Vadose zone modeling to identify controls on groundwater recharge in an unconfined granular aquifer in a cold and humid environment with different meteorological data sources

Silty clay Quaternary deposits inherited from the Champlain Sea are dominant in the lowland portion of the Vaudreuil-Soulanges region, which is mostly occupied by agriculture. This silty clay can reach a thickness of up to 30 m and significantly limit GR for the whole region. Using a spatially distributed surface water budget, Larocque et al. (2015) estimated that GR in the Vaudreuil-Soulanges region can range from 0 mm/y in the clay-covered lowlands to 440 mm/y in areas where substantial sand deposits are present. At the locations of the Saint-Telesphore esker and Saint-Lazare and Hudson hills (Fig. 1), well-drained and thick fluvioglacial deposits represent highly permeable granular aquifers and preferential recharge zones (Larocque et al. 2015). It is estimated that the Hudson and Saint-Lazare hills receive 41% of all regional recharge. Other fluvioglacial deposits of a smaller extent, including the unconfined part of the Saint-Telesphore esker and a recharge zone on Mount Rigaud, receive a mean annual recharge of 256 mm/y. It is known that aquifer recharge occurs in episodic events, mostly in the fall (October to December) and during the spring snowmelt (March to mid-April), but the exact distribution throughout the year has not been quantified yet.

Three sites were monitored in the current study: the Saint-Telesphore (STEL) site and two sites located on the Saint-Lazare hill (SLZA and SLZB) are located on sandy fluvioglacial deposits reaching a thickness of 40 m (STEL) and 86 m (SLZA and SLZB) above the regional bedrock aquifer (Fig. 1). The deposits are composed of thick and complex fluvioglacial deposits that were reworked at the surface and are now characterized as deltaic and littoral glaciomarine deposits. Depending on the energy regime and the progression of the ice margin, the fluvioglacial deposits on SLZ hill are either silty beds or have blocks within the top 1 m. These thick fluvioglacial deposits lay on discontinuous and variable beds of till and ancient Quaternary sand before reaching the sandstone bedrock. This highly variable spatial distribution of sediments leads to alternating confined and unconfined conditions for the deeper aquifer on SLZ hill. The aquifer at STEL site is locally unconfined.

The two stations at the Saint-Lazare site were used to evaluate the effect of vegetation on soil water content and GR. SLZA is located in a small clearing area within the forest and covered with a sparse grass, while SLZB, located 15 m away, is located in a jack pine forest. The STEL is located close to a sand quarry and is dominated by dense prairie grass. The instruments were located 10 m from a forest. At the three locations, well-drained sandy soils were dominant at the surface. The thickness of the unsaturated zone varied between 4.8 and 7.4 m, and the average water table depth was 5.5 and 6.8 m below the surface at SLZ (site A and B; one piezometer between the two sites) and STEL, respectively.

The 1981–2010 climate normal means for the three weather stations in the region (MELCC 2019a) report a mean annual air temperature of 6.1 °C with a maximum mean of 25.7 °C in July and a minimum mean of −15.2 °C in January. The mean total annual precipitation is 980 mm, of which 16% falls as snow between November and March. During the study period, the total annual precipitation was 984 mm/y (2016), 1206 mm/y (2017), and 829 mm/y (2018). The maximum daily mean temperature was 26 °C (2016 and 2017) and 29 °C (2018), while the minimum daily mean temperature was −24 °C for the three years. The year 2018 is considered to be a dry year, while 2017 is characterized as a wet year based on climate normals. The latter was coupled with an early snow melt (mid-February), which led to major flooding events in the southern regions of Quebec (NASA 2017).

The model input data are the volume of liquid water available for infiltration and the potential evapotranspiration at the chosen time step (here both calculated from a meteorological dataset; see section ‘Model setup and parameterization’) in addition to soil moisture as volumetric water content (VWC) for calibration. Soil moisture was measured in situ starting in August 2015 (SLZA), December 2015 (STEL) and October 2016 (SLZB) (Table 1). Soil volumetric water content was measured with capacitance sensors (model EC-5, Campbell Scientific Inc.) located at 10, 20, 50 and 100 cm depths at sites SLZA and SLZB and at 25, 50, 75 and 100 cm depths at site STEL. VWC was recorded every 15 min and then averaged to daily values. The VWC sensors were not calibrated and the generic calibration accuracy of 0.03 cm³/cm³ was used.

Piezometers instrumented with level loggers (Solinst) were installed at the two sites. The borehole piezometer at SLZ was drilled in September 2015 and was 8.5 m deep. The sediments were characterized as medium to fine sand uniformly distributed through the soil column. At STEL, the piezometer borehole was drilled in August 2013 and the sediments were fine to medium sand from 0 to 4.1 m, silty clay between 4.1 and 6.5 m, and again fine to medium sand from 6.5 to 10 m. Below 10 m, the sediments varied between coarse and fine medium sand down to 21.6 m.

Soil cores were sampled at different depths at each site. Soil samples were manually collected using a shovel in spring 2019 at STEL from the 10, 20, 40, 60 and 90 cm depths, and at SLZ in summer 2018 at the 20, 40, 55, 90 and 115 cm depths for grain size analysis by laser sieving. The physical description of the textural group to which the samples belong is based on Folk (1954). The gravel fraction was under 1% for the most part, with a maximum of 4% at STEL from the soil sample at 60 cm depth. The organic matter content was estimated by ignition loss (in an oven at 500 °C) (Dean 1974; Heiri et al. 2001). Hydraulic conductivity was measured at the 20 and 50 cm depths at SLZA and SLZB, and at the 30, 60, and 90 cm depths at the STEL site using a Guelph permeameter (Soil Moisture Equipment Corp 2013).

To evaluate the reliability of simulated GR, it was compared with the results obtained from the water-table fluctuation (WTF) method (Healy and Cook 2002). The WTF method requires knowledge of the aquifer specific yield (S_y) and is based on the assumption that rising groundwater level in an unconfined aquifer is caused by percolating water reaching the water table and recharging the aquifer. The S_y parameter is treated as a storage term that accounts for the instantaneous change in water storage upon a change in total head (Healy 2010). This assumes that the amount of available water in a soil column of unit surface area is S_y times the height of water in the column. Because this method does not rely on meteorological data as an input, it is especially useful for a comparison with the simulated GR.

Meteorological data were available from three different sources (Table 2). First, the automated weather station located at the STEL site records local air temperature, precipitation, solar radiation and wind speed. Secondly, observed daily minimum and maximum temperature (T_min and T_max) and precipitation data interpolated on a 10 km grid by Natural Resources Canada (NRCan) based on the Australian National University Spline (ANUSPLIN) interpolation method (Hutchinson et al. 2009; Hopkinson et al. 2011; McKenney et al. 2011) are also available until 2017. Finally, daily temperature, precipitation, radiation, albedo and wind speed data are available from the North American Regional Reanalysis (NARR) atmospheric and land surface hydrology dataset, which uses the very high-resolution NCEP ETA model together with the Regional Data Assimilation System (Mesinger et al. 2006).

The added value of using NARR and ANUSPLIN data was investigated through hydrological modeling at the STEL station. Two different runs with varying configurations of the ANUSPLIN and NARR datasets as input weather data were performed for the 2016–2018 period. The first run, referred to as the ANUSPLIN scenario, employs T_min, T_max and precipitation data from the ANUSPLIN database. The missing radiation data were estimated from temperature or were set as constant (wind speed was set to 2 m s⁻¹ and albedo set to 0.23 for grass, according to Allen et al. (1998)). The second run, denoted as combined scenario, incorporates data from both datasets (i.e., T_min, T_max, and precipitation from ANUSPLIN and all other variables from NARR). The latter represents a more complete dataset that maximises the number of available variables for GR modeling. Because the ANUSPLIN data have not been produced since 2017, temperature and precipitation data were from the STEL weather station for 2018 because those two variables are very strongly correlated in the data sources. A significant bias was observed for NARR precipitation data, producing a substantial underestimation of the total precipitation (similar results were observed with data from a nearby grid point). Therefore, NARR data were only used in the combined scenario to assess the effect of using albedo, radiation and wind speed from this dataset instead of using constants or estimating these terms.

Because of temporal discontinuities in the STEL dataset, the weather station data were only used for comparative purposes to evaluate the accuracy of the NARR and ANSUPLIN datasets. The comparison was made against measurements from a weather station that includes the main meteorological variables (temperature, precipitation, radiation and wind speed) needed to calculate potential evapotranspiration (PET). Because of the high spatial variability of precipitation, the closest grid point in the ANUSPLIN dataset to both SLZ and STEL sites was selected and used in the calculation to better represent the meteorological conditions at each location.

The physically based model HYDRUS-1D (Simunek et al. 2015) for variably saturated media was used. Using Richards’ equation, the model simulates soil moisture dynamics within a soil column representing the unsaturated zone. The drainage through the root zone leaving the base of the soil column is considered to be GR. The van Genuchten-Mualem model (VGM; Mualem 1976; van Genuchten 1980) as chosen to represent the water retention and hydraulic conductivity characteristics of the soil samples with continuous mathematical functions:

where θ [L³/L³] is the volumetric moisture content; h [L] is the pressure head; θ_r and θ_s are the residual and saturated moisture content respectively; K [L/T] and K_s[L/T] are unsaturated and saturated hydraulic conductivity, respectively; and S_e = (θ - θ_r)/(θ_s - θ_r) [−] is the saturation degree. The fitting factor α [1/L] is inversely related to the pressure at the inflexion point of the retention curve, while n [−] is related to the pore size distribution of the soil with m = 1–1/n, and l [−] is a parameter accounting for pore tortuosity and connectivity.

The initial VGM parameters (θ_r, θ_s, K_s, α and n) were estimated using the ROSETTA software (Schaap et al. 2001). ROSETTA uses pedotransfer functions to predict the VGM parameters using the soil textural distributions obtained from laboratory analyses for the three study sites. K_s values from field tests and the predicted ones from ROSETTA were in the same range, except for the layer comprised between 16 to 30 cm depth at SLZA (Table 3). The hydraulic conductivities of the two layers at STEL, comprised between 41 and 300 cm, were higher than those predicted from ROSETTA, but they exhibited a similar increase with depth. It was hypothesized that field measurements were more representative of field conditions and local heterogeneities of the unsaturated zone than the ROSETTA predictions. For this reason, K_s values for the HYDRUS model were set to field-measurements at the available depth. The pore-connectivity parameter l was set to an initial value of 0.5, corresponding to an average value for many soils (Mualem 1976). Free drainage was set at the base of the soil column as a lower boundary condition. The length of the modeled soil columns was set to 3 m, with a total of 301 nodes evenly distributed between the surface and bottom.

The top boundary condition corresponds to daily vertical inflow (VI) and potential evapotranspiration (PET) used as daily values to drive the unsaturated zone model. VI is calculated from daily snowmelt values added to the liquid fraction of daily precipitation. The selected surface boundary condition allows surface runoff to occur in the model when the surface layer becomes saturated. This excess water leaves the system as runoff and is not available for infiltration. Due to the highly permeable soils at the three stations, saturation was never reached in this study and thus no runoff was simulated, but this could not be verified with field data. The simulated GR is thus considered to be an upper limit for GR, considering that runoff could occur in situ during high intensity precipitation events. Because precipitation data is available as total water equivalent, the calculation of VI as input to the snow model included separation between solid and liquid precipitation, as suggested by Turcotte et al. (2007):where T_min and T_max are the daily maximum and minimum temperatures (°C) and SnowFrac is the snow fraction for daily precipitation events.

A degree-day model was used to assess daily snowmelt available for infiltration:where Melt is the daily snowmelt (mm/day), C_melt represents the snowmelt rate (mm/°C/day), T_air is the mean daily air temperature (°C), and T_melt is the temperature at which the snow starts to melt (was set to 0 °C).

The snowpack density and depth were retrieved from a nearby governmental weather station (MELCC 2019a) to calibrate the snowmelt coefficient from the degree-day model and simulate the evolution of the snowpack during the winter seasons from 2000 to 2017. The daily calculated snowmelt values were added to the liquid fraction of daily precipitation to generate vertical inflow values (VI).

Daily PET was calculated using the Penman-Monteith equation (Allen et al. 1998). Beer’s law (Ritchie 1972) was used to partition PET into potential evaporation (E_p) and transpiration (T_p) directly in the model:where k is an extinction coefficient and LAI is the leaf area index (L²/L²). LAI data were obtained from the MODIS_MCD15A3H dataset, with a spatial resolution of 500 m × 500 m at 4-day intervals (Myneni et al. 2015). Daily LAI data for each site were obtained by linear interpolation between these intervals, coupled with a 30-day window moving average. LAI was used as a primary control of PET among different ecosystems in the same ecozone, such as the forest and pasture (Zha et al. 2010). Therefore, LAI data at the SLZ site were extracted from two points close to the site but each more representative of their respective ecosystem (forest and grassland).

The root water uptake was computed using the Feddes et al. (1978) model:where α(h) is a dimensionless function varying between 0 and 1, depending on soil matric potential, which corresponds to the force with which water is held within the soil matrix, and S_p [1/T] is the potential root water uptake and assumed to be equal to T_p.

The distribution of S_p over the root zone depends on root density distributions (between 0 and 1) attributed to each site and was selected based on literature descriptions of the vegetation’s physiological characteristics (Rudolph 1985; Wang et al. 2009b). In the absence of in situ measurements of root density, a rectangular distribution (homogeneous density distribution with depth) or triangular (linear density decreases with depth) profile are generally recommended. For example, 3-m-deep sandy soil cores analyzed for root biomass by Wang et al. (2009b) showed that in sandy soil cores from a grassland environment, 60–70% of the total root biomass occurred in the top 20 cm depth. At the STEL and SLZA sites, the root density was assumed to be linearly distributed between 1 at the surface to 0 at a depth of 30 cm. Jack pine trees generally develop a lateral root system, and the bulk of the root system is largely confined to the upper 45 cm of the soil, and mostly in the top 15 cm (Rudolph 1985). Therefore, at the SLZB site, the root density was set equal to 1 from 0 to 30 cm depth and a density ranging from 1 to 0 was used between 30 and 45 cm. The root water uptake was then assumed to be equal to actual transpiration. The actual evapotranspiration (AET) was calculated based on the water availability from PET and was the sum of the actual soil evaporation and actual transpiration rates. The soil matric potential values for delineating root water uptake were taken from the database integrated into the HYDRUS-1D model and assigned as alfalfa. Precipitation interception (by plant canopies before reaching the soil) was considered negligeable.

The model was used to simulate daily VWC and those were compared to measured values at the four depths where soil moisture sensors are located. The models were calibrated between 2015 and 2018 (depending on the monitoring period of each site), using the year 2015 as a warm-up period. Because of the limitations of the various meteorological observations, multiple data sources have been combined to calibrate the unsaturated zone model (Shen et al. 2010; Maggioni et al. 2014).

The soil hydraulic parameters (θ_r, θ_s, α, n, l and K_s) were calibrated using the default Marquardt-Levenberg type parameter optimization algorithm (Marquardt 1963) implemented in the HYDRUS-1D model to reproduce measured daily VWC at different soil depths at the three sites. The hydraulic parameters of each soil material were optimized successively using the field monitoring data, here starting with the top layer downwards until there was no further improvement in the objective function (Hopmans et al. 2002). The calibration period excludes the winter months (January to March) because of uncertainties regarding the quality of the VWC measurements during the freezing period. To evaluate how the choice of different meteorological datasets can influence GR estimates, the calibrated model at the STEL site was run for the three meteorological scenarios described above. In many studies, θ_r is not calibrated (Thoma et al. 2014; Turkeltaub et al. 2015) because GR estimates are found to be insensitive to this parameter (Simunek et al. 1998; Scharnagl et al. 2011). In the current study, the VWC at the three sites varied mostly within the dry range (VWC closer to θ_r than θ_s) which made it difficult to calibrate θ_s at a daily time step (rapid drainage and saturation state might not be properly captured). This parameter was thus calibrated only for the first soil layer at STEL and SLZB, where higher VWC were measured (see Fig. 4 and Fig. S2 of the electronic supplementary material (ESM)), but θ_r was calibrated because simulated VWC remained close to this value. In the literature, the pore connectivity parameter l in the VGM model is often set to a value of 0.5 (Mualem 1976; Zhu et al. 2013; Turkeltaub et al. 2015) considered to be an average value for many soils (Simunek et al. 2015). Flow in the unsaturated zone was shown not to be very sensitive to this parameter in general (Wang et al. 2009a). However, because sensitivity was demonstrated for very low VWC similar to those encountered in the current study (Vrugt et al. 2001), this parameter was calibrated here as suggested by Scharnagl et al. (2011). The upper and lower calibration bounds for all parameters (Table S3 of the ESM) were estimated based on the observed VWC at each site or on literature values for coarse materials.

Because of temporal discontinuities in the STEL dataset and the short temporal coverage, the weather station data were only used for comparative purposes to evaluate the accuracy of the gridded meteorological datasets NARR and ANUSPLIN. Comparing ANUSPLIN and NARR data with those of the STEL weather station showed a strong correlation between the datasets for temperature, with R² = 0.99 and slope = 1 for ANUSPLIN and R² = 0.95 and slope = 0.98 for NARR (Fig. 2a and b). The precipitation data from ANUPSLIN had a very similar distribution to the STEL data in 2017 (Fig. 2c and d) and exhibited similar results for 2016 (data not shown). A significant bias was observed for NARR precipitation data, which significantly underestimated the total precipitation (similar results were observed with data from a nearby grid point). The reason for this remains unclear, but could be attributable to the fact that there has been no assimilation of precipitation with observed data over Canada since 2003 (Mesinger et al. 2006). From 2016 and 2018, the average difference between total annual precipitation measured at the STEL station compared to the NARR dataset was 332 mm/y. However, this difference is an underestimation because there were days without measurements at STEL in 2016 and 2017 and the comparison was made on equivalent datasets (days without measurement at STEL were also removed from NARR dataset). The agreement between net radiation from the NARR dataset and that measured at STEL was somewhat scattered but reasonably good (R² = 0.82 and slope = 1.01), with lower agreement for the calculation of net radiation from temperature values (R² = 0.77 and slope = 0.95) (Fig. 2f and e). Both net radiation results showed an overestimation, particularly during the summer months. The overestimation of radiation in the reanalysis dataset may be explained by the uncertainty linked to the small scale of the study area compared to the larger scale (32 km grid) of the NARR dataset (e.g., cloud cover fraction and forest cover limiting solar exposure).

An overview of the hydroclimatic conditions across the study area is given as a statistical summary of the calculated annual mean VI and PET for each site (Table 4). VI values are similar at the three sites, except in 2016, when they were 99 mm lower at STEL (935 mm) than at the SLZ station (1034 mm; different ANUSPLIN point grid) and considered closer to normal conditions than the other two years. Also in 2016, the PET for the SLZ and STEL sites were 680 and 684 mm, respectively. The year 2017 can be characterized as a wet year, with a significant amount of VI (1208–1205 mm for SLZ and STEL). The PET during 2017 was 677 and 681 mm for SLZ and STEL, respectively. The year 2018 can be considered to be a dry year, with 829–830 mm of VI and a PET of 701 mm for SLZ and STEL. PET was the highest in 2018 for all sites because it was the hottest year (maximum daily mean temperature 26 °C in 2016 and 2017, and 29 °C in 2018).

A textural analysis was performed on samples extracted from up to 90 cm depth at STEL and 120 cm at SLZ (Table 3). For modeling purposes, the results of the deepest layer were considered representative of the conditions below that layer down to the bottom of the soil column (300 cm). None of the samples contained a clay fraction. The STEL site showed the most drastic change in soil texture with depth. The first layer, from 0 to 40 cm, was rich in silt (13.2% compared to 7.4% for SLZ sites), with a K_s of 362 cm/day obtained with the Guelph permeameter. The grain size shifted to a silt fraction of 5.3% and a K_s of 1296 cm/day between 41 and 65 cm depths. This is consistent with field observations of a very dense and fine soil horizon between approximately 20 and 50 cm. The two prairie sites showed the least organic matter, with 2.7% at SLZA in the first 15 cm depth and 2.5% in the first 40 cm depth at STEL. In the forest, organic matter content decreased from 6.9% in the first 15 cm depth to 3% in the layer from 31 to 60 cm depth. At SLZ, the soil characteristics of sites A and B are considered to differ only in their percentage of organic matter content and root distribution, here based on their proximity (15 m) and field observations. Therefore, the same textural analysis results (but not the same percentage of organic matter) were used to simulate the two sites. An important difference can be seen in the measured K_s at the 20 cm depth for both sites, with a value of 1036 cm/day in the prairie and 240 cm/day in the forest. This difference illustrates the influence of organic matter content and vegetation on infiltration at shallow depths. It is consistent with field observations, where unconsolidated medium sand at SLZA and a rich organic matter horizon in the forest were observed, enhancing water holding capacity at shallow depths.

Because organic matter content appears to play a significant role in hydraulic properties but is not considered in ROSETTA’s pedotransfer functions, the organic matter content was integrated into the soil textural distribution as a silt fraction. This reflects the fact that moisture storage capacity is positively correlated with organic matter and silt (Bot and Benites 2005). Wang et al. (2009b) found that increasing soil organic matter induced lower hydraulic conductivity in the sandy soils of the Nebraska Sand Hills, suggesting that soil carbon should be considered when estimating the hydraulic conductivity. In the current study, the modification of the textural distribution to include organic matter contents resulted in initial parameters better representing the hydraulic properties of the upper soil layer.

The contrasted moisture dynamics between SLZA and SLZB observed at shallow depths cannot be caused by the soil textural distributions alone. Therefore, the high organic matter content observed in the forest was considered to be an important soil characteristic for representing the soil moisture behavior in the top layers of these permeable sediments. The organic matter content or silt fraction in the uppermost soil layers can explain in part the higher and more variable VWC measured over time within the first 30 cm at every site and the consistently lower and less variable VWC with increasing depth. These observations are the result of the enhanced soil water holding capacity because of soil characteristics (fine fraction and presence of organic matter at shallow depths), in addition to the tighter coupling between soil moisture and land surface processes at shallow depths that directly receive precipitation and are subject to more evapotranspiration (Martínez-Fernández and Ceballos 2003; Guber et al. 2008). At the study sites, the sand fraction increased with depth (Table 3), promoting rapid percolation and lower water holding capacity.

The relationship between moisture and soil characteristics was also demonstrated by the measured VWC dynamic at each site at shallow depths. The lower variability of VWC at SLZA (Fig. 3) is typical of coarse sediments characterized by low holding capacity and favoring rapid drainage. It contrasted with the VWC measured at SLZB, which was highly variable and covered a large range of moisture content (Fig. 4). These observed conditions supported the integration of the organic matter into the silt fraction to account for field differences not reflected in the soil textural distribution alone.

At the STEL site, the soil needed particularly dry conditions for many consecutive days to reach low water content and create drought conditions, as observed by the first sensor in 2016 and 2018 (Fig. 5a). Hence, a higher AET demand was necessary to evacuate the pore water in the denser and finer soil at STEL. These soil characteristics allowed for a higher water holding capacity, comparable to the SLZB site, but with less dynamic variations. Also, the higher evapotranspiration associated to the vegetation present at SLZB must have played an important role in this dynamic behavior of VWC. Therefore, the higher natural silt fraction in the first soil layer at STEL permitted higher water retention and slow drainage. The higher fraction of organic matter in the forest also enhanced the water holding capacity, similarly to STEL, but allowed the soil to drain more easily when mixed with mostly medium sand. Other studies have also shown that many factors influence soil hydraulic characteristics, including bulk density and pore size distribution (Schaap et al. 2001; Wang et al. 2009b). This points to the importance of adequately characterizing the soil properties to simulate VWC and GR with an unsaturated zone model.

The model was able to capture the major processes in the soil moisture dataset obtained from an important portion of the unsaturated zone, extending from the ground surface to well below the root zone. Factors that might contribute to the deviation between observed and simulated VWC may come from a mismatch between the recorded precipitation or evapotranspiration rates, from uncertainties in the observed soil moisture data, or from the unaccounted heterogeneity of the soil layers, such as the presence of macropores. For example, the simulated VWC from the calibrated model at SLZB showed the lowest calibration statistics, which could be caused by a greater presence of natural heterogeneity in the forest ground, which was more likely to create minor local discrepancies between the measurements and simulation. Simulation errors can also be attributed to the significant spatial variability of soil moisture and degree to which the sensors were in contact with the soil material, especially in coarse soil. Nonetheless, the calibration statistics obtained (RMSE between 0.014 and 0.032; R² between 0.71 and 0.98; ME of nearly 0) were within the general ranges of values found in the literature (Jiménez-Martínez et al. 2009; Assefa and Woodbury 2013; Min et al. 2015; Wang et al. 2016), indicating good model performance.

Soil characteristics were shown to influence soil moisture dynamics and, subsequently, recharge processes. The higher GR and lower AET at SLZA were attributed to the coarser nature of the sediments in the root zone and to the vegetation type, which were both associated with a low soil water-holding capacity. Also, preferential flow was not considered in the model and could had an influence on the GR as water can reach the water table faster. With preferential flow, precipitation arriving at the surface typically infiltrates almost instantaneously because of the low organic matter content and can percolate rapidly through the root zone to recharge the underlying aquifer. Water flow of this nature limits plant uptake. The lowest GR was computed for SLZB, which was expected because the site was under forest cover and should take up more water in the soil by transpiration than prairie environments because of the systematically higher AET. The highest GR was computed at SLZA site (703 mm/year) and was probably out of bound. The GR computed at STEL was significantly lower than at SLZA, which was surprising because they had similar land cover. Because of the finer soil texture at STEL in the first 40 cm of depth, water flow was slowed, and the water-holding capacity was increased, permitting higher AET rates compared with the other prairie site. This explained the similar GR estimated at STEL and SLZB, which exhibited similar soil characteristics in the first centimeter depth in the model (Table 3; the higher organic matter content at SLZB here was transformed in silt fraction in the model). Thus, the soil textural distribution in the root zone was an important factor controlling the flow processes in the unsaturated zone.

The analysis of water level data to quantify GR using the WTF method yielded higher GR rates than those simulated by HYDRUS at STEL in 2016 and 2017 (Fig. 6). The interpretation of the numerous small increases as recharge events at STEL in 2016–2017 (wettest years) might have resulted in uncertainty propagation in estimating S_y or recharge events, leading to exaggerated GR rate estimations for those years. Some artifacts can arise in the water level time series from logger inaccuracy and from the Lisse effect, which affects the water level in unconfined aquifers by the entrapment of air between the water table and a wetting front (Weeks 2002; Guo et al. 2008). The Lisse effect is normally prevalent in fine-textured soils with shallow water tables <1–1.3 m (Weeks 2002), which was not the case here. The WTF method is also prone to errors when lateral groundwater inflows from upgradient areas are different from groundwater downflows to downgradient areas, causing potential artificial variations of the water table level that are not directly due to local vertical recharge. The water table time series at SLZ showed a less dynamic behavior, probably leading to fewer misinterpreted recharge events during wet years. This induced a lower estimate GR for STEL with the WTF method. It should also be noted that the SLZA site only represented a clearing area in a large forested zone where PET was probably high. This could explain the similar GR estimated by the WTF and model in the forest because the groundwater level fluctuations were representative of regional hydrogeological conditions. It could also be attributed to an underestimation of the recession curve used in the WTF method because of steady infiltration process in the coarser soil at SLZ. Also, the results from WTF obtained for STEL gave a similar, although slightly lower GR estimate than the HYDRUS model at the other prairie site (SLZA) in 2016–2017. This might be an indication of an overestimation of the fine-textured soil control in simulated GR in top soil layer at STEL.

Despite the uncertainties associated with S_y and the interpretation of recharge events, the WTF method captured the whole interannual range of GR variability simulated by the model (Fig. 6). The year 2017 was the wettest, leading to the highest GR, and 2018 was the driest, with the lowest GR for each site. Consequently, this approach represents a useful tool for estimating the possible range of GR independently of meteorological data or when only few data are available. The range of WTF GR values for different S_y (0.15 to 0.38) framed relatively well the simulated GR but illustrated that GR could be overestimated both by the WTF and by the HYDRUS model. Results should thus be considered as possible maximum ranges of GR. It should also be noted that the baseline S_y was calculated with two optimized parameters from HYDRUS and for this reason WTF GR estimates are not completely independent from those from HYDRUS. The HYDRUS-1D model provides insights into processes taking place in the unsaturated zone at the site scale, while the WTF method provides information at the regional scale and does not rely on meteorological data. Therefore, GR estimates from the two methods should be used with caution to estimate local GR rates at other locations within the area. The main advantage of using HYDRUS is that it is based on flow equations in the unsaturated zone, a reservoir that is still poorly understood, yet crucial for recharge. Another advantage of this type of model is that is uses data from soil moisture sensors that are relatively easy to install and maintain at low cost in a wide network. As pointed out in many studies, it is important to combine different GR estimation methods to reduce the uncertainty on GR estimation (Scanlon et al. 2002; Rivard et al. 2014; von Freyberg et al. 2015).

Based on simulations from a spatially distributed surface water budget model (1989–2009), Larocque et al. (2015) found that the areas of unconfined granular aquifers in the region received a mean GR of 331 mm/y, reaching a maximum of 440 mm/y. This maximum was lower than the mean GR rate found by HYDRUS-1D. The spatially distributed surface water budget model partitioned the water budget into multiple components not accounted for in the current study (runoff and subsurface flow) and considered negligible at the local scale because of the highly permeable soils (between 84 and 100% sand) and flat topography. These characteristics promoted rapid infiltration and prevented runoff and subsurface flow during rain events at the local scale. However, this assumption might have overestimated GR values because most of the infiltrated water reached the bottom of the soil column. Larocque et al. (2015) estimated that the runoff and subsurface flow components accounted for 56% of VI, corresponding to a mean of 540 mm/y for the whole region. However, the model from Larocque et al. (2015) was spatially distributed over the entire Vaudreuil-Soulanges region where topography was not flat and highly permeable sediments were not present everywhere. In the same study, GR from March to May was found to represent 38% of the annual recharge, and in 44% of the annual GR during the fall (October to December). These water budget components generated estimates that fell within the range of GR and GR/VI ratios computed by HYDRUS-1D for the three stations. Larocque et al. (2015) showed for the study area that mean annual AET represented 39% of VI using a surface water budget, and 46% of VI with the conceptual MOHYSE model (Fortin and Turcotte 2007). These ratios were lower than the findings of the present study but still within the same range of magnitude.

January and February GR estimates from Larocque et al. (2015) represent 9 and 5 mm/month, respectively. For the same year, the present study simulated GR ranges from 22 to 30 mm/month in January and from 10 to 14 mm/month in February for the three sites. These simulated GR ranges appeared to be relatively important (especially for January). This could be due to the extremely local simulation of GR in HYDRUS-1D model, compared to the large-scale spatially distributed water budget model of Larocque et al. (2015) which did not considered explicitly soil characteristics.

Barbecot et al. (2018) estimated GR at STEL from 2010 to 2013 with a hydro-isotopic water budget and found a GR of 200–372 mm/y occurring exclusively during the snowmelt period, with an AET of 608–517 mm/y. At STEL, for the same period, a mean GR of 400 mm/y and AET of 542 mm/y were calculated with HYDRUS-1D. The results from the hydro-isotopic water budget and from the model used in the present study were within the same order of magnitude and in good agreement. Mattei et al. (2020a) estimated GR in the same region using water isotopic profiles. The estimated GR at SLZ for 2017 varied between 380 and 395 mm/y and from 158 to 214 in 2018. They used the NARR meteorological dataset, which, as shown here, significantly underestimated precipitation values (an average of 331 mm/y as calculated in this study from 2016 to 2018). This might explain the lower GR estimation in Mattei et al. (2020a).

As seen in Fig. 7, the interannual variability of monthly GR was largest during the spring and fall when PET was negligible and with snow cover conditions being very different from year to year. Fall evapotranspiration and the timing of the spring snowmelt controlled the timing of the preferential recharge periods, while precipitation mainly controlled the magnitude and interannual variability of GR rates. This seasonality of GR processes might be an important factor to consider in the southern areas of Quebec, where precipitation is expected to increase under climate change (Ouranos 2015). It is possible that increased precipitation will be counteracted by increased PET resulting from higher temperatures, especially if higher precipitation occurs during the summer period. Conversely, if more precipitation occurs in spring, when AET is equal to PET and groundwater levels are high, climate change could lead to more runoff and eventually more flooding events. Milder winters might also be an important issue to consider in a climate change context because this will influence the actual seasonality of GR (Allen et al. 2010) by producing episodic recharge events during the winter period and reducing spring recharge volume (Rivard et al. 2009).

Two different model runs using varying configurations of the ANUSPLIN and NARR datasets as input data were performed at the STEL site to assess the impact of their utilization on GR estimates. These runs were performed at the STEL site with the model calibrated using the combined scenario as a sensitivity analysis. The GR computed with the ANUSPLIN scenario (398, 535, and 292 mm/y) was systematically lower than that calculated with the combined scenario (446, 575, and 374 mm/y), even if the same VI was used as an input. This was because of the PET calculation in the ANUSPLIN scenario, which was computed using the temperature data as a proxy for radiation terms. Using the NARR dataset in the combined scenario enabled the calculation of net radiation without approximating it from temperature data. It has been shown that the radiation term has the most impact on PET calculation (Sentelhas et al. 2010). Even though radiation data from NARR were overestimated during the summer compared with the measurements made at the STEL station, they yielded lower PET values (closer to those calculated with STEL measurements), resulting in higher GR rates for the combined scenario. The NARR data significantly underestimated precipitation and were therefore not used as a single input meteorological data source for GR simulation. Similar results were obtained by Wong et al. (2017) who inter-compared several gridded precipitation products over 15 terrestrial ecozones in Canada for different seasons from 1979 to 2012. The overall reliability of NARR was found to be low, probably because the absence of precipitation assimilation over Canada since 2003 (Mesinger et al. 2006). They reported that ANUSPLIN performed well in capturing the key timings and in minimizing the error magnitude of the precipitation.

Langlois et al. (2009) ran three snow models with NARR data in the southern regions of Quebec during three winters between 2004 and 2008. All the models gave accurate results compared with the simulations driven by data from a ground-based station or measurements of snow water equivalent. They demonstrated that the regional reanalysis can be used in snow models to predict snow water equivalent in Quebec. Choi et al. (2009) also demonstrated the applicability of NARR temperature and precipitation data for modeling hydrological processes in the northern regions of Manitoba. Therefore, for areas without direct measurements of radiation or AET, the NARR dataset may provide valuable additional information for the calculation of PET, and the underestimation of precipitation might differ by location and period of study.