C Measurements from 2012 to 2014 at the Environmental Research Station Schneefernerhaus , Germany : Technical Corrections , Temporal Variations and Trajectory Clustering

This study presents continuous atmospheric CO2 and δC measurements by wavelength-scanned cavity ring down spectroscopy (Picarro G1101-i) at the high-mountain station Schneefernerhaus, Germany. δC values were post-corrected for methane and water spectral interferences using accompanying measurements of CH4 and H2O, and CO2 in dried air, respectively. The best precision of ±0.2‰ for δC and of ±4 ppb for CO2 was obtained with an integration time of about 1 hour for δC and 2 hours for CO2. The seasonality of CO2 and δC was studied by fitting background curves for a complete 2-year period. Peak-to-peak amplitudes of the averaged seasonal cycle were 15.5 ± 0.15 ppm for CO2 and 1.97 ± 0.53‰ for δC, respectively. Based on the HYSPLIT Model, air masses were classified into five clusters, with westerly and northeasterly flows being the most and the least frequent, respectively. In the wintertime, northwest and northeast clusters had a higher median level for ΔCO2 and a lower median level for ΔδC (the difference between observed and background concentrations), likely caused by anthropogenic emissions. In the summertime, air masses from the northwest had the lowest ΔCO2 and the highest ΔδC. Potential source contribution functions (PSCFs) were used to identify the potential source and sink areas. In winter, source areas for high CO2 mixing ratios (> 75 percentile) were mainly located in northwestern Europe. In summer, areas with high δC ratios (> 75 percentile), indicating a carbon sink, were observed in the air from Eastern and Central Poland.


INTRODUCTION
The sources and sinks of atmospheric carbon dioxide (CO 2 ) play a critical role in governing global climate.Human activities, especially fossil fuel combustion, cement manufacturing and land use changes trigger anthropogenic emissions and thus contribute to a steady increase of the atmospheric CO 2 mixing ratio (IPCC, 2014;WMO, 2016;Xu et al., 2017).Nearly half of the emitted CO 2 is taken up by both the terrestrial biosphere and absorption in the ocean † Now at Springer Nature B.V., Dordrecht, 3311 GX, Netherlands reservoirs with a similar share each (Le Quéré et al., 2016).The atmosphere and these two major sinks are linked via the balance between photosynthesis, respiration fluxes and CO 2 dissolving in ocean water (Ciais et al., 2013).Considering the risk of the ocean sink to become saturated, the risk of the biosphere sink to turn into a source and of a consequential rapid increase in atmospheric CO 2 growth rates, it is crucial to monitor atmospheric CO 2 mixing ratio continuously with high temporal resolution in order to identify the source and sink contributions (Cramer et al., 2001;IPCC, 2013).
The stable carbon isotope composition of CO 2 (δ 13 C) allows distinguishing the exchange of fossil carbon from the atmosphere and surface reservoir fluxes (Keeling et al., 2011).During photosynthesis, most plants of the terrestrial biosphere prefer to take up the light isotopologue 12 CO 2 and thus discriminate against 13 CO 2 (~18‰) (Farquhar et al., 1989).In contrast, during oceanic uptake of CO 2 , there is almost no discrimination (~2‰) (Mook et al., 1974).Since the discrimination by oceans is small, the carbon isotope signature (δ 13 C) can be applied to study the contribution of the biosphere on carbon cycles in the atmosphere (Keeling et al., 1989;Miller et al., 2003).
Atmospheric CO 2 and δ 13 C have been studied on both regional and global scales at different sites with various techniques (Levin et al., 1995;Tuzson et al., 2011;Sturm et al., 2013;Moore and Jacobson, 2015;Xia et al., 2015;Pang et al., 2016a).For instance, Xia et al. (2015) analyzed measurements of atmospheric CO 2 concentration and its stable isotope ratios (δ 13 C) at the regional background station Lin'an (LAN) in China to identify the isotopic signature of CO 2 sinks and sources.They concluded that during the winter season (Dec-Feb) coal combustion is the major CO 2 source due to domestic heating.The relatively high isotopic signature (~21.32‰) of sources and sinks during the vegetation season (Mar-Nov) was attributed to the significant contribution of biological activities at LAN. Sturm et al. (2013) presented continuous measurements of atmospheric CO 2 isotopes (δ 13 C and δ 18 O) at the High Altitude Research Station Jungfraujoch (JFJ), Switzerland.Based on the high temporal resolution of their measurements (e.g., in comparison to flask samples), diurnal and hourly variations could be analyzed.They determined that diurnal cycles of atmospheric CO 2 and its isotopic compositions were small; however, the day-to-day variability depending on the origin of the air masses arriving at JFJ was larger.Based on this, the footprint clustering by the backward Lagrangian particle dispersion model FLEXPART revealed different CO 2 , δ 13 C, and δ 18 O values depending on the origin of air masses and surface residence time.Moore and Jacobson (2015) differentiated sources of atmospheric CO 2 by records of CO 2 concentration and carbon isotope composition (δ 13 C) at Evanston, an urban site north of Chicago, the third largest city in the United States.
Measurements of CO 2 mixing ratios in combination with stable CO 2 isotopes in the atmosphere at high altitude (background) sites may provide valuable information on carbon source and sink mechanisms (Zhou et al., 2006;Sturm et al., 2013;Yuan et al., 2018).The variability of the atmospheric CO 2 mixing ratio and δ 13 C is related to the history of air masses arriving at a measurement site, such as whether the air traveled through the free troposphere or has been influenced by the planetary boundary layer (PBL).Therefore, many atmospheric studies have applied trajectory statistical methods to identify the sources and their contribution to mixing ratios.Fleming et al. (2012) reviewed more than 150 studies dealing with back trajectories, cluster analyses and residence time.However, most of them applied air-mass history to identify transport pathways and potential sources but not sink contributions.It is well known that the effect of biospherical activity as well as strong influences of thermal convection (local transport processes from the boundary layer) may complicate the study of atmospheric CO 2 and δ 13 C (Zellweger et al., 2003;Sturm et al., 2013).
In the present study, we (1) describe the specific corrections and selections of the data recorded by a nonupgraded G1101-i analyzer, (2) analyze the diurnal and seasonal patterns of atmospheric CO 2 mixing ratios and δ 13 C and (3) identify potential CO 2 source and sink locations with the help of backward trajectories and the potential source contribution function.

Measurement Site
The Environmental Research Station Schneefernerhaus (UFS) is part of the Global Atmosphere Watch (GAW) Global Observatory Zugspitze/Hohenpeissenberg in Germany.It is located in the northern limestone Alps at the border of Germany and Austria (47°25ʹ00ʺN, 10°58ʹ46ʺE) about 90 km southwest of Munich.The UFS is situated at an altitude of 2650 m above sea level (a.s.l.) on the southern slope of Zugspitze mountain massif.It can receive representative free tropospheric air but is also influenced by polluted air masses from the planetary boundary layer (PBL), especially during the day and in summer (Zellweger et al., 2003;Pandey Deolal et al., 2014;Leuchner et al., 2016).Due to its location on a southern slope, northerly winds are mostly blocked (see Fig. 1) while there is wind channeling of western and eastern winds by the west side of the mountain ridge and the Rein Valley, respectively (Gantner et al., 2003;Risius et al., 2015).Due to diverse and intensive human activities in the vicinity of the UFS, from skiing and hiking areas as well as associated tourist services, the GAW site is categorized as a "weakly influenced, constant deposition" site (Henne et al., 2010;Ferrarese et al., 2015).Further detailed information can be found in the UFS station information system (http://www.schneefernerhaus.de/en/home.html).

Instrumentation
A wavelength-scanned cavity ring down spectrometer (G1101-i, Picarro Inc. USA) was installed in the laboratory on the 4 th floor of the UFS building.It was operated from May 2012 to November 2014 (except for a downtime period in May and June 2013 due to a pump defect and its subsequent replacement).CO 2 mixing ratios and δ 13 C were measured at a rate of 0.1 Hz.The air inlet for the instrument was located on a terrace above the laboratory roof at an altitude of 2670 m a.s.l.The inlet cap was constantly heated to prevent ice formation.The inner part of the inlet tube was made of borosilicate glass and was constantly regulated to a low positive temperature of ca.5°C to avoid water condensation.The inlet tube (length: 3.5 m) was connected to a common manifold (glass, 4.2 m in length, 8 cm inner diameter) that was used for all measuring devices and species at the GAW station (Zellweger et al., 2011).A 2 m 1/8ʺ stainless steel pipe connected the instrument to the inlet line, via a VICI (Valco Instruments Company Inc.) rotary valve that switched between ambient air and standard gases.
Isotopic ratio measurements of 13 C/ 12 C are expressed in per mil (‰), defined by the following Eq.(1) as:  Isotopic values were given relative to the international standard VPDB-CO 2 (Vienna Pee Dee Belemnite) (Brand et al., 2010).Measurements of CO 2 and δ 13 C were calibrated using two working standard gases with high and low concentrations (DEUSTE Steininger GmbH) of the certified CO 2 mixing ratios and the isotopic compositions (Standard 1: 350.1 ± 0.50 ppm CO 2 mixing ratio and -3.28 ± 0.164‰ δ 13 C; Standard 2: 503.4 ± 0.50 ppm CO 2 mixing ratio and -20.03 ± 1.002‰ δ 13 C) in synthetic air.Each standard gas was fed into the analyzer for one hour every 7 days using an open-split configuration.After the data collection period, both gas standards were reanalyzed at the laboratory of the Max Planck Institute for Biogeochemistry (MPI-BGC), Jena, Germany, in order to confirm the quality of the calibration and the stability of the standards.The measured δ 13 C values (Standard 1: -3.19 ± 0.009‰ δ 13 C; Standard 2: -20.26 ± 0.01‰ δ 13 C) agreed well with the original values within the given range of GAW data quality objectives (±0.01‰).The standard gases were also analyzed for other trace gas concentrations.Except for N 2 O (Standard 1 and 2: 0.1 ppm) and CO (Standard 1: 6 ppb; Standard 2: 5.47 ppb), no other species (CH 4 , H 2 O, SF 6 ) could be detected.
The German Meteorological Service (Deutscher Wetterdienst, DWD) provided wind speed and wind direction data from the UFS in a one-minute time interval, and for this study, hourly averaged data are presented.In addition, we used CO 2 and CH 4 data measured with a Picarro Envirosense 3000i instrument to correct our measured data (see the section "Correction of CO 2 ").This instrument, using cavity ring down spectroscopy, was operated by the German Environment Agency (Umweltbundesamt, UBA) connected to the same inlet as our Picarro G1101-i, therefore measured identical air sample, and was operated in the same air conditioned laboratory.For this device, water vapor in the ambient air was removed by cold traps.Gas mixtures for working standards of this device were equally delivered by DEUSTE Steininger GmbH, Germany.The station standards for interconnection with the international standard reference were reported on WMO X-2007 scale for CO 2 and WMO X2004a scale for CH 4 by NOAA, Boulder, Colorado, USA.Calibration and quality assurance for these atmospheric compounds followed the standard operating procedures of UBA in accordance with GAW quality standards.Since CO 2 measurements from UBA are reported to the WMO World Data Centre for Greenhouse Gases, the respective CO 2 data in the section "Correction of CO 2 " will be referred to as CO 2 data of Global Atmosphere Watch (CO 2, GAW ).

Optimal Integration Time and Precision
The Allan variance method (Allan, 1966;Werle et al., 1993;Chen et al., 2016) was applied to determine the optimal integration time for the measurements and to determine the best precision, using Standard 1 (CO 2 : 350.1 ± 0.50 ppm; δ 13 C: -3.28 ± 0.164‰; Fig. 2).For this purpose, a long-term measurement of 24 h duration in February 2014 was analyzed.When integrating less than the optimum integration time, the Allan deviation follows a slope of -1/2 in the double logarithmic scale, indicating that white noise is dominating.When integrating beyond the optimum integration time, the Allan deviation rises and follows a slope of 1/2, indicating an instrument drift.
The optimum integration time minimizing the Allan deviation was around 1 hour for CO 2 and around 2 hours for δ 13 C.The best achievable precision (1 sigma) was 4 ppb for CO 2 and 0.2‰ for δ 13 C, respectively (see Fig. 2).However, we used a shorter integration time of 30 minutes for the ambient measurements, which gave a precision of Fig. 2. Allan deviation plots for CO 2 (left) and δ 13 C (right) as a function of the integrating time τ, based on a long-term measurement of Standard 1 (CO 2 : 350.1 ± 0.50 ppm; δ 13 C: 3.28 ± 0.164‰).f stands for frequency and the black dashed lines represent slopes of -1/2 and 1/2, which correspond to power spectral densities S(f) = f 0 and S(f)= f 2 , respectively.The Allan deviation follows a slope of -1/2 up to an integration time of 1 hour for CO 2 and of 2 hours for δ 13 C and then turns over to a slope of 1/2 which defines a drift.0.4‰ for δ 13 C and 5-6 ppb for CO 2 .For the very same analyzer, Wen et al. (2013) found the best precision of 0.08‰ at 2000 s, Vogel et al. (2013) had a precision of 0.2‰ at 5 min averaging, and Pang et al. (2016b) achieved optimum values of 0.08, 0.15, and 0.10‰ at 7600, 1900, and 1900 s for three reference gases.The precision of all studies on Picarro G1101-i shows better precision than the specification provided by the manufacturer (0.3‰).

Data Calibration and Correction
The G1101-i analyzer in this study was manufactured in 2010 and not upgraded during the measurement period to minimize downtime.As an alternative to upgrading, Picarro Inc. recommended later removal of spectral interference caused by CH 4 that can bias δ 13 C by 0.4‰ ppm -1 (Vogel et al., 2013) as well as water interferences.The latter include water vapor dilution, water vapor pressure broadening, and HDO spectral interference effects (Wen et al., 2013).As we did not use any drying system and therefore measured the humid ambient gas, corrections for the dominating water vapor dilution effect were done according to Hoffnagle (2013); however, the smaller effects due to the water vapor pressure broadening and the HDO spectral interference effects were not corrected and their effects are about 2 ppm %v -1 water at 400 ppm of carbon dioxide, and up to 5‰ at ambient humidity (Nara et al., 2012;Rella et al., 2013;Wen et al., 2013).According to Wen et al. (2013), even an upgrade of the analyzer would have led to overcorrections.The dependency of δ 13 C on the CO 2 mixing ratio is of no concern to this study, as Vogel et al. (2013) showed that no dependency could be detected in the range of 303-437 ppm.Background gas concentrations and ratios (e.g., N 2 /O 2 ratio and argon content) are known to influence the CRDS technology in general.While their natural variations usually generate negligible effects, large differences between ambient air and synthetic standards, such as the ones used in our study, may be problematic (Nara et al., 2012;Rella et al., 2013).However, no specific information or corrections are documented for the G1101-i analyzer.

Correction of δ 13 C
The isotopic analyzer reports the peak height of the near-infrared absorption spectrum for the rovibronic transition of 12 C 16 O 2 and 13 C 16 O 2 with arbitrary labeled C12 peak-BookAve and C13 peak-BookAve , respectively.The ratio of the peak absorption values and the ratio of isotopic abundances are linear.

 
A and B are instrument-specific constants and were provided by the manufacturer; they are the linear slope and intercept terms for computing the delta value from the ratio between C13 peak-BookAve and C12 peak-BookAve (Hoffnagle, 2013).The peak height can be affected by the absorption peaks for water, methane and other gases (Nara et al., 2012;Rella et al., 2013).
δ 13 C ratios were corrected for methane and water vapor according to Hoffnagle (2013) H 2 O is the water vapor concentration (in percent) and (-2.98648648 × H 2 O) is the correction for the water interference.CH 4 is the methane concentration (in ppm), measured by the UBA device (Picarro Envirosense 3000i instrument).

Correction of CO 2
To quantify and correct water vapor effects on the CO 2 measurements, we compared the CO 2 mixing ratio for the wet gas stream measured by Picarro G1101-i analyzer (CO 2,meas ) with CO 2 data of the accompanying instrument (CO 2,GAW ).Although similar seasonal cycles were observed in Fig. 3, CO 2, meas mixing ratios only showed a good agreement with CO 2, GAW in wintertime when water vapor concentrations were low (0.39 ± 0.004%), while clear differences were observed in the summertime when water vapor concentrations were higher (1.81 ± 0.0048%).Wen et al. (2013) also measured a wet gas stream and compared the same analyzer as in this study with the Los Gatos DLT-100, observing that the mixing ratio measured by Picarro G1101-i was 2.2 ± 1.0 ppm lower than Los Gatos.They suggested that the water vapor dilution effect was partly responsible for this difference.
We corrected the dilution effect using ordinary least squares regression of the CO 2 mixing ratio difference between CO 2,GAW and CO 2, meas .Concurrent with a pump failure and subsequent replacement (end of May/beginning of June 2013), the extent of the dilution effect changed and thus OLS regressions were carried out separately for the period before (Eq.( 4)) and after (Eq.( 5)) the pump failure, yielding where [H 2 O] is water vapor concentration in percent, [CO 2 ] meas is CO 2 mixing ratio measured by our analyzer and [CO 2 ] calculated is the calculated and corrected mixing ratio.Fig. 4 illustrates the regressions before and after the pump failure.The physical reasons for the change in G1101-i behavior are not known and it seems unlikely that they were caused by the pump failure and replacement alone.
Corrected CO 2 data thus is in accordance with GAW data quality objectives (±0.1 ppm) (WMO, 2016).However, very small differences between the two analyzers (0.01 ± 0.42 and 0.02 ± 0.06 ppm before and after the pump replacement, respectively) remain and may be due to biases from spectral broadening and interferences (Rella, 2012).

Data Coverage and Selection
The temporal resolution of CRDS raw data should be in the order of seconds.In our case, however, the data periodically included measuring intervals far beyond the normal range (55 s and more).Since long measuring intervals are indicative of a slowed ring-down frequency and potential problems with the instrument (such as laser ageing or deficiencies in the optics, detector or data acquisition system), intervals > 55 s duration were removed from the dataset (personal communication by Dr. Renato Winkler, Picarro Inc.), which amounted to about 4.6% of data.The device returned to normal measuring intervals after restarts or laser readjustments.1.8% of the data were missing due to the pump defect and replacement in May and June 2013.Another 3.8% of the data were missing due to the calibration procedures and/or power failures in the lab.At the end, this resulted in 89.8% of valid data over the entire measuring period.
The first 29 measurements of each calibration (ca.3.5 minutes, less than 6% of each calibration) were discarded to ensure that no air sample from the previous measurement was left due to the transient response after valve switching (Vogel et al., 2013).Then, from the remaining data, the average of each calibration was calculated.A smoothing spline was then fitted to all the averages to account for the residual variation in the calibration data and to fill a gap in the calibration measurements (8 months, July 2012-March 2013).This reduced the residual deviation between calibrated standard measurement and the real standard value from 0.05 ± 0.75‰ and 0.03 ± 0.53‰ (linear interpolation) to 0.0 ± 0.44‰ and 0.0 ± 0.32‰ (smoothing spline) for  Standards 1 and 2, respectively.Splines were calculated using the sm.spline function from the pspline package in R (Heckman and Ramsay, 1996).This function fitted a natural polynomial smoothing spline to the calibration values with 25 degree of freedom (out of 108 calibrations).Deviations between spline interpolated calibrations and the real standard values for each calibration are shown in Fig. 5.The smoothed coefficients were used for calibration.The two-point mixing ratio gain and offset calibration strategy of Bowling et al. (2003) was used for each measurement cycle as described above.Data were aggregated to 30-min averages by the statistical program R (R Core Team, 2016), which together with several packages was used for further analyses.At the end, uncertainty of calibration is estimated to be ±0.56 ppm for CO 2 and ±0.53‰ for δ 13 C and combined measurement uncertainty is estimated to be ±0.56 ppm for CO 2 and ±0.56‰ for δ 13 C.

HYSPLIT Trajectory Model
One common method to establish source/sink-receptor relationships is to combine a calculated trajectory path of an air parcel with measured data at the time when the air parcel arrives at the site and consequently to determine locations of sources and sinks from these observations (Stohl, 1996).
In this work, the HYSPLIT (Hybrid Single-Particle Lagrangian Integrated Trajectory) Model was used to calculate backward trajectories to estimate air mass pathways to the UFS.Trajectories were calculated hourly for 96 h backward ending at UFS for the whole measurement period.The backward trajectory calculation was started at an altitude of 1500 m above ground level (a.g.l.) with respect to the model elevation at the coordinates of UFS, thus 3000 m a.s.l., roughly matching the real site altitude of 2650 m a.s.l.(UFS) and 2670 m a.s.l.(sample inlet).The GFS (Global Forecast System) model "Grid 4" forecast weather data with 0.5° resolution was used as an underlying meteorological model.Forecasts from the 18:00 UTC cycle were obtained from the NOAA HAS data portal each day for the next 24 hours and were converted to ARL input model format.In case of the very few gaps (around 1.2%) in the GFS forecast data, forecasts from the previous day(s) were used as a replacement.For each trajectory time step, coordinates, altitude and mixing layer height (from the input model) were extracted.The backward trajectories were further processed by the openair package (Carslaw and Ropkins, 2012) within R. Here, clusters were calculated from all trajectories by applying an angle-based distance matrix method with the k-medoids algorithm (Sirois and Bottenheim, 1995).The number of clusters was set to five since for this number at least 10% of the trajectories were represented within each cluster.In order to filter the trajectories concerning their background air characteristics, they were split into two classes: (1) FT (free troposphere) trajectories when the trajectory height was higher than the mixing layer height for all time steps and (2) PBL influenced trajectories when the trajectory was below the calculated mixing layer height for at least one time step.
The potential source contribution function (PSCF) method using the residence time probability (Ashbaugh et al., 1985;Seinfeld and Pandis, 2016) describes the spatial distribution of probable geographical source locations derived by trajectories.PSCF is defined as the probability that an air parcel with atmospheric component concentrations higher than a specific threshold arrives at the receptor site after having been observed to reside in a certain grid cell.The PSCF value (P ij ) for a given grid cell is then calculated as P ij = m ij /n ij , in which n ij is the total number of trajectory segment endpoints terminating within the grid cell(i,j) over the entire time of measurement and m ij is the number of trajectory segment endpoints terminating within the grid cell(i,j) corresponding to trajectories associated with concentration values at the receptor site greater than a specific threshold (75 th percentile of CO 2 and δ 13 C in this study).

Robust Extraction of Baseline Signal
In order to calculate the long-term background values, the statistical method REBS (Robust Extraction of Baseline Signal) was applied, which is based on robust local regression (Ruckstuhl et al., 2012).REBS is very flexible to derive the background levels of various trace gases at background measurement sites simultaneously (in our case both CO 2 and δ 13 C) due to its non-parametric basis.Sturm et al. (2013) used REBS for high alpine CO 2 and δ 13 C measurements at Jungfraujoch, with results very similar to those obtained from the commonly applied data filtering method by Thoning et al. (1989).The background concentration was extracted using local regression (60 day windows) implemented in the rfbaseline function of the IDPmisc package (Locher and Ruckstuhl, 2012) in R. Missing data points were interpolated by a simple linear interpolation using the values from the hours before and after.The mean peak-to-peak seasonal amplitude was calculated from the background curves for an entire 2-year period (2012)(2013)(2014).

Remote Sensing Data
In order to assess the vegetation properties, information on the vegetation type and activity in summer (June-August) were required.Therefore, we used remotely sensed phenological data based on MODIS Land Cover (LC) data and the Normalized Difference Vegetation Index (NDVI), respectively.The latest LC was issued in 2012 at 0.5° × 0.5° resolution (Friedl et al., 2010;Channan et al., 2014).More information can be found in the Global Land Cover Facility (http://glcf.umd.edu/data/lc/).The LC classes were aggregated to the ones shown and described in Fig. 11(b).NDVI is derived from multiple Advanced Very High Resolution Radiometer (AVHRR) measurements and is an indicator of the greenness of vegetation in each pixel of the satellite image.It ranges from -1.0 to +1.0.NDVI values below zero are excluded from this study because they indicate no vegetation, such as rock, sand or snow.Sparse vegetation such as grassland or cropland results in moderate NDVI values (0.2-0.5) and dense vegetation such as tropical forests results in high NDVI values (0.6-0.9).In this study, we averaged the respective NDVI values for the three months of summer (June to August) in the year 2012.The NDVI data were downloaded from the ECOCAST directory of the NASA Ames Ecological Forecasting Lab, version 3g.v0 (https://ecocast.arc.nasa.gov/data/pub/gimms/3g.v0/) for 2012.

Atmospheric CO 2 and δ 13 C
Time series of CO 2 mixing ratios and δ 13 C recorded from May 1, 2012, to Nov. 2, 2014, are displayed in Fig. 6.This includes both hourly data and fitted background concentrations, using the REBS technique.Fig. 6 shows a clear seasonal variation of hourly mean CO 2 mixing ratios and δ 13 C values.CO 2 mixing ratios simultaneously increase as δ 13 C values decrease, which is associated with seasonal vegetation activity.The mean peak-to-peak amplitudes are 15.5 ± 0.15 ppm for CO 2 and 1.97 ± 0.53‰ for δ 13 C.The minimum CO 2 mixing ratios, as well as the maximum δ 13 C values, occurred in August due to (preferentially 12 CO 2 ) CO 2 terrestrial uptake being dominated by the biosphere.The maximum in CO 2 mixing ratios occurred in March and the minimum in δ 13 C values in February when respiration is dominating.In contrast, the peak-to-peak amplitudes of CO 2 mixing ratios and δ 13 C at the High Altitude Research Station Jungfraujoch (JFJ) on the northern ridge of the Swiss Alps (46°32ʹ53ʺN, 7°59ʹ2ʺE, 3580 m a.s.l.) are 11.0 ppm for CO 2 and 0.60‰ for δ 13 C (Sturm et al., 2013).These differences in seasonal amplitudes between UFS and JFJ are most likely due to altitude since JFJ is located 930 m higher than UFS.Consequently, air masses at JFJ are much less impacted by PBL and mostly are from the lower free troposphere.Although amplitudes of CO 2 mixing ratios and δ 13 C from these two sites were different, the minimum and the maximum of CO 2 mixing ratios and δ 13 C were recorded nearly at the same time of the year.

Diurnal Cycles
The mean diurnal cycles of CO 2 and δ 13 C for each season are shown in Fig. 7.The CO 2 level started to increase at about 06:00 and reached the maxima around 09:00 to 10:00 during spring, summer, and autumn and at around 13:00 during winter, possibly due to local tourist activity, regional respiration, and regional anthropogenic emission potentially due to traffic.The daily peak-to-peak amplitude is 1.4 ppm in winter and 1.6 ppm in spring and autumn.The maximal diurnal change in CO 2 is found during summer months with a peak-to-peak amplitude of about 2.9 ppm.The strong afternoon drop during summer months is due to upward transport of PBL air from the valley.In summer, the CO 2 of the PBL air in lower levels is depleted due to photosynthetic uptake, indicating the influence of vegetation activity in the lower local and regional area around (below) the UFS.During summertime, the PBL influence at UFS has been identified for other parameters as well.Namely, the diurnal variation and high standard deviation of formaldehyde (HCHO) mixing ratio (Leuchner et al., 2016) and the dependency of aerosol concentrations from the altitude of the mixing layer (Birmili et al., 2009) could be detected.The mean diurnal cycle of δ 13 C is very pronounced in summer with an amplitude of 0.4‰.The respective diurnal amplitudes in spring and autumn are 0.2‰ each, showing similar patterns as in summer.In contrast, winter months do not display any distinct diurnal cycles of δ 13 C.
For the Swiss site Jungfraujoch (JFJ), similar start of increase and maxima have been observed in spring, summer and autumn.However, amplitudes of diurnal variations in CO 2 mixing ratios and δ 13 C were different (Sturm et al., 2013).The diurnal peak-to-peak amplitude of CO 2 at JFJ Fig. 7. Diurnal variations of the mean hourly CO 2 mixing ratios and δ 13 C (solid line) relative to the respective daily means for the different seasons.Dashed lines show the 95% confidence intervals of the hourly mean calculated by bootstrap re-sampling.in summer was 2 ppm (0.9 ppm less than at the UFS) and was 1 ppm in the other seasons (0.6 ppm less than at UFS in spring and autumn and 0.4 less in winter).In contrast to UFS, there were almost no diurnal variations of δ 13 C at JFJ in spring, autumn and winter.In summer, JFJ exhibited a diurnal variation of about 0.3‰ less than at UFS with a comparable minimum in the morning and maximum in the afternoon.The larger amplitudes at UFS are likely due to a lower elevation (i.e., the station being closer to sinks and sources in the valley) and more tourist activity in the vicinity of UFS.Compared to urban sites, which are influenced much more strongly by local biogenic and anthropogenic activities, UFS shows much lower diurnal variability, e.g., by factors of 10 and 34 for δ 13 C and CO 2 , respectively, for a site in Kraków, Poland (Zimnoch et al., 2004).

Back Trajectory Clusters Characterization of the Clusters
The back trajectory analysis described in the section "HYSPLIT trajectory model" is shown in Fig. 8. Cluster 3 (C3) comprises the fastest and the most frequent (41.7%) flow at UFS and corresponds to a westerly flow from the mid-Atlantic Ocean.Cluster 4 is the second most abundant (19.5%), comprising flow from the northwest.Cluster 2 (14.9%) represents a moderately fast southwesterly flow.Slow-moving air masses coming from the central and west Mediterranean Basin are grouped in Cluster 1, representing 12.1% of the data.Cluster 5 corresponds to northeasterly flows and accounts for the smallest share (11.9%) of trajectories.

Variation of CO 2 Mixing Ratios and δ 13 C among the Clusters
In order to study whether measured data of CO 2 and its isotopic composition systematically varied with air mass origin, we merged hourly CO 2 mixing ratios and δ 13 C data with the respective trajectory information of the same time step (hourly) and grouped them into the five clusters described above (Fig. 8).Afterwards, data were filtered in two steps in order to separate the short-term deviations from the background concentrations and to derive clear information on potential background air masses.The background concentrations derived from the local regression were first subtracted from CO 2 and δ 13 C values, providing ΔCO 2 and Δδ 13 C. Based on the 96-hour back trajectories of the HYSPLIT Model (see the section "HYSPLIT trajectory model"), the latter data were grouped into air masses in contact with PBL or of completely free troposphere origin.ΔCO 2 and Δδ 13 C by trajectory clusters and seasons are shown in Fig. 9.The highest frequency of air masses contacting the PBL occurred in summer, with 21.3%, while the lowest frequency occurred in winter, with 10.3%.Spring and autumn had contact with the PBL in 14.6% and 12.6% of the cases, respectively.
In summer, CO 2 mixing ratios of clusters from the south significantly differed from the respective median of clusters from the north and west (significant levels of the Student's t-test are p < 0.001).In wintertime, the two clusters from the north (C4 and C5) showed the most pronounced differences in the measured ΔCO 2 and Δδ 13 C. Clusters 4 and 5 had the highest median ΔCO 2 values and, correspondingly, the lowest values in Δδ 13 C.
Using footprint clustering analysis for ΔCO 2 and Δδ 13 C in wintertime for the JFJ site, Sturm et al. (2013) revealed that a cluster representing air masses with surface contact mostly over northern European land masses had the highest median ΔCO 2 value and the second lowest value in Δδ 13 C. Another cluster with high CO 2 mixing ratio had its origin in Eastern Europe.These two clusters according to residence time maps had the same direction as our C4 and C5 clusters, respectively.C4 was always associated with the lowest Δδ 13 C.This can be due to anthropogenic  emissions from this direction (as shown in Fig. 10) being prevalent in all seasons.Cluster 3 had the least differences across seasons, both for ΔCO 2 and Δδ 13 C.In summer, C5 exhibited the highest Δδ 13 C value and the lowest ΔCO 2 value, respectively.

Case Study: Combination of Trajectory Clusters and PSCF
The potential source contribution function (see the section "HYSPLIT trajectory model") was applied for the detection of geographical areas with an influence on the measured CO 2 concentration.This was done explicitly for air masses causing extremely high CO 2 concentrations (> 75 th percentile) in Clusters 4 and 5. PSCF maps (Fig. 10) depict that the air masses from C5 can be traced back to the anthropogenic emissions from coal mining areas in Lusatia and from coal districts in East Germany, while air masses from C4 originate from northwestern Europe, including the high emission regions in the Netherlands and the German Ruhr area.A comparable PSCF map for Δδ 13 C (> 75 th percentile), equally inferred by trajectories, describes the spatial distribution of probable geographical source locations for δ 13 C (i.e., sinks of CO 2 ) in summer (Fig. 11).The combination of this PSCF map with MODIS land cover (LC) and NDVI map improve our understanding of the influences of air origin on seasonal variations in CO 2 andδ 13 C. Fig. 11 clearly illustrates that high values of Δδ 13 C are influenced by Western, Central, and Southwestern Poland, which according to LC are predominantly croplands.Fig. 11(c) displays that this area is characterized by NDVI values greater than 0.6, indicating dense vegetation in these area with CO 2 uptake.

CONCLUSIONS
We presented a 2.5-year measurement time series of CO 2 mixing ratios and δ 13 C at the high altitude GAW station UFS on the northern Alpine ridge.Since the Picarro G1101-i instrument was not upgraded to account for water vapor and methane interferences, comprehensive external data corrections and selections had to be implemented.These corrections would not have been feasible had there not been access to parallel measured carbon dioxide, methane and water vapor.Therefore, from the knowledge obtained with CO 2 measurement devices, adding a reliable drying system on the sample inlet line is strongly recommended.
Clear seasonalities in both the CO 2 mixing ratios and δ 13 C were derived with seasonal amplitudes of 15.5 ± 0.15 ppm for CO 2 and of 1.97 ± 0.53‰ for δ 13 C. Seasonal and diurnal variations of both variables showed similar  patterns as at the High Altitude Research Station Jungfraujoch.However, the higher altitude of JFJ led to receiving more frequent free troposphere air masses, and therefore, amplitudes at UFS were generally larger than at JFJ in all seasons.The most pronounced diurnal variabilities were seen in summer at both sites, with an amplitude of 2.9 ppm and 2 ppm for the CO 2 mixing ratio at UFS and JFJ, respectively, and of 0.4‰ and 0.1‰ for the δ 13 C at UFS and JFJ, respectively.The smallest diurnal variabilities occurred in winter, when the peak-to-peak amplitude was 1.4 ppm and 1 ppm for the CO 2 at UFS and JFJ and there was no discernible cycle for δ 13 C.
HYSPLIT classification of the air mass origins at UFS indicated predominant air masses from the west (41.2%), followed by the northwest (19.7%), southwest (14.8%), southeast (12.5%) and northeast (11.8%).The potential source contribution function, using back trajectories as well as atmospheric measurements, provided helpful indications of the origins of air masses potentially influencing a measuring station.So far, this method has been restricted to the identification of sources in air pollution studies (Begum et al., 2005;Pekney et al., 2006;Kaiser et al., 2007;Pongkiatkul and Oanh, 2007;Zhu et al., 2011), as the probability function in PSCF filters out concentrations below a(n) (arbitrary) threshold.However, in this study, δ 13 C measurements and their inverse relationship to CO 2 allowed us to identify atmospheric CO 2 sinks.Among all clusters, high CO 2 mixing ratios in winter (anthropogenic sources) were associated with air masses originating in the Netherlands, the German Ruhr area and Lusatia, whereas high δ 13 C values and low CO 2 mixing ratios in summer, representing a terrestrial biosphere sink influence, mostly originated in Poland.For an improved understanding of the contribution of the sources and sinks of carbon dioxide and the exchange of CO 2 between the terrestrial ecosystem and atmosphere, greenhouse gas emission models (e.g., the WRF Greenhouse Gas Model) can be used to simulate high-resolution transport of carbon dioxide.
Sinks and sources' isotopic signatures of continuously measured CO 2 (such as the data presented in this publication) or records of distinct CO 2 variations (e.g., a sudden enhancement in atmospheric CO 2 ) could be further identified and quantified using the Keeling plot method (Keeling et al., 1989;Vardag et al., 2016).Access to other related tracers, e.g., carbon monoxide and radon-222 measured at the same site (as is done at UFS), may be a great help in this venture (Hirsch, 2007;Tuzson et al., 2011).

Fig. 1 .
Fig. 1.Windrose at UFS, obtained from hourly wind data over the entire study period (May 2012-November 2014).Wind speed and direction are classified into colored segments; gray circles show the cumulated percentage of occurrence.

Fig. 3 .
Fig. 3. Comparison of CO 2 mixing ratio measured by Picarro G1101-i (in red) without drying system to CO 2 measured by Picarro Envirosense 3000i (in black) dried by cold traps.

Fig. 4 .
Fig. 4. Scatterplot of water vapor concentrations and difference of GAW and Picarro CO 2 measurements in the two phases before (a) and after (b) the pump replacement.The dashed line is an ordinary least squares regression line (coefficient of determination of linear regressions in (a): R 2 = 0.97, p-value < 0.001 and (b): R 2 = 0.92, p-value < 0.001).

Fig. 5 .
Fig. 5. Histogram of the deviation between spline and true standard value for (a) Standard 1 and (b) Standard 2.

Fig. 6 .
Fig. 6.Time series of hourly mean values of CO 2 and δ 13 C for the period May 2012 to November 2014.Gray dots and solid lines represent the measurements and fitted background curves, respectively.

Fig. 8 .
Fig. 8. Clustering of trajectories arriving at UFS in the entire period of May 2012 to November 2014 with their percentages of trajectories.

Fig. 9 .
Fig. 9. Boxplots of ΔCO 2 mixing ratio and Δδ 13 C for all seasons classified for the five clusters of Fig. 8 for the entire period.ΔCO 2 and Δδ 13 C are filtered for free troposphere air and background concentrations are subtracted.

Fig. 10 .
Fig. 10.Potential source contribution function plot for ΔCO 2 in wintertime (December to February).(a) PSCF map of Cluster 4 (NW) and (b) PSCF map of Cluster 5 (NE).The position of the site is shown by black circles.

Fig. 11 .
Fig. 11.(a) Potential source contribution function plot for Δδ 13 C in the summertime (June to August) and (b) MODIS Land Cover for the year 2012.Land cover types on this map are grouped and labeled as follow: 1: water; 2: evergreen, deciduous as well as mixed forests; 3: closed and open shrublands, woody savannas, savannas, grassland; 4: croplands; 5: cropland/natural vegetation mosaic and 6: urban and build-up, (c) NDVI map.The position of the UFS site is shown by white circles.