Recent advances in hydrological modelling have included more explicit process-based conceptualization of ecohydrological interactions (Fatichi et al., 2016) and the integration of tracer-based data (Birkel and Soulsby, 2015). The EcH2O model (Maneta and Silverman, 2013) was developed as an ecohydrological model coupling land-surface energy-balance models with a physically based hydrologic model. This explicitly includes the dynamics of vegetation growth and vertical and lateral ecohydrological exchanges.

Hydrology in cold regions can be greatly affected by the freeze–thaw cycles of soil water during the winter, resulting in reduced liquid water storage capacity during the spring melt and a restricted capability for infiltration due to the expansion of ice in pore spaces (Jansson, 1998). The depth of the soil frost can have a large influence on the timing of snowmelt runoff and can provide an estimation of the liquid water available within a soil layer (Carey and Woo, 2005). The Stefan equation is a simple energy-balance approach to estimate the progression of soil water freezing (Jumikis, 1977): 1Δzf=2kfTs-Tfλθ12, where Δz)f (m) is the change in depth of the frost and is a function of the thermal conductivity of the frozen soil layers between the frost depth and the soil surface (kf, W m-1 ∘C-1), the soil surface temperature below the snowpack (Ts, ∘C), the temperature of freezing (Tf, ∘C), the latent heat of freezing (λ, J m-3), and the liquid soil moisture (θ, m3 m-3). As with previous approaches (Jumikis, 1977; Carey and Woo, 2005), the progression of the soil frost is estimated by discretizing the total soil depth into smaller layers. Within EcH2O-iso, the sub-surface soil regime is discretized into three soil layers, layer 1 (near the surface), layer 2, and layer 3 (groundwater to bedrock), to resolve the water balance and estimate soil moisture. Here, the depths of layers 1, 2, and 3 were used as the layers since they intrinsically incorporate the soil moisture estimations without additional parameterization. The thermal conductivity of frost-affected layers is dependent on the moisture content of the soil: 2kf(i)=ksat-kdry⋅θ(i)ϕ(i)+kdry, where kf(i) (W m-1 ∘C-1) is the thermal conductivity of frozen soil in layer i, ksat (W m-1 ∘C-1) is the thermal conductivity of saturated soil, kdry (W m-1 ∘C-1) is the thermal conductivity of dry soil, θ(i) (m3 m-3) is the soil moisture in layer i, and ϕ(i) (m3 m-3) is the soil porosity in layer i. The saturated thermal conductivity was estimated from the proportions of soil comprised of ice, liquid water, air, organic material, and mineral soil (Carey and Woo, 2005): 3ksat=∏n=15k(j)f(j), where j is the thermal conductivity of each volume proportion, f is the fraction of total soil volume, and k is the thermal conductivity of volume j. Without proportions of soil organic and mineral material, the bulk soil thermal conductivity (kdry) is considered the weighted average of organic and mineral thermal conductivity (only four total volumes in Eq. ). Implementations of Eqs. ()–() for freezing layers below the ground surface are ideal for EcH2O as the model estimates the parameters (Ts and θ) or includes parameterization of physical properties (λ, kdry, ϕ, kwater, kair), and only requires the addition of the thermal conductivity of ice (2.1 W m-1 ∘C-1, Waite et al., 2006). Within EcH2O, the estimation of surface temperature (above the snowpack) is assumed to be isothermal with the snowpack and conduction through the snowpack is not considered. However, conduction through snowpack is important for the Stefan equation (Eq. ) as the surface temperature used is below the snowpack, which is generally thermally insulated by the snowpack. To address the conduction through the snowpack without snow depth or density, the estimated surface temperature above the snowpack (TEst) was damped with a single unitless parameter (D) such that Ts=TEst⋅D. To account for the reduction of the infiltration rate due to ice, models have previously adjusted the soil hydraulic conductivity (e.g. Jansson, 1998). Here, the reduction in hydraulic conductivity is estimated using an exponential function: 4Kwf=10fc⋅FKsat, where Kwf (m s-1) is the hydraulic conductivity of the soil influenced by ice, Ksat (m s-1) is the saturated hydraulic conductivity of ice-free soils, fc is a unitless ice-impedance parameter, and F is the fraction of frost depth to total soil depth. Equation () has two key assumptions: no ice lenses or frost heaving, and no soil volume expansion due to lower ice density (assumed 920 kg m-3 at ice temperature 0–5 ∘C).

As soil frost progresses through the layers, the proportion of liquid water is assumed to decrease at the same rate as the proportion of unfrozen soil. Similar to other approaches estimating the moisture content of frost-affected soils (Jansson, 1998), a minimum liquid soil moisture was retained in all frozen soils. This minimum was assumed to be the residual soil moisture (θr), the minimum moisture content required for evaporation and root uptake. Following the estimation of the soil water infiltration and redistribution of soil water within EcH2O, the change in soil moisture due to freezing in each layer is estimated: 5Δθ=θ(i)-θr⋅Δzfd(i)-dF(i), where Δθ (m3 m-3) is the change in liquid water and ice content, θ(i) (m3 m-3) is the initial liquid content in layer i, θr (m3 m-3) is the residual moisture content, d(i) (m) is the total depth of layer i, and dF(i) (m) is the depth of frost in layer i. Step-wise estimation of freeze and thaw for each layer is provided in more detail in Sect. S1 in the Supplement. The water age of the ice is estimated in a similar way to the liquid water ages of the soil layers (Kuppel et al., 2018b): 6Vrest+ΔtArest+Δt-VrestArest=qinAint+Δt-qoutArest+Δt, where t is time (s), Δt is the time step (s), Vres (m3) is the volume of ice in storage, qin (m3) is the volume of water from the change in soil moisture during freeze-up (using Δθ in Eq. 5), qout (m3) is the volume of water from the change in soil moisture during thaw (from Eq. ), and A (s) is the water age (subscripts “res” and “in” are the water ages in storage and inflow, respectively). Similar to the isotope and vegetation modules in EcH2O, the frost dynamics (i.e. frost depths and water ages) were implemented as an option within EcH2O.

Isotopic fractionation of snowmelt can have a significant influence on the composition of streams (Ala-aho et al., 2017). Previous successful applications of a simple approach equation to estimate the isotopic fractionation of snowmelt at multiple locations have shown that low-parameterized fractionation models can be used to spatially approximate snowmelt fractionation. One of the noted limitations of the simple snowmelt fractionation approach used in Ala-aho et al. (2017) is the dependence of the snowmelt fractionation on the past snowmelt volumes (total days of melt, dmelt) rather than current snowmelt rate (δ2Hmelt=δ2Hpack-Mfrac/dmelt); Mfrac is a fractionation parameter in Ala-aho et al. (2017). The approach was modified to include the snowmelt rate with one additional parameter using an exponential function: 7δ2Hmelt=δ2Hpack-S⋅exp⁡-S⋅1-SWE-MSWEmax⋅C, where δ2Hmelt (‰) is the isotopic composition of the snowmelt, δ2Hpack (‰) is the composition of the snowpack at the beginning of the time step, SWE (m) is the snow water equivalent at the current time, SWEmax (m) is the maximum snow water equivalent before melt, M (m) is the total depth of snowmelt in the current time step, S is a unitless slope parameter describing the shape of the exponential change in the snowmelt fractionation, and C (‰) is an amplification factor. Equation () serves as the mass balance for the snowpack isotopes throughout the winter and spring melt period. In comparison to Ala-aho et al. (2017), the exponential form works to temporally change the shape of the fractionation as a relationship with the amount of melt (i.e. replacing 1/dmelt). Higher values of S (10–20) result in larger early melt fractionation and limited late melt fractionation, while low values of S result in a lower but more consistent fractionation throughout the melt period. The amplification factor behaves as a simplification of atmospheric effects on the snowmelt fractionation. The isotopic composition of the snowpack is updated at the end of each time step.

Svartberget (C7, 0.49 km2) is a small subcatchment situated in the headwaters of the Krycklan catchment (64∘14′ N, 19∘46′ E) in northern Sweden. Svartberget is a well-studied site with long-term data collection including streamflow (1991–present), stream chemistry (2000–present), and hillslope transect measurements (soil moisture and water chemistry). Svartberget has two subcatchments, Västrabäcken (C2, 0.12 km2) and mire (C4, 0.18 km2) (Fig. ). The topographic relief of C7 is 71 m (235–306 m a.s.l.), with 57 m of relief in C2 (247–304 m a.s.l.) and only 26 m of relief in C4 (280–306 m a.s.l.) (Fig. ). The climate is subarctic (in the Köppen classification index), with annual precipitation of 614 mm, evapotranspiration (ET) of 303 mm, mean relative humidity of 82 %, and a 30-year mean annual temperature of 1.8 ∘C (Laudon et al., 2013). The relatively low topography results in no observable influence of elevation on precipitation (Karlsen et al., 2016b). The catchment experiences continuous snowpack development throughout the winter, accounting for approximately a third of the annual precipitation and lasting on average 167 d (Laudon and Ottosson Löfvenius, 2016). The large quantity of snowfall results in a dominant snowmelt-driven freshet period (Karlsen et al., 2016a). Till (10–15 m thick) covers the majority of the downstream catchment area (C7, 92 % downstream of C4) with intermittent shallow soils in the headwaters of C2 (Fig. a). The catchment is predominantly forest-covered (82 % total, 98 % downstream of C4), with Scots pine (Pinus sylvestris), Norway spruce (Picea abies), and birch (Betula spp.). The mire (Fig. b) is dominated by Sphagnum mosses.

All simulations were conducted on a daily time step between January 2005 and September 2016. The period from January 2005 to December 2009 was used as a spin-up period with measured hydrologic data, to stabilize δ2H, δ18O composition, and water ages in each of the model storage units. Initial analysis of the measured discharge from 2000 to 2016 revealed the highest and lowest annual discharge years were between 2010 and 2014. Consequently, calibration was carried out for the period between January 2010 and December 2014. The validation set used was the remaining period from January 2015 to September 2016. The C7 catchment was defined with a grid resolution of 25×25 m2 to balance adequate differentiation of multiple locations on the soil water transect while maintaining computational efficiency. The 25 m grid includes adjacent soil pixels for S12 and S22, with sites S04 and S28 within the same grids as S12 and S22, respectively. Within the biomass module, the vegetation dynamics for leaf growth and carbon allocation were held at steady state to minimize the parameterization and focus on the soil freeze–thaw cycles. As temperature effects and water stress are less sensitive for conifer trees, a relatively constant leaf area index and needle growth/decay rate were maintained (Liu et al., 2018). Evaporative soil water fractionation was activated using a similar parameterization to Kuppel et al. (2018b), as this has previously been identified as an influential summer process in the catchment (Ala-aho et al., 2017). Soil relative humidity was estimated using Lee and Pielke's (1992) approach, and values of kinematic diffusion were estimated as presented by Vogt (1976) (0.9877 and 0.9859 for H2/H1 and O18/O16 ratios, respectively). Parameterization of the model was conducted for each soil type (three soil types, Fig. a) and vegetation type (four types, Fig. b).

A sensitivity analysis established the most sensitive parameters to be used in calibration using the Morris sensitivity analysis (Soheir et al., 2014). Parameters were assessed using 10 trajectories using a radial step for evaluating the parameter space. The parameter sensitivity was evaluated using the mean absolute error. Results of this are shown in Sect. S3. Sensitive parameters were calibrated using Latin hypercube sampling (McKay et al., 1979) with 150 000 parameter sets and a Monte Carlo simulation approach to optimize the testing of the model parameter space.

The model output was constrained using measurements of stream discharge (three sites, Fig. ), stream δ2H (three sites, Fig. ), and soil moisture (two sites, Fig. a). The eight measurement datasets were combined into a multi-criterion calibration objective function using the mean absolute error (MAE) with the cumulative distribution functions (CDFs) of the model goodness-of-fit (GOF) (Ala-aho et al., 2017; Kuppel et al., 2018a, b). The MAE moderated over-calibration of peak flow events, typical for functions like the root mean square error, and Nash–Sutcliffe efficiency, as well as being consistent with previous studies in the region (Ala-aho et al., 2017). To focus on the dynamics of soil moisture, given the coarse model grid, measured and simulated values were standardized by their respective mean values prior to analysis. From the CDF method, the 30 “best” simulations were selected for evaluation and are presented using a 95 % spread of predictive uncertainty (Kuppel et al., 2018b). The parameters achieved through calibration are shown in Sect. S4. Model results were verified against the remaining years of discharge, soil moisture, and streamflow δ2H, as well as independent time series of soil isotopes (bulk and lysimeter), net radiation, sensible heat, latent heat, and frost depth (estimated from depth-dependent soil temperatures).

The evaluation of changes to water ages due to soil frost was conducted by comparing the ages within the catchment for simulations of the 30 “best” parameter sets with and without frost. These were conducted without frost by turning frost dynamics off within the model. Freeze–thaw effects on evaporation and transpiration ages were evaluated as the difference between frost and non-frost simulations. Positive values indicate older water with the frost, while negative values indicate older water with frost-free simulations. The age differences were only considered on days when both frost and non-frost simulations simulate a flux greater than 0 mm d-1.

Calibration captured dynamics of both high- and low-flow discharge periods through both the calibration period (2010–2014) and validation period (2015–2016), with a maximum mean streamflow MAE of 2×-3 m3 s-1 for C7 and a maximum mean stream δ2H MAE of 5.8 ‰ at C4 (Table ). Due to extreme high- and low-flow periods in the calibration period, it was unsurprising that the resulting MAE was higher than in the validation. The MAE of the soil moisture calibration was also reasonable, with an average MAE of 0.05 and 0.09 for sites S12 and S22, respectively. With the standardization of the soil moisture, the low MAE indicates that the dynamics in the model correspond well to those measured. The optimization of the GOF for three measures (discharge, stream δ2H, and soil moisture) at eight locations resulted in a compromise for all streams. Simulations yielded better (lower) MAE for discharge and isotopes of individual streams.

Temporal variability of δ2H in each of the streams was captured quite well throughout the calibration and validation periods (Fig. a–c). The largest offsets in modelled isotopic composition occurred during the winter low-flow conditions. The simulated stream isotopes tended to retain a slight “memory” effect from the contributions of more enriched inflow in late summer. This was likely due to the underestimation of discharge during winter (Fig. d–f) which slowed the flushing of the more enriched water. Overall though, discharge was adequately simulated for each site, notably during the spring melt and summer months. While flows were underestimated during the winter, the difference between simulations and measurements was typically <1×10-3 m3 s-1. The weight-median water ages of each of the three streams were broadly similar, 2.8, 2.6, and 3.1 years for C7, C2, and C4, respectively (Fig. g–i). These stream ages were generally older than previous estimates, with deeper soil layers and complete mixing in each compartment tending to increase the average age. The depths of the soil layers in the peat and podzolic areas are the primary drivers of water age, with a ∼1:1 relationship (Sect. S5). Water age decreased during the annual freshet, driven by the younger snowmelt and frozen soil water ages (typically 150–200 d old). The rapid runoff during the freshet limited the long-term influence of the younger water ages on the stream water at each of the sites as older groundwater dominated low flows. Rain-on-snow events resulted in some rapid, yet un-sustained, influences on the soil water ages, as observed with the sudden decreases in stream water age during winter months (Fig. g–i; log scale with a lower bound of 250 d).

Simulated soil water isotopes (note that the model did not use isotopes during calibration) mostly captured those measured in both bulk water (2015–2016) and lysimeter water (2012) within the 90 % simulation bounds at the S12 and S22 sites (Fig. a and b). Isotope dynamics were best captured at site S12, with early season depletion due to snowmelt and enrichment of the previous summer. While most variability was captured within the 90 % bounds, the magnitude of the intra-annual contrasts at site S22 was not fully reproduced. Similar to the soil isotopes, dynamics of simulated soil moisture (calibrated) were captured at both S12 and S22, with better simulation performance at S12 (Fig. c and d). The model struggled to simultaneously reproduce the more damped soil moisture at S12 with the relatively dynamic soil moisture post-melt at S22 in the adjacent cell under the same soil parameterization. Rather, the same parameterization resulted in balancing of the conditions observed at S12 and S22. The large declines in measured soil moisture during the winter months were captured with the soil frost module (Fig. c and d). The modelled decline in the soil moisture resulted from the transition of soil water from liquid to ice. Water ages in layers 1 and 2 at each site showed noticeable intra-annual variability, and gradually declined during the growing season (May–September) and increased during the winter due to negligible water inflow (Fig. e and f). The gradual decrease in the soil water age during the summer was the result of younger rainfall flushing the older snowmelt and pre-winter soil water as the growing season progressed. The variability of the soil water ages in layers 1 and 2 was similar, though the ages in layer 2 were significantly older. While S12 is closer to the stream, water ages in S22 were generally older in both layers 1 and 2.

Simulations of frost depth revealed large inter-annual variability throughout the catchment (Fig. a–d), depending on winter temperatures, snowpack depth, and the soil moisture conditions. Wetter conditions in the mire generally show shallower frost depths than the podzolic soils elsewhere in the catchment. Similar soil conditions for the podzolic and thin podzolic soils (Fig. a) resulted in negligible differences for estimated frost depth. Overall, estimated frost depth was generally limited by the total number of freezing days. Colder winters (larger numbers of freezing degree days) resulted in deeper frost depths for an equivalent snowpack depth (e.g. Fig. a vs. Fig. c). Conversely, a deeper snowpack (higher maximum SWE) resulted in a shallower simulated frost depth for years with similar temperatures (e.g. Fig. a vs. Fig. c), as the deeper snowpack was a larger storage for incoming radiation. Using 0 ∘C in the soil temperature probes at the ICOS tower as a proxy for the depth of the soil frost, a direct comparison of simulated frost depth and the measured catchment frost depth was completed without calibration. Simulated frost depth showed good agreement with observed 0 ∘C soil temperature depth, imitating the rapid increase in frost depth in 2014 and a more gradual increase in 2015 (Fig. e). Late winter soil frost depth was estimated to be shallower and varied more rapidly than the observed 0 ∘C soil temperature depth (Fig. e). The median estimated soil depth against the measured 0 ∘C soil temperature depth showed that estimated soil thaw was too rapid, and thaw completed too early.

While the evaporation, transpiration, and energy-balance datasets were not included in the calibration, modelled energy-balance components (sensible heat, latent heat, and net radiation) showed reasonable agreement with observed values in 2014–2016 at the ICOS tower. There was an underestimation of net radiation and sensible heat throughout the growing season (Fig. b and c) and an underestimation of latent heat late in the year (Fig. a). While the MAE of the latent heat was relatively small (13.1 W m-2) considering that they were not used for calibration, net radiation and sensible heat had a notable maximum bias (∼30 W m-2) during summer. Simulations of total daily evaporation (soil and interception) and transpiration had a similar pattern, with transpiration accounting, on average, for 54 % of total evapotranspiration. Throughout the year, the simulated proportion of transpiration to total evapotranspiration ranged from 31 % to 72 % except for the spring periods (Fig. d). The late onset of evaporation resulted from the assumption that soil evaporation was negligible while the snowpack remains, which potentially leads to an underestimation of evaporation during the melt.

Ages of soil evaporation and transpiration decreased throughout the year (Fig. a and b), tracing the decline in soil water ages estimated with the addition of precipitation (age of 0 d). Older water present in evaporation and transpiration water at the start of the year was a mixture of the snowmelt water age and frozen water ages (from the previous summer). Spatial differences in evaporation and transpiration ages were evident throughout the catchment, shown by the difference between the forested ICOS tower site (blue, Fig. a and b) and the average for shrubs in the mire (green, Fig. a and b). The annual flux-weighted median water age of transpiration was 200±42 and 141±40 d for the ICOS tower and mire, respectively, while evaporation ages were 48±11 and 85±36 d for the ICOS tower and mire, respectively.

Differences between the evaporation and transpiration ages were determined by comparing water ages with the soil frost module activated, against those with the frost module deactivated. Generally, including frost in the simulations resulted in older water (water age difference > 0 Fig. c) for both evaporation and transpiration. Differences in evaporation age were not as pronounced as transpiration ages due to the slight bias of the evaporation timing (always following the snowmelt). Due to the estimated completion of soil thaw prior to the snowmelt period, the difference between the water ages of evaporation with the influence of frozen ground was modest. Rapid flushing of the soil water due to large snowmelt inputs and spring precipitation resulted in a rapid decline in the differences of transpiration water ages. Within the first month of transpiration, the difference for the frost and non-frost simulations were more notable and approached 200 d when frost limited water movement. However, the relatively lower transpiration rates, which occurred during the spring within these simulations, resulted in a moderate effect on the overall annual transpiration water ages. The effects of soil frost on stream water ages showed little effect, with negligible differences given the relatively old water bias in the stream that only shows some flashes of younger water influence (Fig. g–i). While the soil frost increased the stream water ages throughout the year, the effect is well within the relatively large uncertainty bounds of the stream water ages.

Hydrologic models are powerful tools for exploring the internal functioning of catchments, particularly when intensive and long-term monitoring programmes are in place to help calibration and testing (Maxwell et al., 2019). Here, the development of a spatially distributed, process-based tracer-aided model for northern climates produced encouraging results, reproducing soil frost dynamics despite the model not being directly calibrated to match frost depth observations. The use of streamflows, stream isotope ratios, and soil moisture dynamics in calibration proved to be adequate for estimating the dynamics of soil frost depth and timing of the frost onset (Fig. ) and revealed spatial differences in frost depth due to contrasting soil types and moisture conditions. However, there are limitations with the current approach that result in some uncertainty of the effect of soil freeze–thaw on catchment hydrology. To improve the computational efficiency of the model, the temperature of the snowpack was assumed to be isothermal (Maneta and Silverman, 2013) and was modified here to include only a single temperature damping parameter. However, snowpacks may have a variable thermal gradient (e.g. Filippa et al., 2014), and are dependent on snow density (e.g. Riche and Schneebeli, 2013), snow surface albedo, wind speed, and liquid water component, among others (USACE, 1956; Meløysund et al., 2007; Sturm et al., 2010). While these additional components may contribute to an improvement in the estimation of soil frost, it likely would not have a significant improvement compared to the simple temperature damping used here with additional calibration to constrain snow water equivalent for more dynamic energy exchange (e.g. Lindström et al., 2002). The simplistic consideration of negligible soil sensible heat storage effects on the soil freeze–thaw processes, consistent with other process-based cold-region models (e.g. CHRM, Krogh and Pomeroy, 2018), may result in dampened rates of freezing and rapid melting during the spring (Kurylyk and Hayashi, 2016). More delayed melting of the soil frost may have implications for snowmelt runoff, increasing the dynamics of the streamflow isotopic compositions towards more depleted isotopic compositions (Fig. a–c). Finally, the simplification of a single soil frost front may have some implications for the snowmelt infiltration to the soil. The single front does not allow for near-surface soil thaw to occur prior to deeper soils and thereby has implications for shallow root-water uptake and evaporation.

Energy fluxes in northern catchments can be highly sensitive to the timing of snowmelt, yielding differences in the surface and canopy net radiation due to changing albedo, and to turbulent fluxes due to alterations in surface temperature. Here, the underestimated sensible heat flux during the spring and the growing season could be the result of many different processes. Some of these processes include the aerodynamic resistance (ra) to transpiration, an underestimated thermal gradient between the simulated soil and air temperature, an overestimation of the incoming short- or long-wave radiation from the ERA-Interim dataset, sublimation and snowpack energy storage, or an overestimation of a heat sink of ground heat through the thermal layers of the soil. An overestimated aerodynamic resistance can simultaneously reduce the transpiration (increase the surface temperature) as well as decrease the sensible heat flux (Maneta and Silverman, 2013). An overestimated surface temperature can result in both a decreased thermal gradient from the soil to the air used for the sensible heat flux estimation as well as in the shallower, and earlier, simulated soil frost melt relative to the measured 0 ∘C soil temperature depth. While the short- and long-wave radiations from the dataset had no consistent deviation from the shorter measured time series at the ICOS tower (1:1 ratio for both short- and long-wave radiation, not shown), intermittent periods of underestimated ERA-Interim radiation could contribute an underestimation of the net and sensible heat flux. Lastly, the lack of snowpack heat storage and sublimation could have an influence on the energy balance by limiting incoming winter radiation into the soils (i.e. decreasing soil temperatures).

The adaptation here of a process-based, spatially distributed model to additionally incorporate fundamental cold regions processes provides the opportunity both to improve the representation of key hydrologic functions of cold catchments and to assess the effect that these processes have on transit times and ages of ecohydrological fluxes. While studies in northern catchments have aimed to assess the partitioning and transit or residence times of ecohydrologic fluxes (e.g. Sprenger et al., 2018a), few studies have included the influence of frozen conditions on the water movement. Frozen conditions may be significant for the effective transit times during the spring freshet period (Tetzlaff et al., 2018) and flowpath modelling (Laudon et al., 2007; Sterte et al., 2018). Traditionally, isotopic tracer waters at catchment outlets have been the primary metrics for assessing the water ages in streams. Here, the relatively old age of stream water and the underestimation of soil thaw result in only slightly older water ages when soil frost conditions are considered, potentially due to the smaller proportion of wetland areas (Sterte et al., 2018). The limited effect of soil frost effects on stream water age was compounded from both the wide uncertainty of the stream water ages (Fig. g–i) and the late deviation of soil water frost ages from the soil water (Fig. e and f). Generally, the uncertainty bounds of the stream water ages were greater than the difference of the soil water and soil frost ages. The smaller different of soil water and soil frost ages thereby resulted in small effective changes in stream water age. The deeper frost depth in the forested regions likely did not reduce the spring infiltration due to the low moisture content in the soil relative to the more saturated wetlands (Laudon et al., 2007). Additionally, the relatively wide uncertainty bounds of stream water age estimates present difficulties in assessing the relatively moderate effects of soil frost on the stream water age (Fig. ). The large dependence of the flows and stream water ages at C7 on the outlet of the large mire at C4 indicates that the water age progressing through the mire will be a strong determinant of long-term change. The flux-weighted median water age estimations for the streams here were estimated to be substantially older than other tracer-aided hydrologic models for the catchment (Ala-aho et al., 2017), though they were on the upper end of other stream and hillslope transit times from transit time methods (Peralta-Tapia et al., 2016; Ameli et al., 2017). The reasons for this are largely 3-fold. Firstly, the model was calibrated with soil depths comparable to those observed in the catchment. The calibrated model used soil depths ranging from 1.5 to 6 m, where the shallower soil depths yield stream water ages comparable to previous studies. Secondly, the complete mixing assumption within the model does not allow for rapid preferential movement of young water that has been observed in numerous other recent studies (e.g. Botter et al., 2010; Harman, 2015). Incomplete mixing within the model framework would allow for deeper soil profiles to yield younger water fluxes, as estimated from isotopes alone, albeit at the expense of additional parameterization. Lastly, the previous transit time estimations (Peralta-Tapia et al., 2016; Ameli et al., 2017) do not account for older water ages of the snowpack, or the immobility and aging of frozen soil water, which would increase the estimated water ages.

Unlike stream or soil water ages, low uncertainty of transpiration and soil evaporation ages helps bring new understanding to how soil frost affects the source contributions of these ecohydrological fluxes which were the focus of the study. Ages of both transpiration and soil evaporation are consistent with soil profile modelling conducted in the region using the SWIS model (Sprenger et al., 2018b). However, the dynamics of the age variation are notably different due to the differences in the input water, where the snowmelt input to the SWIS model assumes a water age of 0 d and does not account for the “green” water fluxes during the spring months. While the transpiration ages show notable differences when frost and the corresponding discontinuity of transit times are included in the simulation, the evaporation water ages are not greatly affected. Transpiration fluxes are influenced by the frost due to the reduced liquid soil water availability and increased soil resistance to transpiration. The higher soil water deficit for transpiration does not fully restrict the transpiration flux, but reduces the fluxes until soil frost thaws and mixes the older frost water in the upper soil layers. The differences are reduced for both fluxes due to a few potential reasons. Firstly, the timing of the soil thaw has a significant influence on age estimation of soil water available for both evaporation and root uptake. While the general timing and magnitude of the soil frost depth development seems appropriately captured by the model, even without calibration, soil thaw in late winter was simulated faster than observations (Fig. e). There are notable differences between the ages of soil water, soil frost, and the snowpack, where soil frost is representative of the previous fall soil water, soil water is a younger water mix of the fall soil water and newer precipitation (e.g. from rain-on-snow and early spring snowmelt), and snowpack is the amount-weighted age of solid precipitation. Here, shallower soil frost and early melting of soil frost in the spring results in step-wise mixing, firstly of soil frost (oldest water, e.g. 200–300 d in Fig. ) and soil water (moderate age, e.g. 100–125 d in Fig. ), then of the soil water mixture and snowmelt (youngest water, e.g. 70–90 d). Since evaporation, and its corresponding age, only begins following the end of snowmelt, the greater degree of older soil frost with the younger soil water and snowmelt reduces the influence of the soil frost on the evaporation ages. Delaying the simulated timing of soil thaw would result in a larger influence of the soil water ages on both the evaporation and root uptake.

While the influence of soil frost on stream water ages was limited in this catchment, the results have potentially significant implications for modelling other catchments with frozen soils. The effect on water ages will likely be greatest in catchments where winter precipitation is limited, allowing the soil frost depth to increase from the surface, delaying the soil thaw until after the primary snowmelt. For evaporation and transpiration water ages, notable spatial differences highlight an essential consideration for northern climates in the influence of vegetation type on the source of water fluxes. In many northern areas, past glaciation results in significant wetlands typically dominated by shrub and herbaceous vegetation. Reductions in soil frost will result in greater water availability throughout the year, aiding in vegetation growth (Woo, 2013). With increased water availability throughout the year, the water use of vegetation will likely increase and thereby limit the amount of young water percolating through the rooting zone. A reduction in the amount of young water percolating through the rooting zone will likely increase the age of soil water and catchment outflows. Finally, the timing of the evaporation and root uptake needs to be strongly considered at both seasonal and diurnal timescales. Soil frost had a strong influence on the timing of evaporation and transpiration, where the magnitude of both fluxes was greater in simulations without soil frost and timing of the root uptake and soil evaporation was delayed due to ice-restricted pore spaces. While such changes are anticipated, many studies have focused on plot-scale studies, and with estimated long-term reductions of soil frost depth, larger-scale estimations of these differences are essential to understanding how catchment ecosystems will respond.