We conducted our study in the Krycklan catchment in northern Sweden (64∘14′ N, 19∘46′ E) along a 1.5 km stream reach located between the gauging stations C5 and C6 (Fig. 1) (Laudon et al., 2013). The gauging station C5 is the outlet of Lake Stortjärn (4.2 ha) and has a catchment area of 65 ha. The gauging station C6 is situated 1.5 km downstream of C5 and has a catchment area of 110 ha. The catchment contributing to the C5–C6 reach consists of pine-dominated forest, mostly underlain by post-glacial till soil (72 %). Iron podzols and thin soils can be found in the upland areas, while the shallow subsurface soils of the riparian zone (<1.2 m deep) are dominated by peat. Furthermore, soil wetness and flow accumulation maps based on 2×2 m digital elevation models identify seven wet corridors of 1–10 ha contributing area that extend from upland areas to the stream (Ågren et al., 2014). From those, five are considered DRIPs based on previous field validation using vegetation surveys and thermal and isotopic tracing (Kuglerová et al., 2014; Leach et al., 2017). The five DRIPs collectively account for approximately 60 % of the lateral groundwater inflows along the reach, while the remaining lateral inflow is diffuse (Leach et al., 2017). No tributaries are present along the stream reach, and deep groundwater inflows are minimal in this catchment (Tiwari et al., 2017).

The average annual temperature in Krycklan is 1.8 ∘C (period 1981–2010; Laudon and Ottosson-Löfvenius, 2016), and the average annual precipitation is 614 mm, of which 35 %–50 % falls as snow (Laudon et al., 2013). Approximately 50 % of the annual precipitation translates to streamflow. The hydrological regime at the C5–C6 reach is dominated by the annual snowmelt peak, occurring around May (100–200 L s-1). In summer and fall, low flows (<10 L s-1) alternate with medium to high flows (25–75 L s-1) in response to rain events. During winter, the stream is snow- and ice-covered, with flows < 3 L s-1. At C5, streamflow is mostly driven by lake-level variations. As a result, peak flow events are dampened and recession limbs decline gradually (Leach and Laudon, 2019; Fig. 2). At C6 (1.5 km downstream), streamflow responds much more quickly to hydrological events compared to C5 and is characterized by steep rising limbs (Fig. 2; Ploum et al., 2018).

Field measurements were collected between May 2017 and May 2019. In total, we conducted 14 sampling campaigns with different streamflow conditions, which ranged from drought to peak flow conditions (Fig. 2). Nine sampling campaigns were centered around the snowmelt periods of 2017–2019 and five around a lake damming experiment in summer 2017. In this experiment, the upstream lake was blocked and, after a period of artificial drought, a series of controlled flows were released using a pump. During the course of the artificial drought, the strength of DRIP-stream hydrological connections declined, generating a patchy distribution of lateral DOC inputs similar to those occurring under natural droughts (Gómez-Gener et al., 2020). For each sampling campaign, stream water was collected along the stream reach at approximately 50 m intervals over 1200 m, dividing the stream reach into 25 sections (Lupon et al., 2019). Five of the 25 sections had a DRIP discharging into it, while the other 20 sections were fed by small diffuse groundwater inputs (Figs. 1 and 3). For 10 of those sections, we sampled riparian groundwater inputs from a well network setup, which included five pairs of DRIP and non-DRIP wells located 1–5 m from the stream edge (Ploum et al., 2020). Therefore, we sampled the phreatic groundwater of all DRIPs but not all the diffuse groundwater inputs discharging into 15 reaches. The PVC wells (30 mm diameter) had a mean depth of 95 cm (σ=37 cm) below the soil surface and were fully screened every 5 cm.

Stream water was collected from the thalweg with acid-washed high-density polyethylene bottles. Groundwater was sampled from PVC wells using suction cup lysimeters and evacuated glass bottles or using a peristaltic pump to fill acid-washed high-density polyethylene bottles. The wells were pre-pumped to ensure we did not sample stagnant water. Bottles for both stream water and groundwater were rinsed three times before filling with minimal headspace. Within 24 h, all samples were filtered (0.45 µm mixed cellulose ester (MCE) syringe filters, Millipore^®) and kept refrigerated at 4 ∘C until analysis (<7 d after filtering). DOC analysis consisted of acidification of the sample for removing inorganic carbon, followed by combustion using a Shimadzu TOC-VCPH (analytical error: 2 %; Laudon et al., 2011). The analysis was repeated at least three times per sample, resulting in a DOC concentration (mg L-1) and a percent standard deviation.

We used a mixing model that considered the stream DOC concentration at location i to be a result of upstream DOC flux (location i-1) and the net lateral riparian groundwater flux that is gained along the stream section between locations i-1 and i. In addition, we considered that riparian DOC inputs were subjected to in-stream uptake (Eq. 1). 1DOCswi=DOCsw,i-1×Qi-1+DOCgwi×Qi-Qi-1-uptakeiQi, where DOCsw,i and DOCsw,i-1 are the stream DOC concentration measured at locations i and i-1, respectively, Qi and Qi-1 are the estimated streamflows at locations i and i-1, respectively, the difference (Qi–Qi-1) is the net groundwater inflow for that stream section, DOCgw,i is the estimated groundwater DOC concentration between locations i-1 and i, and uptakei is the in-stream DOC uptake associated with lateral groundwater labile DOC inputs (see below).

We modified the abovementioned model (Eq. 1) to represent different assumptions about catchment hydrology (diffuse vs. distributed groundwater inputs) and in-stream DOC uptake (no uptake vs. in-stream uptake downstream DRIPs), resulting in four different models (Table 1). The “Diff” model assumed diffuse groundwater inputs and no in-stream DOC uptake. The “Diff-Bio” model also assumed diffuse groundwater inputs but accounted for in-stream DOC uptake downstream from DRIPs. The “UCA” model assumed that groundwater inputs were distributed proportionally to their UCA and no in-stream DOC uptake. Finally, the “UCA-Bio” model assumed groundwater inputs proportional to UCA and accounted for in-stream DOC uptake downstream from DRIPs. Below, we outline the approaches used for estimating riparian groundwater DOC concentrations, groundwater inputs, and in-stream uptake.

For each model, we accounted for uncertainty in modeled stream DOC concentrations (Eq. 1) by incorporating errors in Q observations, water sample analysis, and estimates of both DOCgw and DOC Vf. We assumed normally distributed errors for Qc5 and Qc6 (±10 %, based on repeat streamflow gauging; Karlsen et al., 2016), DOCgw (either ±2 % for sites with measurements based on laboratory analytical precision or ±1 SD – standard deviation – of the mean for sections that rely on estimates), and Vf (±10 %, based on Mineau et al., 2016). For each date, each model was run 10 000 times using random values selected from these parameter distributions. Error estimates in DOCsw,i were tracked downstream since the values become DOCsw,i-1 in the computation for the next stream reach.

We evaluated the simulation of each run using two goodness-of-fit metrics, computed using the “hydroGOF” R package (Zambrano-Bigiarini, 2020). First, we computed the percent bias (PBias, %), which measures the average tendency of the simulated values to be larger (PBias > 0 %) or smaller (PBias < 0 %) than their observed ones. We considered that a model successfully simulated the magnitude of stream DOC concentrations if the median value of all the runs was within -5 % and +5 % bias. Second, we calculated the Spearman correlation (R), which shows whether the longitudinal patterns of simulated DOC concentrations mimicked the observed ones. In this case, we considered that a model was capturing the general direction of stream DOC concentrations if the median R of all runs was higher than 0.70.

Across the 14 sampling campaigns, hourly Q ranged from 0 to 116 L s-1 and from 2 to 152 L s-1 at C5 and C6, respectively, which were comparable to the range of flows observed throughout the whole ice-free period (C5: 0–150 L s-1, C6: 2–200 L s-1; Fig. 2). At both gauging stations, maximum Q occurred during the snowmelt, whereas minimum Q was observed during the artificial drought in summer 2017. As a result, the net gain in streamflow along the C5–C6 reach ranged from 8 % (artificial flood, event G) to 90 % (artificial drought, event H) (Fig. 2). During the other sampling campaigns, the net gain in streamflow along the reach was between 20 % and 50 %, with a mean of 37 % (Fig. 2).

During the study period, stream DOC concentration ranged from 15 to 32 mg L-1 and varied over time as well as along the stream reach (Fig. 4). Seasonally, average DOC concentration decreased as the snowmelt period progressed during each year – events A–E (2017), J–L (2018), and N–M (2019). In summer 2017, stream DOC concentrations were relatively constant at C5 (19-20 m L-1), whereas they decreased during the same period at C6 (from 28 to 18 mg L-1) (events F–I). Spatially, stream DOC concentrations generally decreased along the C5–C6 reach (8 out of 14 sampling campaigns) (Fig. 4). However, stream DOC concentrations clearly increased along the reach for the increasing limb of snowmelt 2017 (event B), the summer storm event (event F) and the artificial drought (event G). During the recession limb of the snowmelt peak 2017 (event D) and the lake flooding experiment (events H and I), stream DOC concentrations were relatively constant along the reach.

Abrupt changes in stream DOC concentrations occur in those sections affected by DRIPs (Fig. 4). During most snowmelt campaigns (events A–C, J–N) and summer baseflow conditions (event E), stream DOC concentrations sharply decreased in sections fed by DRIPs. The only exception was the section affected by the last DRIP, from which stream DOC concentrations tended to increase. Peaks in DOC concentrations immediately downstream from DRIPs also occurred during the summer rain event and the experimental drought (events F and G). For the other sampling campaigns (events D, H, and I), both increases and decreases in DOC concentrations occurred at DRIP locations.

The ability of the models to simulate the magnitude and longitudinal patterns of stream DOC concentrations varied across sampling campaigns (Fig. 4). For most sampling campaigns performed during the snowmelt period, at least one of the models was able to capture either the magnitude or spatial variations in stream DOC concentrations (Fig. 4). For six events (B–D, K, and M–N), all the models captured the magnitude of stream DOC concentrations (median PBias from -5 % to 5 %; Fig. 5), yet only for event K were all of them also able to simulate their longitudinal patterns (median R>0.80; Fig. 6). Indeed, none of the models captured spatial patterns for events B, D, and M (median R<0.50), although the Diff-Bio model tended to perform better than the others (Fig. 6). For event C, spatial patterns were captured by both the Diff-Bio and UCA-Bio models (median R∼0.75; Fig. 6), whereas models UCA and UCA-Bio were able to simulate patterns of stream DOC concentrations for event N (median R∼0.70, Fig. 6). For the three other field campaigns performed during the snowmelt period (events A, J, and L), the magnitude of stream DOC concentrations was only successfully captured by models UCA and/or UCA-Bio (median PBias ∼ 0 %, Fig. 5), despite all models being able to simulate the spatial patterns of DOC concentrations (R>0.70; Fig. 6).

For the sampling campaigns performed during summer 2017 (events E–I), there were large inconsistencies across models (Fig. 4). For summer baseflow conditions (event E), the Diff-Bio and UCA-Bio models successfully simulated both the magnitude (PBias < 3 %) and spatial patterns (R>0.90) of stream DOC concentrations (Figs. 5 and 6). For the natural rain event (event F), all the models underestimated stream DOC concentrations (PBias < -5 %) and also failed to predict their spatial variation (median R<0.35). Yet the UCA model performed better than the others (Figs. 5 and 6). None of the models was able either to capture spatial patterns during the lake flooding (event H, R<0.5), even though all the models captured the overall magnitude of DOC concentrations (Figs. 5 and 6). For the experimental drought (event G), the Diff and UCA models successfully simulated spatial patterns of stream DOC concentrations (R>0.70; Fig. 5), yet only the Diff model accurately captured their magnitude (median PBias = 1 %; Fig. 6). Similarly, only the Diff model was able to simulate both the magnitude and spatial pattern of stream DOC concentrations for the post-flooding campaign (event I; Figs. 4–6).

DRIPs are an important, and often primary, source of water and C to headwaters (Briggs and Hare, 2018; Demars, 2019) and play a major role in regulating spatial variation in stream DOC concentration, processing, and export. Our spatially explicit surveys revealed that longitudinal patterns of stream DOC concentrations varied across flow conditions. In general, DOC concentrations tended to decrease along the C5–C6 reach, indicating that DOC was generally diluted or taken up along the stream segment. However, we observed step changes in stream DOC concentrations at DRIPs, indicating that these locations play a key role in stream C export. Similar patterns have been observed in boreal headwaters (Duvert et al., 2018; Lupon et al., 2019), yet our study is the first to reveal the mechanisms by which DRIPs shape spatial patterns of DOC concentrations and fluxes in these streams.

Our results show that accounting for spatial variability in lateral groundwater inflows in the models (i.e., UCA) improved simulations of stream DOC concentrations for five out of the nine sampling campaigns performed during the snowmelt period (events A, C, J, L, and N), when groundwater inputs were high (>20 L s-1) and/or contributed significantly (>40 %) to streamflow. During these events, sharp decreases in DOC concentrations were observed in sections fed by DRIPs, suggesting that these preferential groundwater flow paths mostly diluted the DOC concentrations in the stream. Further, UCA improved the simulations of stream DOC concentrations during the summer rain (event F), when the net gain in streamflow along the reach was 46 %. In this case, however, DOC concentrations increased in most DRIP locations, indicating that these flow paths were acting as important sources of C to the stream. Previous studies have observed that high groundwater tables associated with rain events often increase groundwater DOC concentrations by activating the dominant source layer (Ledesma et al., 2018), which might explain the observed increase in DOC concentrations along the reach observed for event F. Regardless of the process (i.e., DOC delivery or dilution), these findings corroborate that the spatial variability in groundwater flow paths related to landscape topography has a major influence on stream C patterns when streams are mostly fed by groundwater flow (Covino et al., 2021; Dupas et al., 2021; Rocher-Ros et al., 2019).

Accounting for spatial variability in groundwater inflow was important during some, but not all, events. Representing UCAs did not improve model simulations for the sampling campaigns close to a snowmelt peak (events B, K, and M) despite groundwater inputs being elevated (25–36 L s-1). Our explanation is that, during these events, increases in the groundwater level might homogenize groundwater inflows along the reach, potentially generating overland flow because of soil frost causing impervious conditions (Ploum et al., 2020). Similarly, a potential homogenization of overland flow during the snowmelt can explain the decline in DOC concentrations during events K and M (Ploum et al., 2018; Laudon et al., 2011). In any case, from our work it is evident that model frameworks that integrate the spatial arrangement of groundwater flow paths (i.e., DRIPs) can help represent the variability in the hydrological connectivity along the stream.

Model simulations also revealed that accounting for in-stream uptake downstream of DRIPs improved predictions of longitudinal patterns of stream DOC concentrations for the five sampling campaigns occurring from May to August (events C–E, H, and M), but especially during summer low-flow conditions (event E). It is likely that these results are explained by the seasonal pattern of microbial activity in boreal streams, which often mirror the temporal variation in water temperature (Burrows et al., 2017). However, in-stream uptake did not improve model simulations during those dates in summer characterized by very low (event G) or high (event I) flows. These results concur with other recent studies (Lupon et al., 2019; Seybold and McGlynn, 2018; Demars, 2019) and suggest that aquatic biological activity is enhanced at the transition between low and high flows due to increases in labile DOC supply from terrestrial systems. Conversely, in-stream DOC uptake may be minimal during low flows due to C limitation (Burrows et al., 2017) or overwhelmed by high water velocities during rain events (pulse-shunt concept; Raymond et al., 2016). Most importantly, our findings support the idea that spatial patterns of DOC concentrations along headwaters are associated with aquatic biological activity and stream water permanence driven by terrestrial flow path organization (Hale and Godsey, 2019). Collectively, these results suggest that DRIPs are important hotspots of DOC uptake and, thus, the capacity for processing DOC of boreal headwater streams is closely tied to the spatial arrangement of lateral inputs of DOC from riparian zones.

Our model framework represented the source of lateral DOC inputs based on groundwater samples from a riparian well network that compared DRIP and non-DRIP groundwater chemistry (Ploum et al., 2020). This allowed us to distinguish between the spatial variability in riparian groundwater chemistry associated with different soil wetness regimes (Vidon, 2017). For example, during the experimental drought (event G), the Diff model provided better simulations of both the magnitude and spatial patterns of stream DOC concentrations compared to the assumption of uniform inputs along the reach, suggesting that the representation of spatial variability in groundwater DOC concentrations was more important than hydrology or in-stream uptake. Hence, under these conditions, stream DOC patterns might not be directly related to groundwater fluxes but rather to the thermal and chemical conditions that groundwater discharge creates at the local level (Briggs and Hare, 2018). However, there are also limitations in our groundwater sampling approach. For example, our groundwater sampling was not able to represent temporal DOC dynamics associated with variability in groundwater travel times (Heidbüchel et al., 2020), event-scale variability in riparian DOC mobilization (Werner et al., 2019), or the activation of DOC from different soil layers (Ledesma et al., 2018).

Apart from the limitations of our groundwater sampling, we identified some limitations in the hydrological and biogeochemical components of the model as well. The sampling campaign of the summer rain event (event F) is a clear example of mismatch between our simulations and the observations. During this event, stream DOC concentrations increased along the reach, but most of our models simulated a decreasing pattern. Similarly, none of the models properly simulated the longitudinal patterns for two sampling campaigns performed around the snowmelt peak (events B and M). These examples suggest that the model framework has some limitations in representing the complex hydrologic and biogeochemical dynamics occurring in headwater streams (Ambroise, 2004; Klaus and Jackson, 2018). For instance, our models do not account for local conditions affecting snowmelt rates on hillslopes (i.e., shading, sun exposure) or local variations in precipitation, interception, or infiltration that are relevant during rain events (Laudon et al., 2004; Lyon et al., 2010). For the biogeochemical component of our model, we did not take into account processes that produce (i.e., resuspension) or remove (i.e., photodegradation, sorption, flocculation) DOC from the water column (Droppo et al., 1998; Kaiser and Kalbitz, 2012). Further, in-stream DOC uptake was assumed to occur only downstream of DRIPs and at a uniform rate across flow conditions. Previous studies have shown that uptake rates can vary over time as a function of temperature, DOC composition, and microbial assemblages (Berggren et al., 2009; Mineau et al., 2016). While the use of other values for Vf generally resulted in similar model output (Fig. 5 and 6), we cannot rule out the idea that Vf varied among DRIPs and/or over time due to changes in groundwater DOC composition and temperature. To better understand the role of DRIPs in stream hydrology and biogeochemistry, future empirical studies testing how DRIPs affect specific processes are needed. Nevertheless, our study, even with its limitations, demonstrated that both lateral discrete and diffused inputs as well as biological activity are essential components of the DOC patterns in boreal streams. These findings shed new light on the understanding of C dynamics across the boreal aquatic–terrestrial interface.

Another major limitation of our models is their large uncertainty, especially during events with large groundwater contributions such as event B. For six events (A–C, G, J, and L), all the models showed large inconsistencies among runs, resulting in simulated DOC concentrations at C6 that vary over 10 mg L-1. Moreover, the uncertainty in groundwater DOC concentrations was large, because not all stream sections were sampled and groundwater inputs of DOC had to be estimated based on means of the available DOC concentrations from non-DRIP wells. For future studies, we have identified two more directions that can be useful for improving the simulations of stream DOC dynamics along boreal headwaters. For the representation of the spatial heterogeneity in riparian hydrochemistry, the hydrological representation of lateral groundwater inputs through the distinction of DRIP and non-DRIP riparian zones can be further developed. For this matter, integrative hydrochemical frameworks that represent fluxes from various soil layers would be useful to include, especially at non-DRIPs, because here groundwater levels are more dynamic compared to DRIPs (Seibert et al., 2009; Ploum et al., 2020). Furthermore, it can be of interest to downscale the number of riparian groundwater chemistry samples to understand which minimum set of samples is required to represent the spatial heterogeneity in sources of lateral DOC inputs from riparian zones to streams. A preliminary analysis indicated that the most optimal strategy for reducing model uncertainty was to monitor DRIPs individually while averaging DOC concentrations at non-DRIPs (Ploum, 2021). However, given that non-DRIP groundwater chemistry changes with groundwater table fluctuations (Ledesma et al., 2015; Ploum et al., 2020), it is likely that optimizing groundwater sampling campaigns requires careful consideration of the antecedent groundwater conditions.