of the major ion chemistry results indicates that groundwater has a strong influence on surface water quality at the study site. The pH of the surface water ranged from 7.1–9.4, and the average pH of the groundwater was 7.1. The dissolved oxygen (DO) concentration of surface water ranged from 7.2–17.0 mg/L, while the average DO of the groundwater was 9.5 mg/L (see also Supplementary Table S2).

To assess whether the Mill Pond wetland was functioning as a nitrate sink or source, the weekly NO₃-N loads imported and exported to the wetland during the 52-week study period were compared. As shown in Figure 7, the imported NO₃-N load exceeded the exported NO₃-N load during 50 of 52 weeks. The t-test results indicated that 51 out of 52 weeks had statistically significant differences (p < 0.05) between the imported and exported loads. The wetland had the greatest imported and exported loads in late winter and early spring; however, it also had the greatest removal rate during this time, with a maximum weekly removal rate of 82% in February-March 2024. The lowest imported and exported loads were observed in July 2024. During the two weeks when the wetland behaved as a nitrate source, the maximum increase in NO₃-N load was 8%.

Unfortunately, missing data within long-term, high-frequency monitoring studies is a common hindrance to compiling a continuous data set (e.g., Pavlovskii et al., 2023). As such, using machine learning models to fill data gaps within datasets is an emerging sub-field of water quality studies. Sierra-Porta (2024) recently evaluated the suitability of ten machine learning models for filling water quality data gaps, finding that the gradient boosting model was among the most effective and accurate in predicting missing values. In the current study the gradient boosting model was also found to be effective at predicting the NO₃-N loads at the Upstream monitoring location, as shown by the results of the model cross-validation (see Section 4.3.2). Given the increased use of in situ nitrate monitoring instruments, such as SUNAs, and the continual challenges related to long-term field deployment, it is likely that the gradient boosting model, or other machine learning models, will become commonplace in water quality studies.

Throughout the study period, groundwater contributed 66% of the total freshwater discharge to the wetland; however, as seen in Table 1, the groundwater contribution varied seasonally. The maximum relative contribution occurred in the fall (72%), while the minimum occurred in the winter (48%).

In PEI, groundwater recharge typically occurs in late October due to reduced evapotranspiration, and again in March to early May due to snowmelt/rainfall (Jiang et al., 2015; Danielescu et al., 2022). These recharge periods correspond to peak groundwater baseflow periods and, consequently, peak groundwater nitrate contributions to the wetland (Figures 4B, H). Figure 8 shows the groundwater level measured in Well #6, upgradient of the wetland, and the computed groundwater discharge to the wetland. The occurrence of two principal recharge periods is reflected in both time series. The earlier increase observed in groundwater discharge at the Springs location may be explained by shorter flow paths adjacent to the wetland discharge area, whereas the water level in Well #6 shows a slower response due to the presence of a thick unsaturated zone (approximately ∼16 m) (Figure 8).

As seen in Figure 4G, in late November 2023, the NO₃-N load at the Upstream location reached a maximum; however, as seen in Figure 4D, the NO₃-N concentrations decreased due to a dilution effect caused by heavy rain events (Figure 4A), indicating that the increases in surface water NO₃-N loads were driven by discharge. Moreover, the groundwater nitrate concentrations remained relatively constant throughout the study period (Figure 4E), whereas the groundwater NO₃-N load (Figure 4H) trended with the groundwater discharge (Figure 4B). Hence, as mentioned previously, discharge variability is the dominant control on nitrate loads delivered to the wetland via surface and subsurface pathways. This is further confirmed by the SHAP values obtained from application of the machine learning model, which indicated that the Upstream discharge had the greatest impact on the NO₃-N load at that location (Figure 6).

As seen in Figure 7, the NO₃-N load to the wetland was high in the fall, then again in March and April, corresponding to the periods with the highest groundwater discharge rates (Figure 4B). These findings are consistent with a study by Miller et al. (2016), which found a positive relationship between base flow index (BFI) and in-stream NO₃-N loads. Therefore, estimating BFI over short periods (i.e., monthly) within PEI watersheds could aid in understanding NO₃-N load dynamics within groundwater-influenced systems. The seasonal trends and magnitude of loads in this study should be interpreted with caution due to the inter-annual variability in weather patterns (e.g., precipitation and temperature) that can significantly impact local and regional hydrological processes (Danielescu, 2023; Pavlovskii et al., 2023). For example, Oliver et al. (2024), using a PEI-based hydrologic model and future climate projections, showed that the fall peak in nitrate loads may shift from late fall to early December, while the spring peak may occur earlier in the season. Furthermore, Oliver et al. (2024) found that baseflow may increase during the winter and early spring as a result of increased precipitation as rainfall, more mid-winter thaws, and an earlier spring melt period, and a projected five-fold increase in February by the end of the century (2070–2100). In the current study the groundwater contribution to the Mill Pond wetland bypasses the riparian zone and enters the wetland via underwater springs, which allows for nitrate-rich groundwater to directly enter the receiving water body. As such, a potential five-fold increase in bypass flows during colder months (January and February), when the removal of NO₃-N within the wetland is at a minimum (Figure 7), could significantly increase the impact of nitrate in these freshwater systems.

Danielescu et al. (2024) simulated a 3-year potato-barley-clover crop rotation for a field at the Harrington Experimental Farm and found that ∼90 kg NO₃-N/ha/yr leached below the root zone in the potato year, while the barley and red clover (catch crops) years leached ∼29 kg NO₃-N/ha/yr each. Interestingly, the load data from this study results in a catchment-based loading rate of 54.1 kg NO₃-N/ha catchment/yr (i.e., using the catchment area upgradient of the study site, Figure 1), which falls within the simulated leaching loads for potatoes and catch crops. In 2023 only 5.5% of the upgradient catchment was cropland in potato phase (Agriculture and Agri-Food Canada, 2024), which suggests that the NO₃-N loads we have determined are related to agricultural land use in previous years (i.e., legacy NO₃-N). For the study period the total NO₃-N load to the wetland was approximately 30,000 kg NO₃-N/yr. Steiness et al. (2021) estimated that a 1.3 ha wetland within an agricultural landscape in Denmark received an average annual NO₃-N load of 1,033 kg NO₃-N/yr (41 kg NO₃-N/ha catchment/yr) between 1990 and 2008, noting that these loads were, at the time, on the high end of reported values. A recent study by Oliver et al. (2024) in the Basin Head Lagoon, located in northeastern PEI, found that the annual NO₃-N load to the lagoon between 1990–2020 was 7 kg NO₃-N/ha catchment/yr, where groundwater contributed 83% of the load. The annual NO₃-N load we have determined for the Mill Pond wetland (54.1 kg NO₃-N/ha catchment/yr) is thus relatively high in comparison.

The small (1.2 ha) groundwater-influenced wetland investigated here attenuated approximately 39% of the total annual NO₃-N load (30,000 kg NO₃-N/yr); on catchment area basis this is equivalent to 21.1 kg NO₃-N/ha catchment/yr. The wetland behaved as a NO₃-N sink for approximately 94% of the study period, or 49 out of 52 weeks, a NO₃-N source for two weeks during January and February, and neither a NO₃-N source nor sink for one week in January (Figure 7). As seen in Figure 9, there is notable seasonal variability in the NO₃-N attenuation capacity of the wetland, with the winter (Dec-Feb) having removal rates in the range of 8 to −41% (where the data point in late February-early March with a removal rate of −82% is excluded), the spring (Mar-May) having removal rates in the range of −15% to −71%, the summer (Jun-Aug) having removal rates in the range of −15% to −55%, and the fall (Sep-Nov) having removal rates in the range of −17% to −65%, where negative values indicate a net reduction in load. Although previous studies did not report nitrate removal rates on a seasonal basis, Steiness et al. (2021) found that a 1.3 ha wetland attenuated between 43% and 83% of the annual NO₃-N load, while Cheng et al. (2020) found that strategically placed wetlands could reduce NO₃-N loads by 54%. The ranges of removal rates observed in the Mill Pond wetland are similar to previously reported values; however, they also highlight the significant seasonal variability in NO₃-N dynamics which may be missed with low-frequency monitoring.

As shown in a previous study in a constructed wetland (Kuschk et al., 2003), seasonal variation in nitrate attenuation within wetlands in temperate climates should be expected, given that lower water temperatures would reduce the amount of vegetative uptake and biological activity. However, as seen in Figure 9, the NO₃-N load reductions for all four seasons generally overlap in the range of −15% to −55%, indicating that the wetland investigated here may attenuate nitrate regardless of season. Figure 9 reveals a positive, non-linear trend between the imported NO₃-N load and the change in NO₃-N load and, by extension, the impact of increased discharge to the wetland. These results suggest that the wetland attenuates disproportionately more nitrate during high imported loads, especially during the spring and fall periods. Veraart et al. (2011) found that temperature disproportionately increases denitrification rates within constructed wetlands, which could be a potential factor producing the trend observed in the Mill Pond wetland.

As discussed in Mitsch and Wilson (1996), constructed wetlands could take 15–20 years to reach their full biologically viability and sustainability. The Mill Pond wetland was constructed in the 1950s and is considered well established. However, wetlands continuously evolve over time, notably due to sediment accumulation driven by factors such as geomorphology and hydrology (Anderson and Mitsch, 2006). The findings reported here provide a one-year snapshot of the wetland functioning, but it is possible that the nitrate attenuation capacity may change over longer time periods due to changes in wetland vegetation, land use and climate.

Previous studies by Jiang and Somers (2009) and Malenica (2015) have found that denitrification is not significant in attenuating nitrate within the PEI sandstone aquifer. This finding, combined with groundwater flow paths that apparently bypass the riparian zone, indicates that nitrate in groundwater is likely not attenuated before entering the Mill Pond wetland. This is also supported by the significant DO concentrations (>8.0 mg/L) determined at all locations, including the groundwater springs within the wetland (see Supplementary Table S2). While it was beyond the scope of the current study to identify specific NO₃-N removal mechanisms, and there is limited evidence of denitrification (Crossley et al., 2024), it is possible that denitrification may be occurring in regions of the wetland bottom sediments not influenced by groundwater discharge, where low DO levels (i.e., <0.5 mg/L; Savard et al., 2010) may exist. In addition, the role of wetland vegetation, such as filamentous green algae, which proliferates during the summer months (e.g., Supplementary Figure S1) and is commonly observed in PEI freshwater wetlands (Cynthia Crane, personal communication, 17 April 2024), is not well understood. Additionally, it is understood that wetlands may remove other contaminants such as phosphorus and pesticides that were not investigated in this study (e.g., Kennedy and Mayer, 2002; Nowell, 2000). In future work, a multi-year study investigating the attenuation of various agricultural contaminants within a constructed wetland may provide greater insight into the removal potential of small wetlands as they evolve.