Dynamics and potential drivers of CO2 concentration and evasion across temporal scales in high-alpine streams

Carbon dioxide (CO2) evasion from streams greatly contributes to global carbon fluxes. Despite this, the temporal dynamics of CO2 and its drivers remain poorly understood to date. This is particularly true for high-altitude streams. Using high-resolution time series of CO2 concentration and specific discharge from sensors in twelve streams in the Swiss Alps, we studied over three years the responsiveness of both CO2 concentration and evasion fluxes to specific discharge at annual scales and at the scale of the spring freshet. On an annual basis, our results show dilution responses of the streamwater CO2 likely attributable to limited supply from sources within the catchment. Combining our sensor data with stable isotope analyses, we identify the spring freshet as a window where source limitation of the CO2 evasion fluxes becomes relieved. CO2 from soil respiration enters the streams during the freshet thereby facilitating CO2 evasion fluxes that are potentially relevant for the carbon fluxes at catchment scale. Our study highlights the need for long-term measurements of CO2 concentrations and fluxes to better understand and predict the role of streams for global carbon cycling.


Introduction
Inland waters are now recognized as important components of the global carbon cycle (Cole et al 2007, Battin et al 2009, Drake et al 2018 with total carbon (C) evasion fluxes to the atmosphere possibly as high as 3.88 Pg C yr −1 (Drake et al 2018). Among the inland waters, headwater streams-the smallest but most abundant streams in fluvial networks-are estimated to contribute approximately one third to the global carbon dioxide (CO 2 ) evasion flux (Marx et al 2017). Our understanding of the role of headwater streams for large-scale carbon fluxes is largely based on the study of headwater streams draining biomes with large carbon stocks (Johnson et al 2008, Lauerwald et al 2015 and generally attributed to the close connectivity with the terrestrial environment, which delivers large amounts of carbon, including CO 2 from soil respiration, to the headwaters (Hotchkiss et al 2015, Tank et al 2018. However, not all headwater catchments are rich in organic carbon, which is particularly true for mountain catchments above the tree line. Discharge is a master variable controlling ecological and biogeochemical processes in stream ecosystems. At catchment scale, the response of streamwater solute concentrations (C) to discharge (Q), or specific discharge (q) (here we use q for the sake of crosscatchment comparability), provides information on the sources of solutes within the catchment, their size and arrangement, and mobilization and transportation to the streams (e.g. Godsey et al 2009, Meybeck andMoatar 2012). Invariant responses of C to q are indicative of chemostasis and may reflect a uniform distribution of solutes within the catchment (e.g. in soils), where changes in hydrological connectivity and flow-paths position do not alter C in the streamwater. Chemostatis may also be linked to mineral weathering and its associated processes (e.g. Clow and Mast 2010). Chemodynamic responses indicates a change of C Original content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence.
Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
with q, where increasing q can either dilute or concentrate a solute in the streamwater. Responses of C to q are typically described by a simple power function, C=aq b , where the exponent (b) indicates whether the response is chemostatic (b=0) or chemodynamic (concentration: b>0; dilution: b<0) (e.g. Godsey et al 2009, Meybeck andMoatar 2012). This approach has been widely used to understand event-scale (e.g. storms) behavior of solutes and more recently also inter-annual solute dynamics in streams, and to infer drivers (e.g. source versus transportation limitation) that act at catchment scale (e.g. Godsey et al 2009, Meybeck andMoatar 2012). Numerous studies have focused on the behavior of conservative solutes, including dissolved ions, but few have adopted the C-q scaling approach to gases, such as CO 2 (Liu and Raymond 2018), to understand their temporal and spatial dynamics. Unlike non-gaseous solutes that are transported downstream through advective flow, CO 2 may not only be converted to other carbonate species such as HCO 3 − , but may also be outgassed vertically from the streamwater into the atmosphere (figure 1). The CO 2 evasion flux (F CO2 ) depends on the gas exchange velocity (k CO2 ) and the gradient between atmospheric and streamwater CO 2 concentrations. Streamwater CO 2 concentration and F CO2 are mutually dependent because (at oversaturation) the latter can deplete the CO 2 pool within the stream through atmospheric loss, which in turn can diminish F CO2 (Rocher-Ros et al 2019). This relationship is further complicated by the dual role of q. As shown above, q drives solute dilution and concentration behavior in streams and at the same time it influences the gas exchange velocity through the water surface (Raymond et al 2012. A recent survey by Liu and Raymond (2018) shows that roughly 50% of the streams and rivers throughout the USA had positive responses of CO 2 concentration to q, which would suggest increased CO 2 deliveries from the catchment outbalancing F CO2 from these systems. This is in line with observations that hydrological connectivity, as encapsulated by changes in q and its relationship with groundwater, can affect the transportation of CO 2 from various sources (e.g. soil respiration, geogenic origin) within the catchment to the streams Based on the previous considerations, we present a framework based on the relationship between the responsiveness of CO 2 concentration to q (b C ) and the responsiveness of F CO2 to q (b F ), with the aim to gain mechanistic understanding of the dynamics of CO 2 evasion from streams and its linkage to processes operating at catchment scale ( figure 1). An underlying premise to this is that F CO2 scales with q similarly as C. That is, a b F >0 would indicate a responsiveness of F CO2 owing to increasing gas exchange velocity, Figure 1. Conceptual framework proposed to study the CO 2 dynamics in streams. (a) At catchment scale, sources (S) and transportation (T) can shape the responsiveness of solutes concentration to specific discharge (q). (b) There is a mutual relationship between CO 2 concentration and the CO 2 evasion flux (F CO2 ). The concentration is also influenced by source and/or transportation, while both the CO 2 concentration and the F CO2 are dependent on q. (c) The responsiveness of CO 2 concentration (b C ) and F CO2 (b F ) to q can be quantified by partial least squares regressions on log-log transformed data. The b C to b F space serves as a frame to explore the potential drivers on the CO 2 dynamics in streams. For instance, the domain defined by b C <0 and b F <0 indicates that source/ transportation limits CO 2 concentration and hence F CO2 . Depending on the limitation the system state can move to an alternative domain (as indicated by the blue arrow) where limitation relief results in a higher responsiveness of F CO2 to q.
possibly also because of no CO 2 limitation in the streamwater. A b F <0 would indicate that CO 2 depletion in the streamwater in combination with high gas exchange velocity would drive the responsiveness of F CO2 to q. We postulated that low responsiveness of both CO 2 concentration and F CO2 to q would limit F CO2 through low CO 2 concentration (i.e. dilution) as it happens for instance when CO 2 from terrestrial deliveries and/or in-stream respiration are reduced. Alternatively, CO 2 dilution but high F CO2 responsiveness to q indicates enhanced F CO2 with low CO 2 concentrations but high turnover. On the other hand, high responsiveness of both CO 2 concentration and F CO2 to q also enhances F CO2 but because of elevated CO 2 concentrations in the streamwater. Our conceptual framework serves as guidance to understand the balance between CO 2 supply to the stream and its outgassing from the stream, which ultimately affects the role of streams for large-scale CO 2 fluxes (Liu and In this study, we use high-resolution temporal data over two consecutive water years to assess the CO 2 dynamics across a range of twelve streams in the Swiss Alps across different timescales. On a yearly basis, we anticipated an overall limitation on the supply of CO 2 from the sources within the catchments, as they are often devoid of major vegetation coverage and soil horizons rich in organic carbon. We also expected low streamwater CO 2 concentrations resulting from the combination of low CO 2 supply and high gas exchange velocities. Furthermore, we postulate that there are windows when supply limitation becomes relieved, and streams receive larger deliveries of CO 2 from sources within the catchment, which would increase the responsiveness of both CO 2 concentrations and F CO2 . We complemented sensor data with occasional measurements of the isotopic composition of CO 2 to explore its potential sources. Our findings shed new light on the CO 2 dynamics in high-altitude streams, further contributes to a better understanding of the role of these ecosystems for global carbon fluxes.

High-frequency measurements and isotope sampling
In each stream, we measured streamwater pCO 2 , temperature and depth, as well as barometric pressure every 10 min across three years. We prepared the pCO 2 sensors (Vaisala CARBOCAP ® , GMT220, Vantaa, Finland) according to Johnson et al (2010), where sensors were contained within a polytetrafluoroethylene (ePDFE) semi-permeable membrane sealed with liquid electrical tape. Sensors were further protected with a metal casing and powered by a solar panel. Sensors were maintained and data downloaded on average every month. Raw data were corrected according to the manufacturer's recommendations for streamwater temperatures (HOBO U24-001 Conductivity Logger, ONSET, Bourne, USA) and barometric pressure (Track-It TM Logger, Monarch Instrument, Amherst, USA), as well as for hydrostatic pressure differences caused by varying water depths. Before deployment, we tested all pCO 2 sensors in the laboratory using certified gas mixtures of CO 2 diluted in synthetic air to final concentrations of 0, 400 and 2000 ppmv. We also performed a laboratory calibration with two of our sensors, which revealed sensor accuracy of −5% and sensor response times between 2.5 and 13 min. Water depth (Odyssey ® Logger, Dataflow Systems Ltd, New Zealand; TruTrack Data Logger, Intech Instruments LTD, New Zealand) was converted to Q from sodium chloride (NaCl) slug additions (Gordon et al 2004). Thereby, we established rating curves for each individual stream (ranging from 4 to 13 NaCl additions in FED and VID, respectively, with an average of 7). From the rating curves, we obtained discharge, which we converted to q by normalizing for drainage area. Using the same approach, we also established rating curves between Q and streamwater velocity (V, ms −1 ) for each site. We determined stream channel slopes (S, mm −1 ) with a dGPS, and the catchment area in ArcGIS 10.5 (Environmental Systems Research Institute, USA) from a 2 m 2 digital elevation model (Geodata © Swisstopo).
When accessible, streams were sampled on a monthly basis for the determination of the isotopic composition of streamwater CO 2 (δ 13 C-CO 2 ; expressed as ‰ VPDB; Vienna Pee Dee Belemnite) using glass vials sealed with rubber stoppers and metal caps. Samples were stored in the dark (4°C) pending analyses within 24 h. In the laboratory, we created a headspace (in the 60 ml sample vials) with synthetic air, shook the samples (2 min) and let them equilibrate (2 h). We measured CO 2 concentrations and δ 13 C-CO 2 using a cavity ring-down spectrometer (Model G2201-I, Picarro Instruments, Santa Clara, CA, USA). Samples for atmospheric CO 2 were collected next to the study streams into glass vials sealed with rubber stoppers and metal caps; we injected additional 50 ml of ambient air to over-pressurize the samples, which were measured on the same Picarro G2201-I as above.

CO 2 concentration and evasion flux calculations
At the 10 min basis, we multiplied monitored streamwater pCO 2 with Henry's constant (K H, mol l −1 atm −1 ) (Plummer and Busenberg 1982) as a function of streamwater temperature and atmospheric pressure (P atm , atm) to obtain streamwater CO 2 concentration. F CO2 were calculated from the CO 2 gradient (ΔCO 2 , mol l −1 ) between the streamwater and the atmosphere, and the gas transfer velocity for CO 2 (k CO2 , m d −1 ) (equation (1)).
To estimate ΔCO 2 , we first derived site-specific time series of atmospheric CO 2 from the air samples that we had collected in proximity of the streams (in average 11 measurements per stream; minimum 5, maximum 16), using linear interpolation between sampling dates. The atmospheric CO 2 was multiplied with K H and P atm at 10 min time steps to obtain atmospheric CO 2 concentrations at saturation ([CO 2sat ], mol l −1 ). We subtracted CO 2sat from the streamwater CO 2 concentrations to obtain ΔCO 2 . Ulseth et al (2019) recently described gas transfer velocities (k 600 , m d −1 ) in turbulent mountain streams as function of energy dissipation (eD, m 2 s −3 ; turbulent stream: eD>0.02 m 2 s −3 ), where eD is calculated as a function of stream flow velocity, stream channel slope and gravity acceleration (g, m s −2 ) (the three variables are multiplied). We performed all sensor corrections and flux calculations in Matlab R2017b.

Data analyses
We analyzed time series covering two water years (defined according to the US Geological Survey as a time period ranging from 1 October to 30 September) using data for a period from 1 October 2016 to 30 September 2018, in order to capture two snowmelt periods, as well as two 'freshet' periods. We define the freshets as the first flushing events during the onset of the snowmelt. We collapsed the 10 min interval data from the time series to daily median values of CO 2 concentration, q, gas transfer velocity of CO 2 (k CO2 ) and F CO2 . We followed the same approach as Liu and Raymond (2018) to identify the responses of CO 2 concentration, k CO2 , and F CO2 to q, respectively. We used power law functions with transformed data (log (x+50)) to analyze whether the responses were chemostatic or chemodynamic, as well as the magnitude of the response (Godsey et al 2009, Meybeck and Moatar 2012). We fitted the data using partial least squares regressions. We fitted each site and each water year separately, for which we derived statistical parameters. We used α=0.05 as the threshold for statistical significance and the coefficient of determination, R 2 , to determine the goodness of the fit. We used JMP 13 (SAS Institute Inc., Cary, USA) for all statistical analyses.

Hydrological regimes
The hydrological regimes of our study streams are typical for high-altitude systems (Hannah et al 2005, Milner et al 2009. After an extended winter baseflow, where q was relatively stable, the snowmelt period started between March and April depending on the altitude and exposition of the catchments. Snowmelt further shaped the hydrological regimes throughout spring and summer. As expected, q in summer was higher in the glacier-fed streams due to the glacier ice melt. The 2016/2017 winter was milder with less precipitation than the 2017/2018 winter, which resulted in overall higher spring and summer q in the water year of 2018 (figure S2).

Streamwater CO 2 concentration dynamics
Overall, we found low streamwater CO 2 concentrations ( (table S1), with 6.5 and 77.6 μmol l −1 as minimum and maximum concentrations, respectively. Daily median CO 2 concentrations covered a similar range as the 10 min time step (ranging from 21.8 to 36.0 μmol l −1 across all study streams and both water years). Therefore, we used the daily median for all further analyses, which would also smooth potential outliers and make estimates more robust. Despite often incomplete time series owing to sensor malfunctioning or loss (average data coverage across the two water years of 41%; minimum data coverage of 23% at PEU; maximum data coverage of 62% at VAU), our analysis of streamwater CO 2 concentration revealed recurrent seasonal patterns. Specifically, the CO 2 concentrations increased during the recession of the snowmelt and ice melt and further into fall baseflow. The freshet (i.e. onset of the snowmelt period) in early spring was marked by a transient increase in both q and CO 2 concentration in several of our streams (e.g. VAD, VAU, VIM, FED, AND, ANU).

Annual responsiveness of CO 2 concentration and evasion fluxes to specific discharge
On an annual basis, we consistently found inverse relationships between CO 2 concentration and q with a Inter-annual differences in b C may be caused by differences in q, however due to the gaps in the time series, we attribute most the inter-annual differences to missing data, particularly when where different periods of the year excluded from the analyses ( figure 2; figure S2).
Overall our b C values are lower than those reported for 1st-to 6th-order streams (b C ranging from −0.05 to 0) and larger rivers (b C ranging from 0.02 to 0.24) throughout the USA (Liu and Raymond 2018). Low b C values are indicative of dilution as a response to increasing q, which in mountain streams is certainly Figure 2. Response of streamwater CO 2 concentrations to specific discharge in twelve streams in the Swiss Alps, for two water years (from 1 October to 30 September). Shown are daily median CO 2 concentration and daily median specific discharge (log-log). Data are color-coded according to the Julian day and the lines represent the linear regressions for each water year (2017, 2018); see table S2 for regression statistics. facilitated by an acceleration of k CO2 that also increases with discharge (hence also with q) and related hydraulics. The b C values from our study streams are comparable to those reported from small boreal streams draining forested and peatland catchments (b C ranging from −0.36 to −0.08) . Wallin et al (2010) found more negative responses of CO 2 concentration to discharge in streams with lower pH and hence lower carbonate buffering capacity. We did not find a similar trend, possibly also because the pH in our study streams was typically >8. Only the streams in the Valsorey catchment were transiently undersaturated in CO 2 with respect to the atmosphere. We tentatively attribute this to low dissolved inorganic carbon concentrations in these streams (2-3 times lower compared to the other catchments) and thus less potential for inorganic carbon buffering.
The overall low b C values reported here may also be linked to the sparse sources from where CO 2 may emanate. This notion is supported by the trend showing b C values decreasing with increasing altitude across both water years (R 2 =0.39, n = 9, P=0.07; VEL and VID were outliers and excluded from this analysis; VIU did not show a significant response in CO 2 concentration to q), which could be linked to decreasing vegetation coverage with altitude in those two catchments. This notion is supported by increasing b C values with increasing vegetation coverage (R 2 =0.44, n=10, P=0.05; VID and FED were outliers and excluded from this analysis).
Near-chemostatic behavior can occur when CO 2 supply balances CO 2 export through evasion and downstream transport fluxes. We attribute the annual near-chemostatis observed in some of our streams to substantial deliveries CO 2 via groundwater. In fact, groundwater has been shown to be important for the CO 2 dynamics in the streams in the Vallon de Nant catchment (RIC, AND, ANU) (Horgby et al 2019). Beyond our mountain streams, groundwater is now being increasingly recognized to drive CO 2 concentration and fluxes in various headwater streams (Duvert et al 2018, Lupon et al 2019. Because of the inherent link between gas exchange velocity and discharge, we found positive responses in k CO2 to q (100%, P<0.05) and F CO2 to increasing q , respectively (53%, P<0.05) ( Table S2). Of the 24 b F , only 2 were from non-significant relationships (from VAD 2017 and PEU 2018). The median b F value for the 22 significant F CO2 -q relationships was −0.004, ranging from −0.20 to 0.32 across the streams (table S2). Liu and Raymond (2018) found, in their systematic survey on US streams and rivers, the highest b F values (0.23-0.31) in small streams. We relate the different relationships observed in their study and ours to the overall low CO 2 concentrations in our mountain streams. In fact, low streamwater CO 2 concentration reduces the CO 2 gradient between the streamwater and the atmosphere, and hence F CO2 . This is further supported by the positive relationship between b F and vegetation coverage (n=11, R 2 =0.33, P=0.05; VID excluded as outlier) for the same reasons as discussed above.
The notion of increased CO 2 deliveries during freshet (i.e. the onset of snowmelt), is further supported by our stable isotope analyses. Across all streams, stable isotope analysis consistently revealed depleted CO 2 compositions (δ 13 C-CO 2 ) in spring (April: median δ 13 C-CO 2 : −15.81‰; March: −15.47‰; May: −14.39‰) but more enriched composition in August (−11.73‰) and later in February (−11.83‰) and January (−11.90‰) (streams were not accessible in December) (figure 3). These isotopic compositions indicate that during the spring freshet, there are proportionally less contributions of CO 2 from geogenic sources (i.e. more enriched compositions). Moreover, it suggests CO 2 from soil respiration as a source to the streams during the spring freshet, coinciding with the hydrological activation of the headwater network. In fact, respiratory CO 2 ultimately from the heterotrophic breakdown of organic matter typically has isotopic compositions ranging from −34 to −24‰ (Wang et al 1998) and is hence more depleted than atmospheric CO 2 and CO 2 with carbonate origin (Clark and Fritz 1997). Our observation of a respiratory CO 2 pulse during the freshet is in agreement with previous observations from high-altitude and high-latitude catchments (e.g. Dinsmore and Billett 2008;Dinsmore et al 2013), and further corroborates the notion of microbial activity underneath the snow cover leading to CO 2 accumulation (Mast et al 1998). Increasing hydrological connectivity during snowmelt facilitates the transportation of this CO 2 via shallow groundwater flow paths to the streams (Doctor et al 2008). The delivery of CO 2 from the terrestrial environment to streams is analogous to the DOC flushing during snowmelt (sensu Boyer et al 2000), and as it has been observed in our study streams as well (Boix Canadell et al 2019). This further supports the relevance of this short window for carbon fluxes at the scale of high-altitude catchments.
Our findings reveal the window of the freshet as potentially important for carbon fluxes in high-altitude catchments. During freshet, the F CO2 increased rapidly over a short time period. For instance, during 5 d of freshet in 2017, F CO2 at FED increased from 1.5 g C m −2 d −1 (11 May, 2017) to 2.2 g C m −2 d −1 (16 May, 2017). Similarly, during 16 days of freshet in 2018, F CO2 at ANU increased from 3.0 g C m −2 d −1 (1 April, 2018) to 8.0 g C m −2 d −1 (16 April, 2018).

CO 2 dynamics across temporal scales
This study reveals different responsiveness of CO 2 concentration and fluxes to q depending on the temporal scale. In line with our conceptual framework (figure 1), we found a clear transition of the CO 2 dynamics and its potential drivers from an annual to the event-driven (i.e. freshet) scale (figure 4). On an annual basis, the relationships between b F and b C remained largely restricted to the domain where CO 2 concentration apparently limited F CO2 . Here, the streams draining the catchments with the highest altitudes tended to have lower F CO2 (average −0.002±0.79, n=6) compared to the streams draining the catchments with lower altitudes (average 0.009±0.64, n=6). This is intuitive as CO 2 source limitation is more pronounced in these catchments as in those at lower altitude and with higher organic  . Responsiveness of F CO2 (b F ) and CO 2 (b C ) concentration to increasing specific discharge across time scales. On an annual basis, the relationship between b F and b C was largely constrained to the domain where CO 2 dilution governs the evasion fluxes (see also figure 1). This was particularly true for high-elevation catchments. During the freshet (onset of the snowmelt period) the CO 2 source limitation became relieved, which resulted in a higher F CO2 responsiveness to specific discharge. carbon stocks. The latter have higher potential to generate CO 2 from soil respiration, for instance, that can be delivered to the streams. The freshet is a window where the CO 2 source limitation becomes transiently relieved enabling an increase in F CO2 . This is furthermore facilitated by an accelerated gas exchange due to increasing q.

Conclusions
Across the 12 studied mountain streams, we found varying and temporally dynamic responsiveness of CO 2 , and F CO2 , to q. Streamwater CO 2 dilution, likely a consequence of source limitation, was the general response to increasing specific discharge on an annual basis. The spring freshet was found to be a window where source limitation was relieved and CO 2 from soil respiration could be flushed to the streams. This window is potentially relevant for carbon fluxes at the catchment scale as it is certainly susceptible to snow dynamics owing to climate change in the Alps.