Stream metabolism controls diel patterns and evasion of CO2 in Arctic streams

Abstract Streams play an important role in the global carbon (C) cycle, accounting for a large portion of CO2 evaded from inland waters despite their small areal coverage. However, the relative importance of different terrestrial and aquatic processes driving CO2 production and evasion from streams remains poorly understood. In this study, we measured O2 and CO2 continuously in streams draining tundra‐dominated catchments in northern Sweden, during the summers of 2015 and 2016. From this, we estimated daily metabolic rates and CO2 evasion simultaneously and thus provide insight into the role of stream metabolism as a driver of C dynamics in Arctic streams. Our results show that aquatic biological processes regulate CO2 concentrations and evasion at multiple timescales. Photosynthesis caused CO2 concentrations to decrease by as much as 900 ppm during the day, with the magnitude of this diel variation being strongest at the low‐turbulence streams. Diel patterns in CO2 concentrations in turn influenced evasion, with up to 45% higher rates at night. Throughout the summer, CO2 evasion was sustained by aquatic ecosystem respiration, which was one order of magnitude higher than gross primary production. Furthermore, in most cases, the contribution of stream respiration exceeded CO2 evasion, suggesting that some stream reaches serve as net sources of CO2, thus creating longitudinal heterogeneity in C production and loss within this stream network. Overall, our results provide the first link between stream metabolism and CO2 evasion in the Arctic and demonstrate that stream metabolic processes are key drivers of the transformation and fate of terrestrial organic matter exported from these landscapes.

mineralized in soils and subsequently transported to streams in gas form (Öquist et al., 2009), respired within the stream ecosystem (Fisher & Likens, 1973;Hedin, 1990), and/or photo-oxidized in the water column (Cory, Ward, Crump, & Kling, 2014). Resolving these different pathways is necessary to determine the fate of OC at regional scales, including the magnitude of CO 2 evasion and water-borne C export to recipient systems (Webb, Santos, Maher, & Finlay, 2018).
While processes delivering CO 2 to streams have been extensively researched, individual studies often reach different conclusions in terms of assigning relative contribution to any one mechanism. A potential reason for these differences is that each pathway operates within distinct compartments of fluvial ecosystems, and thus, studies on specific mechanisms often fail to capture others (but see Demars, 2018;Lupon et al., 2019;Rasilo, Hutchins, Ruiz-González, & del Giorgio, 2016). For instance, studies of photo-oxidation suggest this as an important CO 2 source in streams (>70%; Cory et al., 2014), but these typically consider only processes that occur in the water column. Other studies indicate that the contribution from soil respiration to stream CO 2 evasion is more than 90% (Winterdahl et al., 2016), but these often neglect the potential role of benthic and hyporheic processes. Finally, while there has been decades of research on stream metabolism (Hoellein, Bruesewitz, & Richardson, 2013), these rates have not been integrated with estimates of CO 2 evasion until recently (Hotchkiss et al., 2015). Overall, while the different conclusions drawn from these studies likely reveal real variation in contributing processes among systems, the large variability also reflects the challenge of partitioning these sources at meaningful spatial and temporal scales.
One way to partition the different pathways contributing to CO 2 evasion is to couple continuous measurements of stream CO 2 dynamics with independent and simultaneous estimates of aquatic ecosystem metabolism based on O 2 measurements. Ecosystem metabolism in streams has been measured and modelled for decades using diel measurements of O 2 concentrations in water (Hall & Hotchkiss, 2017;Odum, 1956). This approach assumes that the concentration of O 2 in water is affected by three processes: (a) gross primary production (GPP) that produces O 2 , (b) ecosystem respiration (ER) that consumes O 2 and (c) stream water turbulence that affects the air-water exchange of O 2 . Recent advances in O 2 sensor technology, together with new modelling tools, make it possible to estimate daily GPP, ER and net ecosystem production (NEP; NEP = GPP − ER) using continuous time series of O 2 , light and hydrological parameters (Appling, Hall, Yackulic, & Arroita, 2018;Hall & Hotchkiss, 2017;Holtgrieve, Schindler, Branch, & A'mar, 2010).
Importantly, GPP and ER also consume and produce CO 2 , respectively, and thus provide estimates of aquatic C processing rates that can be compared to independent measures of CO 2 . In this way, estimating metabolism modelled from O 2 data is a powerful tool to understand CO 2 sources to streams (Hotchkiss et al., 2015), yet few studies have coupled high frequency measurements of O 2 and CO 2 with the goal of resolving these different pathways (but see Gómez-Gener, von Schiller, et al., 2016;Stets et al., 2017).
The few existing studies that address how stream metabolism may contribute to CO 2 evasion are from boreal (Crawford, Striegl, Wickland, Dornblaser, & Stanley, 2013;Rasilo et al., 2016) or temperate ecosystems (Cole & Caraco, 2001;Crawford et al., 2014;Gómez-Gener, von Schiller, et al., 2016;Hotchkiss et al., 2015), while studies of these processes in the Arctic tundra are lacking. Tundra streams are characterized by cold temperatures, long days with high incident light during a short summer and winters that span for more than 6 months. Yet, these streams often drain soils with large C stocks (Schuur et al., 2015), and export vast quantities of OC to the Arctic ocean (Cooper et al., 2008). Furthermore, streams represent considerable sources of CO 2 evasion in the Arctic landscape (Lundin et al., 2016;Stackpoole et al., 2017) and emit more C than is exported to the ocean (Serikova et al., 2018). Given that climate change is drastically altering the hydrology and biogeochemistry of Arctic landscapes (Drake et al., 2018;Kendrick et al., 2018), understanding how C is mineralized and evaded within streams (e.g. Giesler et al., 2013) is necessary to understand and predict the effects of environmental change on C cycling in this region.
In this study, we ask: how do stream metabolic processes affect CO 2 dynamics and evasion in Arctic stream networks? To answer this, we measured CO 2 and O 2 concentrations continuously in six streams in an Arctic catchment during the summer of 2015 and 2016. Specifically, we (a) quantified the contribution of the stream NEP to CO 2 evasion and (b) explored whether GPP can explain diel changes in CO 2 evasion. To achieve this, we modelled metabolic rates using the O 2 data and estimated CO 2 evasion simultaneously.

| Site description
The Miellajokka catchment (52.5 km 2 ) in north-western Sweden se/abisko). Climate in the Miellajokka catchment is characterized by long winters with precipitation as snow from October to May and a short terrestrial growing season from June to September (Christensen et al., 2012). Hydrologic patterns reflect the seasonal climate regime, with a spring flood in May or June during snow melt (discharge at the outlet of 20-25 m 3 /s), and base flow of about 0.05-0.1 m 3 /s during the autumn and winter (Lyon et al., 2018). Dissolved organic carbon (DOC) in Miellajokka can reach 8-10 mg C/L during spring flood and decrease to about 2 mg C/L during summer base flow (Giesler et al., 2014). Dissolved inorganic carbon is around 4 mg C/L during summer base flow conditions, but is lower during spring flood (<2 mg C/L; Giesler et al., 2014).
The pH in the catchment is circumneutral and with little seasonal variation . The stream network ranges from first to fourth Strahler order streams (Table 1), with a total length of 44.6 km and a total stream surface area of 0.151 km 2 (Rocher-Ros, Sponseller, Lidberg, Mörth, & Giesler, 2019). Streams are moderately steep with slopes ranging from 0.07 to 0.32 m/m (Lyon et al., 2018), and several medium-sized waterfalls. There are two lakes in the catchment, covering in total 0.69 km 2 .
The Miellajokka catchment is north-facing and elevation ranges between 384 and 1,731 meters above sea level (m a.s.l.). In this F I G U R E region, sporadic permafrost occurs at low elevations and discontinuous to continuous permafrost at high elevation zones (Gisnås et al., 2017). At elevations above 1,200 m a.sl., the land is mostly barren, with several permanent snowfields. Between 700 and 1,200 m, the landscape is characterized by tundra vegetation and cryoturbated soils (Becher, Olid, & Klaminder, 2013). The tree line is at approximately 700 m, and below this elevation, the landscape consists of sparse mountain birch forest (Betula pubescens spp. Czerepanovii) with mixed tundra heath vegetation. Below 400 m, there is a more productive birch forest with a denser canopy cover. The sites M1 and M16 are located here, and the riparian forest cover results in less incident light compared to streams draining tundra vegetation (Myrstener et al., 2018).

| Continuous measurements
We recorded water temperature, water level and dissolved concentrations of CO 2 and O 2 at six stream locations from late June to early September in 2015 and 2016 ( Figure 1). Water temperature and water level were recorded hourly using HOBO water level loggers (model U20-001-04; Onset Computer Corporation). Stream CO 2 concentrations were measured hourly using infrared gas analyser (IRGA) adapted for wet environments. In streams M1 and M16, we used a Vaisala GMT220 sensor (Vaisala) covered with a PTFE layer highly permeable to dissolved gasses but not to water, following (Johnson et al., 2010). At sites M6, M9, M10 and M17, we used eosGP CO 2 concentration probes (Eosense Inc.). The eosGP sensor uses the same technique as the Vaisala, but with a PTFE membrane included by design. The Vaisala and eosGP sensors were connected to CR1000 data loggers (Campbell Scientific Inc.), powered with 12 V lead-acid batteries. The sensors were calibrated with standard gases in the lab before and after deployment in the field, using gas concentrations of 400, 2,000 and 5,000 ppm of CO 2 . Sensors were placed with protective casings to avoid damage due to floods and rock movements in the water and were inspected and gently cleaned every 3 weeks.
Due to the fragile material of the membrane and the extreme conditions in some streams, several malfunctions and subsequent data loss occurred, particularly at M1. We monitored O 2 concentrations every 10 min using miniDOT oxygen loggers (Precision Measurement Engineering Inc.). The loggers were installed with a copper mesh to avoid biofouling, and the sensor was placed in the opposite direction of the flow to prevent accumulation of debris and impact of stones.
Prior and post deployment, the sensors were intercalibrated using reaerated water to achieve a 100% saturation of O 2 , and then by adding dry yeast to decrease the O 2 saturation to 0%.
All loggers were attached in the stream using a perforated steel pipe attached to a heavy metal platform to prevent movement.
The temperature/water level loggers were placed firmly inside the pipe, the CO 2 sensor outside but downstream of the pipe, to be exposed to flowing water, and the O 2 sensor parallel to the flow with the sensor facing downstream. We selected these sites taking into account three criteria: (a) a suitable location within the thalweg to install loggers so that they would not be exposed to air during base flow conditions, while also avoiding deep pools; (b) lack of upstream tributaries (in all streams except M16 the distance to the nearest tributary was >1 km); and (c) minimal groundwater inputs immediately upstream of deployment sites. On two to six occasions, we quantified local groundwater inputs by comparing discharge estimates made with salt slugs at the deployment site with those made upstream; 50-500 m, depending on stream size. For each site, we observed similar discharge values, differing less than 10%, for example, the precision of the slug discharge measurements (Moore, 2005). This indicates low rates of groundwater input within the likely footprint of the metabolism estimate.
Snow/ice cover and peak flow conditions during snow melt restricted the time period of our measurements to June-September.
Due to these climatic constraints, in 2015, we installed the loggers between 5 and 7 July until 7 September, and in 2016, between 15 and 17 June until 8 September. Other climatic variables used in this study were atmospheric air pressure and light irradiance. We used data measured in the meteorological station in Stordalen (SITES Sweden monitoring station, circa 4 km from the catchment outlet).
To obtain atmospheric pressure in each site, the atmospheric pressure was corrected by the elevation difference following the barometric formula (Hall & Hotchkiss, 2017).

| Discharge and the gas exchange coefficient (K 600 )
Discharge (Q) was measured at every site on several occasions with the salt slug method (Moore, 2005). At M1, M6 and M16, we obtained more than 10 measurements, while in sites M9, M10 and M17, we performed four discharge measurements. The discrete measures were then related to depth that was continuously monitored with a pressure logger to obtain continuous discharge estimates. The relationship between depth and discharge used was linear, with an R 2 > .85 in all streams. To relate the depth of the logger position to the average channel depth, we measured depth every 5-20 cm (depending on the stream size) along 8-10 cross sections upstream of the sensors at each site.
Briefly, at sunset when GPP approaches zero, O 2 in water decreases as there is no biological input. The rate of decrease in O 2 concentrations is therefore dependent on the rate in which O 2 can reach a new equilibrium with the atmosphere, and thus proportional to the K 600 . During the period when this occurs, K O 2 is approximated by the slope of the relationship between the rate of change in O 2 concentration and the O 2 deficit in the water (Odum, 1956), that can be converted to K 600 (Aristegi, Izagirre, & Elosegi, 2009). Given that the length of the night shifts strongly through the summer at high latitudes, we used an algorithm to perform six linear regressions each day at different periods to capture the night-time drop of O 2 using an R script (https ://github.com/rocher-ros/night time_regre ssion_multiple). We selected days when night was at least 2 hr long (from 25 July onwards), and days when the night-time regression had an R 2 > .7. The K 600 values obtained with this method were then related to discharge, usually a major predictor of K 600 within a site (Raymond et al., 2012). In our case, K 600 was significantly related to discharge in all sites ( Figure S4). At three sites, we also performed several propane releases, following the method described in Wallin et al. (2011), and the K 600 obtained by this method agreed well with the night-time regression estimates ( Figure S4a,b,e).
To calculate the specific gas aeration coefficients for O 2 and CO 2 , K O 2 and K CO 2 , we used the following approach. The K 600 is a standardized measure of the gas exchange coefficient for a Schmidt number of 600, which can be converted to a specific gas following (Wanninkhof, 1992): where K x is the gas exchange coefficient for a given gas x, and SC x is the Schmidt number of that gas (in this study CO 2 or O 2 ). The Schmidt numbers for each gas were calculated using the published Schmidt coefficients (Raymond et al., 2012), for O 2 was calculated as: And for CO 2 as: where T is the water temperature in °C. With the gas specific Schmidt numbers (Equations 2 and 3), it was therefore possible to calculate the K O 2 and the K CO 2 (Equation 1).

| Stream metabolism modelling
Stream metabolism was modelled based on the open channel diel oxygen method (Odum, 1956). Stream NEP is the balance between GPP and ER, and these two processes affect the diel oxygen concentration. These diel patterns can be used to estimate GPP and ER by analysing O 2 time series. We used a Bayesian inverse model from Hall and Hotchkiss (2017), governed by the following equation: where O 2 t is the oxygen concentration at time t (in g O 2 /m 3 ), z is the channel depth (in m), PAR is the photosynthetically active radiation (in mol m −2 s −1 ), K O 2 is the gas exchange coefficient of O 2 (in day −1 ), Δt is the time steps of the time series (10 min) and O 2 sat is the concentration of O 2 in the water if it would be 100% saturated. GPP and ER are obtained as areal rates (g O 2 m −2 day −1 ) and were converted to C assuming that 1 mol of O 2 is produced/consumed for 1 mol of CO 2 (Demars et al., 2016). We acknowledge that the conversion between O 2 and CO 2 depends on the chosen respiratory or photosynthetic quotient and could thus bias results (Berggren, Lapierre, & Del Giorgio, 2012;Williams & Robertson, 1991).
We modelled the three parameters (GPP, ER and K O 2 ), but using priors for K that were strongly constrained to minimize the problem of equifinality (Appling et al., 2018). Models that predict the three parameters avoid errors associated with estimating K 600 empirically (Aristegi et al., 2009;Holtgrieve, Schindler, & Jankowski, 2016), but can give multiple solutions where different combinations of GPP, ER and K 600 reproduce the same O 2 data, so-called equifinality (Appling et al., 2018).
A solution for this is to relate K 600 to hydrological measures such as discharge, which should be a proxy for K 600 within a site (Appling et al., 2018). We used the relationship between K 600 and Q from each site obtained from the night-time regression method (see above), to obtain an approximate K 600 for each day with its error associated. Then, for each day, the prior distribution of K O 2 was defined by the mean and standard deviation (SD) obtained from the K 600 -Q relationship ( Figure S4). The priors for GPP and ER were largely uninformed, with a mean of 1 and −5 g O 2 m −2 day −1 , respectively, and an SD of 2. The priors of GPP and ER were chosen to be similar to the mean values measured in another Arctic stream in Alaska (Huryn, Benstead, & Parker, 2014).
To simulate the posterior distributions of the parameters, we used the concentrations. If the MAE was larger than 0.2, we discarded that day. The threshold of 0.2 was determined after visually inspecting the plot of O 2 concentrations and was similar to the threshold used in another study (Lupon et al., 2019). (b) One of the model assumptions is that depth and K 600 are constant throughout the day (Odum, 1956). We removed days when depth (which is also the proxy used for K 600 ) changed more than 10% within the day. (c) Finally, daily outputs were plotted to visually inspect that the model reproduced O 2 concentrations accurately. Here, we inspected each day manually and removed any days showing poor model fit. After this, 165 observations of daily metabolism were removed from a total of 875.

| Estimating CO 2 evasion
The CO 2 exchange with the atmosphere (E CO 2 ) was calculated as (Raymond et al., 2012): where K CO 2 (in day −1 ) is the gas exchange coefficient, z is the channel depth (in m), [CO 2 ] w is the concentration of CO 2 measured in the water and [CO 2 ] a is the CO 2 concentration in equilibrium with the atmosphere (in mol/m 3 ). We used an atmospheric CO 2 concentration of 380 ppm, obtained from the mean of several air CO 2 measurements performed in the field. The concentrations of CO 2 in mol/m 3 were calculated using the pCO 2 measurements and Henry's law, using the temperature measured in the oxygen sensor. The units of E CO 2 were converted from mol C m −2 day −1 to g C m −2 day −1 .

| Mass balance along a single stream reach
In a previous study in this catchment, we measured CO 2 evasion and discharge at a high spatial resolution (Rocher-Ros et al., 2019). Here, we used this data set to do mass balance calculations for CO 2 in order to generate estimates of net CO 2 production along a stream reach (2.1 km) that loses a major fraction of water into the nearby forest as it crosses an alluvial deposit ( Figure S13). We used these independent estimates and compared them with estimates derived from metabolism modelling. Along this reach, CO 2 concentrations, discharge, K 600 and channel hydraulics (wetted width, depth and velocity) were measured every 300-480 m. Therefore, it is possible to use a mass balance calculation for CO 2 within each segment of this reach: where C out is the CO 2 leaving the segment, calculated as the product of discharge at the downstream end (Q out ; in m 3 /day) and the CO 2 concentration (CO 2 out ; in g C/m 3 ); C in is the CO 2 entering the segment, calculated as the product of discharge at the entrance (Q in ; in m 3 /day) and the CO 2 concentration (CO 2 in ; in g C/m 3 ); C GW is the CO 2 input from groundwater (GW), as the product of groundwater flow (Q GW ) and the groundwater CO 2 concentration (CO 2 GW ; in g C/m 3 ); P is the production of CO 2 within the stream segment; and E is evasion of CO 2 in the stream segment. Thus, to estimate the unknown P (stream production of CO 2 ), Equation (6) can be rearranged as: Equation (7) can be further decomposed in its components as: where E CO 2 is the CO 2 evasion rate (in g C/m 2 ) using the average CO 2 concentration and A is the stream segment area (in m 2 ). Q GW can be estimated as the difference between Q in and Q out for each stream segment. Importantly, because this is a losing reach, there is no net increase in groundwater contribution; therefore, all CO 2 produced originated within the stream channel, and the parameter CO 2 GW is the mean stream CO 2 concentration. All these parameters were measured in the field and therefore used to estimate the internal stream CO 2 production (P).

| Data analysis and statistics
All data were analysed using R (R Core Team, 2017; version 3.5.1), the data set with daily summary data and an R script to reproduce these figures can be found in the Supporting Information. Linear regressions were performed to test the prediction that GPP is related to diel changes in CO 2 evasion. The ΔCO 2 to summarize the diel change in CO 2 concentration was calculated as the difference between the highest and lowest CO 2 concentration within each day. The diel change in CO 2 evasion was calculated as the cumulative CO 2 evasion occurring between sunrise and sunset, and subtracting the CO 2 evasion before sunrise. The coefficient of variation (CV) was calculated as SD/average × 100, where SD is the standard deviation. Significant differences refer to the p < .05 level unless otherwise stated.

| Physical and chemical characteristics of streams
Overall, streams were clearly separated by K 600 , with the more turbulent sites (M6, M9 and M10) having the highest values, ranging between 21 and 57 day −1 (Figure 2). By contrast, in less turbulent streams (M1, M16 and M17), K 600 values were considerably lower, that is, ranging between 4 and 17 day −1 (Figure 2). Henceforth, the streams are labelled so that those that have a low K 600 (M1 LK , M16 LK , M17 LK ) are easily separated for those with high K 600 (M6 HK , M9 HK , M10 HK ).
All streams were supersaturated in CO 2 , with average concentrations ranging from 740 to 2,460 ppm (Table 1). We observed the highest average CO 2 concentrations in the two smallest streams, M17 LK and M16 LK , and the lowest concentration in M10 HK (Table 1). The amplitude of diel change in CO 2 concentration ranged from 0 to 920 ppm; this varied throughout the measuring periods, but also differed markedly among streams ( Figure S1). Specifically, the diel change in pCO 2 (ΔCO 2 ) was more pronounced in M1 LK , M16 LK and M17 LK compared to M6 HK , M9 HK and M10 HK (Figures 3 and 4). The average diel change of

| Stream metabolic rates
Rates of ER were an order of magnitude higher than GPP ( Figure S6).
Average ER across all streams was −1.8 g C m −2 day −1 , with individual site averages ranging from −1.35 (M17 LK ) to −2.63 g C m −2 day −1 (M1 LK , Table 2). Temporal variation in ER, described by the % CV, was greatest at M10 HK (59%) and lowest at M1 LK (27%). Average GPP across all streams was 0.22 g C m −2 day −1 with averages for individual sites ranging from 0.19 to 0.28 g C m −2 day −1 (Table 2). GPP also varied over time within sites with the highest % CV in M9 HK (85%) and the lowest in M16 LK (41%). GPP and ER were significantly and linearly related in four of the six sites, with a degree of explanation (R 2 ) ranging from .28 to .5 ( Figure S9). In all sites, we found that ER was linearly related with discharge, with an R 2 ranging from .65 to .93 ( Figure S10). GPP was also significantly related to discharge in four of the six sites, with an R 2 ranging from .22 to .64 ( Figure S11). There was also a strong relationship between ER with K 600 in all sites, with an R 2 ranging from .62 to .87 ( Figure S12).

| GPP and diel patterns of CO 2 concentration and evasion
All streams had higher CO 2 concentrations at night compared to day, displaying a clear diel change in pCO 2 (ΔCO 2 ; Figures 3 and 4). The diel pattern in pCO 2 resulted in higher night-time CO 2 evasion rates compared to daytime rates (Table 2). Not surprisingly, this effect on CO 2 evasion was highest in the streams with a strong diel pattern F I G U R E 3 Daily variations in pCO 2 (panels a, c and e) and the relationship between the diel change in CO 2 evasion and gross primary production (GPP; panels b, d and f), in the three streams with low K 600 (Figure 2). ΔCO 2 is the daily change in pCO 2 from midnight, where each solid line represents 1 day, and the in pCO 2 (Figure 3). Specifically, CO 2 evasion at midnight compared to noon was 45%, 37% and 34% higher in sites M1 LK , M16 LK and M17 LK respectively. The impact on CO 2 evasion for the streams with a weaker diel pCO 2 pattern was lower but still important, with midnight evasion rates 26% and 24% higher than noon in sites M9 HK and M10 HK respectively. In site M6 HK , the pCO 2 diel pattern was the weakest, and CO 2 evasion rates at midnight were just 1% higher than noon. The magnitude of diel change in evasion was positively related to GPP rates in all streams, with significant relationships in all cases except M6 HK during the year 2015 (Figures 3 and 4). For the streams with large diel changes in pCO 2 (M1 LK , M16 LK and M17 LK ), GPP explained between 31% and 78% of the variability in the diel change in CO 2 evasion (Figure 3b,e,f). For the streams with low K 600 , the degree of explanation of GPP was weaker, ranging from 4% to 58% (Figure 4b,e,f).
Furthermore, the effect of GPP was also visible directly on diel changes in pCO 2 in the three streams with low K 600 . In M1 LK , GPP explained 74% and in M16 LK 83% of the amplitude in ΔCO 2 , with values close to the 1:1 line ( Figure S8). For stream M17 LK , GPP also had a significant, linear relationship with the diel pattern in pCO 2 and explained 65% and 32% of the variability for the years 2015 and 2016 respectively ( Figure S8). In the other three F I G U R E 4 Daily variations in pCO 2 (panels a, c and e) and the relationship between the diel change in CO 2 evasion and gross primary production (GPP; panels b, d and f), in the three streams with high K 600 (Figure 2 Figure 4a,c,e). At these sites, where K 600 was higher than 20 day −1 , and hence, streams were more turbulent, GPP had no significant relationship with the diel change in CO 2 concentrations ( Figure S8).

| CO 2 evasion and the contribution of stream metabolism
Average daily CO 2 evasion rates of the sites ranged from 0.1 to 6.2 g C m −2 day −1 (Table 2), with an average of 1.4 g C m −2 day −1 .
The highest average evasion rate was at M16 LK and the lowest at M10 HK (Table 2). All streams were undersaturated with O 2 and supersaturated with CO 2 relative to the atmosphere (Figure 5a).
Consequently, all streams had negative NEP (ER > GPP) and were therefore net sources of CO 2 , with NEP rates comparable to CO 2 evasion rates (Figure 5b). Average NEP among streams ranged from −1.2 to −2.4 g C m −2 day −1 ( and CO 2 concentrations also captured similar results but without the effect of K 600 and its potential uncertainties (Figure 5a). All streams were close to the 1:1 line and were significantly related, with the highest R 2 found in the streams M1 LK , M9 HK , M16 LK and M17 LK (.73, .5, .85 and .45, respectively), while for M6 HK and M10 HK , the R 2 was .06 and .07 respectively. Although the departure from O 2 and CO 2 equilibrium does not incorporate the effect of the K 600 , its effect determines the potential for the departure.
Here, the streams with high K 600 (M6 HK , M9 HK , M10 HK ) are closer to saturation for both O 2 and CO 2 than the streams with low K 600 (M1 LK , M16 LK , M17 LK ). Additionally, the spread along the 1:1 line was also larger for low compared to high K 600 streams (Figure 5a).
The streams M16 LK and M17 LK showed an offset relative to the 1:1 line, indicating that there is an external source of CO 2 uncoupled from O 2 dynamics. This external source of CO 2 for these same streams is also detected by comparing NEP and CO 2 evasion rates (Figure 5c), where NEP accounts for <100% of CO 2 evasion rates.

| CO 2 mass balance along a stream reach
Mass balance calculations along single segments of the stream reach provided evidence for net CO 2 production within the stream. The water lost along the entire reach ( Figure S13) was more than 50% (Figure 6a), as discharge decreased from 1.33 to 0.68 m 3 /s. Along this same distance, pCO 2 increased more than twofold, from 400 to 1,000 ppm (Figure 6b).
K 600 also decreased markedly along the reach, with K 600 values dropping (300-480 m) ranged from 0.86 to 5.46 g C m −2 day −1 , with an average of 2.6 g C m −2 day −1 (Figure 5c). This average CO 2 production in the reach was similar to the average CO 2 evasion (2.7 g C m −2 day −1 ) and close to the NEP (−2.81 g C m −2 day −1 ) measured the same day at the site M1 LK which is 800 m downstream of this reach.

| D ISCUSS I ON
In this study, we simultaneously assessed continuous O 2 and CO 2 data to show that aquatic biological processes play an important role in the C cycle of these Arctic streams. In the Swedish northern landscape, the signature of aquatic metabolism was imprinted upon stream CO 2 dynamics in two distinct ways: photosynthesis created a clear daynight difference in CO 2 evasion and in-stream respiration sustained CO 2 evasion from streams throughout the summer. Streams were consistently heterotrophic, indicating that respiration in these ecosystems relies on organic C exported from land. Thus, through both autotrophic and heterotrophic processes, aquatic metabolism has the potential to regulate the transformation and the fate of terrestrial organic matter exported from Arctic landscapes.

| Diel patterns in CO 2 evasion
We observed a consistent and sometimes dramatic day-night change in pCO 2 (Figure 3) with night-time evasion rates that were between 24% and 45% higher than during the day in five of the six streams, a similar magnitude as reported in other studies in lower latitude regions (Peter et al., 2014;Reiman & Jun Xu, 2018;Schelker, Singer, Ulseth, Hengsberger, & Battin, 2016). Our results further indicate that this diel change in CO 2 evasion was caused by photosynthetic activity during the day (Figures 3 and 4). The effect of GPP was also visible directly in diel changes in pCO 2 , but only in streams with less turbulence and lower K 600 ( Figure S8). This suggests that degassing in more turbulent streams conceals the effect of GPP on stream CO 2 concentrations, as observed for O 2 concentrations (Appling et al., 2018). Regardless, despite relatively low GPP rates ( Figure S6), photosynthesis acts as important, short-term C sink in these streams.
Furthermore, this day-night pattern implies that estimates of CO 2 evasion based on daytime observations may grossly underestimate the total daily efflux, in this study by as much as 27%. By showing how low and high K 600 environments differ in their capacity to support strong diel patterns, these results may help to correct regional and global estimates of CO 2 evasion.
Our results show that aquatic photosynthesis drives diel changes in CO 2 evasion and pCO 2 in these Arctic streams (Figures 3 and 4).
However, in the Alaskan Arctic, it has been suggested that photooxidation can account for as much as 70%-95% of the CO 2 production in the water column of streams and rivers (Cory et al., 2014).
If this light-dependent process was the main driver of CO 2 production in our streams, we would expect to see an increase in pCO 2 from night to day, that is, in contrary to our observations ( Figure 3).
The discrepancy of our results with Cory et al. (2014) could be due to the clear, low DOC water in Miellajokka streams (Giesler et al., 2014), as compared to the more coloured and DOC rich waters in Alaska. Still, photochemical measurements are performed in the water column, which represents a minor fraction (<5%) of C mineralization from benthic and hyporheic sediments (Demars, 2018).

F I G U R E 6
Patterns of CO 2 concentrations and fluxes in a losing water stream. (a) Downstream change in discharge along the 2 km stream reach. The stream reach loses water as it passes through an alluvial deposit (see Figure S13 for a spatial version of this figure). We therefore expect that the contribution of terrestrially respired CO 2 is negligible as there are no groundwater inputs. (b) How the pCO 2 increases a twofold along this reach. By assuming that lateral inputs are negligible, we can do a mass balance to quantify the CO 2 produced within the stream. (c) Calculated inputs and export of CO 2 for five stream segments of the 2 km stream reach. The CO 2 was produced at a rate of 2.6 g C m −2 day −1 in this reach, and the net ecosystem production the same day measured at the site M1 (~800 m downstream) was 2.8 g C m −2 day −1 0.6 0.8 Evasion Import upstream CO 2 production Inputs: Outputs: Distance downstream (m) Export dowstream + GW Indeed, even considering the highest rate of photo-oxidation from Alaska (0.3 g C m −2 day −1 ; Cory et al., 2014), this process would only account for 20% of average CO 2 evasion in our streams, and an even lower fraction in other Arctic sites that have reported considerably higher evasion rates (Denfeld, Frey, Sobczak, Mann, & Holmes, 2013;Lundin, Giesler, Persson, Thompson, & Karlsson, 2013;Serikova et al., 2018).

| Contribution of stream NEP to CO 2 evasion
While GPP can have a strong impact on stream CO 2 dynamics, rates of ER were an order of magnitude higher (Table 2), and therefore had a stronger overall effect on the stream C cycle. Indeed, NEP in our streams was strongly negative due to high ER rates, a common observation across riverine ecosystems (Hoellein et al., 2013), and was the major contributor to CO 2 evasion ( Figure 5). This indicates that these streams mineralize substantial amounts of the organic C received from land that otherwise would have been exported downstream to lakes or marine systems. Our reported values of the contribution of aquatic NEP to CO 2 evasion are high compared to other studies of small streams in high latitudes (40%-75%; Lupon et al., 2019;Rasilo et al., 2016), and typically the largest contributions to date have been observed for considerably larger rivers (85%-97%; e.g. Cole & Caraco, 2001;Lynch, Beatty, Seidel, Jungst, & DeGrandpre, 2010). Therefore, our results seemingly contradict the expected minor contribution of stream NEP to CO 2 evasion in headwaters (Hotchkiss et al., 2015), although we did find an increase in the average contribution of NEP with stream size (Figure 5c).
The discrepancy of our results with other studies reporting smaller contribution of aquatic NEP to CO 2 evasion may reflect constraints imposed on site selection when estimating stream metabolism. Importantly, we avoided reaches for metabolism modelling that had high rates of groundwater input and/or areas with very high gas exchange (e.g. waterfalls), which are both likely hotspots of C inputs or evasion (Lupon et al., 2019;Rocher-Ros et al., 2019). This decision may explain our relatively low CO 2 evasion rates compared to other studies in the Arctic (Denfeld et al., 2013;Lundin et al., 2013;Serikova et al., 2018). Even within the Miellajokka catchment, the CO 2 evasion rates observed here are lower (median 1.4 g C m −2 day −1 ; Table 2) than those reported in a previous study based on synoptic sampling of 168 locations in this same catchment (median: 3.3 g C m −2 day −1 ; Rocher-Ros et al., 2019). Similarly, the median gas transfer velocity (K 600 standardized by depth) in that synoptic study was much higher (54.5 m/day) than the median gas transfer velocity in this study (7.4 m/day), suggesting that reaches selected for metabolism estimates do not represent the most important locations for CO 2 evasion in the network. Indeed, if the median NEP from this study (1.4 g C m −2 day −1 ) is representative of the catchment, this would indicate that stream NEP only accounts for 40% of the CO 2 evasion estimated from the more spatially extensive sampling effort. This contribution is more similar to other studies (Hotchkiss et al., 2015;Lupon et al., 2019;Rasilo et al., 2016). Thus, without a spatial assessment of CO 2 evasion that included other hotspots of C inputs and evasion (Rocher-Ros et al., 2019), the conclusion of this study would have overestimated the contribution of in-stream metabolism. This stresses the importance of combining different tools, approaches and scales that capture unique pathways for C processing and evasion in stream networks.
While stream respiration appears to be important for CO 2 evasion, capturing these rates and understanding their underlying drivers remain sources of uncertainty. Respiration rates in streams can be regulated by temperature (Demars et al., 2016;Song et al., 2018) and organic C supply: either autochthonous (i.e. GPP ;Huryn et al., 2014) or allochthonous (i.e. litterfall or DOC; Demars, 2018;Roberts, Mulholland, & Hill, 2007). In our study, ER was strongly related to discharge ( Figure S10), which has been reported elsewhere for other small northern streams (Demars, 2018;Lupon et al., 2019). These authors suggest that discharge could regulate the activity of stream heterotrophs through the delivery of terrestrial organic C. Consistent with this supply mechanism, previous work in the Miellajokka catchment has shown that DOC increases with discharge (Giesler et al., 2014). Thus, the positive relationship between ER and discharge reported here could reflect real hydrological processes that drive the supply and processing of terrestrial organic C in these streams.
Despite this plausible mechanism, the close correspondence between ER and discharge needs to be taken with caution because it also reflects covariance between K 600 and ER ( Figure S12). In this study, the observed relationship between ER and discharge (or K 600 ) emerges from a persistent deficit of O 2 across a large range of flow conditions (see Figures S2 and S3). The covariance between ER and another parameter such as K 600 can be problematic when studying within site variability of ER, and so we are conservative and focus on average rates of ER, as other studies have done (Blaszczak, Delesantro, Urban, Doyle, & Bernhardt, 2018). Regardless, since K 600 is used both for metabolism modelling and for CO 2 evasion (Equations 4 and 5), potential biases arising from K 600 estimates would affect NEP and CO 2 evasion rates to a similar extent and direction. This is also reflected in the similar departure from equilibrium for both O 2 and CO 2 (Figure 5a), which indicates a strong coupling of both gases in these streams. Finally, mass balance estimates of CO 2 production provided an independent validation of NEP rates, which were remarkably similar to NEP measured via metabolism modelling in the same stream on the same day (2.6 g C m −2 day −1 vs. 2.8 g C m −2 day −1 ; Figure 6). Together, these multiple observations provide additional confidence in our conclusions regarding the important role of aquatic respiration to CO 2 evasion. Strikingly, our results further suggest that rates of NEP can exceed evasion locally, leading to an accumulation and downstream export of CO 2 ( Figure 5). While this condition (NEP > E CO2 ) was evident from our continuous, modelled data, we also tested whether this is reasonable using a mass balance approach. In this case, along a 2 km stream reach, we observed a large increase of pCO 2 ( Figure 6).
Given that this is a hydrologically loosing reach, most of the CO 2 must be produced internally and thus originate from stream processes. Furthermore, continuous sensor data identified this reach as one of the sites where NEP > E CO2 (Site M1 LK in Figure 5). Together, these observations suggest that in-stream biological processes can actively generate CO 2 in streams and override lateral transport.
Overall, whether a stream reach can or cannot export CO 2 downstream will ultimately be controlled by the turbulence of the water and the capacity to evade CO 2 , which is highly variable at fine spatial scales (Rocher-Ros et al., 2019). The interplay between stream reaches that are importers or exporters of CO 2 creates a strong heterogeneity and dynamism within stream networks that has important implications for understanding how C is processed and evaded along the aquatic continuum.

| Understanding the effects of climate change for C cycling in high-latitude streams
The Arctic is currently confronted by a wide array of changes due to global warming, with increased temperatures that result in the mobilization of old OC in soils (Schuur et al., 2015) and increased discharge into the Arctic Ocean (Peterson et al., 2002). These changes are currently altering the functioning of stream ecosystems (e.g. Kendrick et al., 2018) and also appear to have strong effects on CO 2 evasion from fluvial networks (Serikova et al., 2018). Our results suggest that Arctic streams and rivers not only play an important role in C export and CO 2 evasion, but are active components in the biological mineralization of OC.
However, the Arctic is large and diverse, with variation in permafrost extent and soil C storage (Hugelius et al., 2014), as well as regional differences in vegetation structure and growth trends (Huang et al., 2017), which together underpin largely unknown variability in stream biogeochemistry and aquatic ecosystem dynamics. Thus, multiple Arctic regions may respond uniquely to global change, and this variability needs to be captured in future studies, given current focus on few Arctic areas (Metcalfe et al., 2018).
Regardless, owing to the importance of the Arctic C feedback on climate change (Schuur et al., 2015) and the dependence of stream respiration to discharge and C supply (Demars, 2018), we suggest to include stream metabolism and its response to environmental change (e.g. Song et al., 2018) in future scenarios for the prediction of the effects of climate change.

ACK N OWLED G EM ENTS
The authors thank Albin Bjärhall and Belén Díaz for help in the field and the lab. The authors are also grateful to the Abisko Research Station and SITES Sweden for the meteorological data. This study was supported by the Swedish Research Council (VR; 2013-5001) and the Swedish Research Council for Environment, Agricultural Sciences and Spatial Planning (FORMAS; 2014-00970) to R.G.
Finally, we thank three anonymous reviewers who provided critical feedback that improved this manuscript.

CO N FLI C T O F I NTE R E S T
The authors declare no competing interests.