The locations of Anmyeondo (AMY), Jejudo Gosan Suwolbong (JGS), and Ulleungdo (ULD) stations are shown in Fig. 1, and their details are summarized in Table 1.

AMY is located in the west part of South Korea, about 130 km southwest of the megacity of Seoul. Within a 100 km radius, the semiconductor industry and relevant industries exist. Also, the largest thermal power plants fired by coal and heavy oil in South Korea are within 35 km of the northeast and southeast of the station. Local activity is related to agriculture, such as rice paddies, sweet potatoes, and onions, while the area is also known for its leisure opportunities during summer. The west and south side of AMY is open to the sea and along the coast, and there is a large tidal mudflat with many pine trees.

JGS is located in the west part of Jeju Island, which is the biggest volcanic island (1845.88 km2) in the southwest of South Korea and about 90 km from the mainland. Jeju is popular for tourists regardless of the season, while the region of Suwolbong is famous as a Global Geopark due to the outcrops of volcanic deposits exposed along the coastal cliff where JGS is located. In Jeju, there are two major power plants fired by heavy oil at approximately 47 km northeast and 16 km southeast of the stations. The side of the station from the southwest to the northwest is open to the sea, where there are volcanic basalt rocks. The sea to the south is connected to the East China Sea and the sea to the west is linked to the Yellow Sea. Next to JGS there is a wide plain where mainly potatoes, garlic, and onions are harvested.

ULD is located in the east part of Ulleung Island, which is in the east of South Korea and about 155 km from the mainland. In the southeast of the Korean Peninsula, there are cities very famous for steel, chemical, and petrochemical industries along the coastline, and these cities are located about 200–250 km southwest of the island. Ulleung Island is 72 km2 and of volcanic origin, and it is a rocky steep-sided island, with the top of a large stratovolcano reaching a maximum elevation of 984 m. This peak is located northwest of ULD. There are a few small mountains whose heights are about 500 to 960 m a.s.l., within 5 km to the north and southeast of the station. Due to those geological features, ULD is mainly affected by airflow up over the hill from the southwest and downslope winds from the northeast. There is also a small town in the valley northeast of the station with a small port, which is 810 m away from the station. In the southwest area, there is a small brickyard 200 m from the ULD. Farming and fishing industries are very active on the island, though there is no farm in the southern area. An automatic weather station (AWS) was installed at AMY near the inlet, and 10 m above the station at JGS and ULD, but separate from the air inlet tower.

The measurement system consists of three main parts: the inlet, the drying system, and the instruments (Fig. 2). The intake is an inverted stainless steel box (15 cm (W) × 25 cm (D) × 30 cm (H)) with a stainless steel filter (D 4.7 cm, pore size 5 µm) mounted on a plastic mesh holder and connected to the Dekabon sampling tubing (Nitta Moore 1300-10, I.D. 6.8 mm, O.D. 10 mm). Over times longer than 1 month, a significant pressure drop occurs between the inlet and the pump, so the filter is replaced monthly.

Sample air is dried with a system that has a cold trap (CT-90, Operon, South Korea), which is connected to the pump (KNF N145.1.2AN.18, Germany, 55 L min-1, 7 bar in AMY; KNF N035AN.18, Germany, 30 L min-1, 4 bar in JGS and ULD). The cold trap is set to -80 ∘C and maintains its temperature. When the sample air comes from outside into the drying system, the inner temperature increases. Therefore ambient air is cooled down to -20 ∘C in the first chamber, and then to -50 ∘C in the second chamber. To increase dehumidification efficiency, the second chamber is filled with stainless steel beads (Fig. 2).

One of the dual traps is used to dry ambient air for 24 h while the other is warmed and drained. The dehumidification and water drain sequence is as follows: Step 1: pump/cold trap A is employed to dry ambient air for 24 h. Step 2: pump/cold trap B is turned off to melt ice at ambient temperature for 20 h. Step 3: pump B turns on to pressurize and allow water to drain for 2 h. Step 4: cold trap B turns on and cools to operating temperature for 2 h. The difference between this system and a typical cryogenic one is that it was designed with a dual mode, with one trap drying while water is automatically drained from the other. Therefore it avoids the cold trap impinger clogging during long-term, unattended monitoring. This drying system was developed by KMA and the Korea Research Institute of Standards and Science (KRISS) in 2011 for the remote monitoring stations so that it can be easily accessed remotely.

Even though the H2O monitored using CRDS was not calibrated, hourly mean H2O through the drying system is 0.004±0.005 % at AMY, 0.001±0.002 % in JGS, and 0.001±0.004 % in ULD during 2012 to 2016. Laboratory standard gases prepared by the Central Calibration Laboratory (CCL), which is operated by the National Oceanic and Atmospheric Administration, Global Monitoring Division in Boulder, Colorado, USA, typically contain less than 0.0001 % H2O (http://www.esrl.noaa.gov/gmd/ccl/airstandard.html, last access: 1 August 2018). When we sampled them directly using CRDS without this drying system, mean H2O (10 min average) was 0.0009 % regardless of the CO2 level across the KMA monitoring stations.

For example, when there is a difference in H2O at AMY between laboratory standard gases and ambient samples of 0.003 %, this creates a small bias of 0.012 ppm on 400 ppm CO2 according to the equation suggested by Rella et al. (2013): CdilutionCdry=1-0.01Hact, where C is the CO2 mole fraction and Hact is the actual water mole fraction (in %). Since working standards showed almost the same level of H2O to laboratory standards through using CRDS, we considered the CO2 mole fraction dilution offsets between calibration standards and sample air when the uncertainty was estimated (Sect. 3.1).

After the drying system, ambient air flows through the 1/8 in. (O.D.) stainless steel tubing to an 8-port multi-position valve (VICI, EMTMA-CE), which selects among standard gases and ambient air. A leak test of all lines is performed every month. CRDS is well known for its highly linear and stable response (Crosson, 2008). The G2301 model (Picarro, USA) was installed in October 2011, and it became our official CO2 measurement at AMY starting on 1 January 2012. Picarro models G1301 and G2401 have been used to measure ambient CO2 and CH4 since 1 January and 12 February in 2012, at JGS and ULD, respectively. Those analyzers monitor CO2 every 5 s across the KMA greenhouse gas (GHG) network.

At AMY, a non-dispersive infrared analyzer (NDIR; Ultramat 6, Siemens, Germany) was used to monitor atmospheric CO2 every 30 s from 1 February 1999 to 31 December 2011. During the period, we used a three-step dehumidification system, (1) -4∘ cold trap, (2) nafion, and (3) Mg(ClO4)2, before installing the new system.

Variability in CO2 observed at KMA's stations includes contributions from natural atmospheric variability and variability related to the air handling and measurement procedures. Natural atmospheric variability is represented, for example, by the standard deviation of all measurements contributing to a time average, after the contribution of experimental noise is accounted for. Here we develop methods to calculate practical realistic measurement uncertainties. Based on measurements of target cylinders and a co-located comparison of measurements at AMY, we assume systematic biases are negligible. According to the previous studies, the total measurement uncertainty consists of multiple uncertainty components (Andrews et al., 2014, Verhulst et al., 2017). However, in this paper, we assess the measurement uncertainty based on the following components: (UT)2=UH2O2+(Up)2+Ur2+(Uscale)2, where UT is the total measurement uncertainty in the reported dry-air mole fractions, UH2O is the uncertainty from the drying system, Up is repeatability, Ur is reproducibility, and Uscale the uncertainty of propagating the WMO-XCO2 scale to working standard gases.

UH2O is computed from the differences in H2O (%) between the ambient airstream through the drying system and standard gases injected directly, bypassing the drying system. According to the GAW recommendation, the standard gases should be treated through the same system to air sample (WMO, 2016). However, our drying efficiency is not constant, so we injected standard gases directly as a reference value. Here, we define H2O from the standard gases as 0.0009 %. This value was constant and stable during 2012 to 2016. On the other hand, the drying system efficiency is not constant, so this uncertainty component was time dependent. Equation (1) was applied to this factor, where Hact is the difference between H2O in samples and standard gases (0.0009 %). Hourly CO2 dilution offsets range from -0.05 to 0.09 ppm at AMY, -0.02 to 0.07 ppm at JGS, and -0.05 to 0.08 ppm at ULD during 2012 to 2016. Since positive and negative values are found, we use the following equation: Ux=∑i=1N(xi)2N, where Ux represents UH2O, x is the hourly CO2 dilution offsets from Eq. (1), and N is the total number of hourly mean values. UH2O is tabulated for each station in Table 2.

Up is determined from the standard deviations of working standard measurements, as described in Sect. 2.3.1 and expressed by a pooled standard deviation: Up=∑i=1NNi×Si2Ni-Nt, where Si is the standard deviation of 10 min averages of working standard measurements, Ni the number of data during 10 min (based 5 s intervals), and Nt is the total number of calibrations during the period. Si varied from 0.02 to 0.09 ppm at AMY, 0.02 to 0.07 ppm at JGS, and 0.01 to 0.05 at ULD. The pooled standard deviations (Up) are shown in Table 2.

Ur is the drift occurring between 2-weekly calibration episodes, which was mentioned in Sect. 2.3.1. We determined it as the differences in CO2 measured from cylinders with subsequent calibrations over 2 weeks. It ranged from -0.08 to 0.1 ppm at AMY, -0.07 to 0.09 ppm at JGS, and -0.16 to 0.11 ppm at ULD. We expressed Ur as the standard deviation of all drift values during the experimental period using Eq. (), where Ux represents Ur, xΔCO2 during 2 weeks, and N is the total number of data. They are tabulated with other uncertainty terms by site in Table 2.

According to the Zhao and Thans (2006) the uncertainty of working standards can be calculated by the propagation error arising from the uncertainty of primaries with maximum propagation coefficient (γ=1) and repeatability. Similarly Uscale for working standards is determined by Uscale=Up2+Ulab2, where Ulab is the uncertainty of laboratory standards, which CCL (NOAA/ESRL) certified. Here, Ulab has the same value as the uncertainty of secondaries, 0.070 ppm, in the 1σ absolute scale. These values are the same for all stations since they are calibrated by a central lab in AMY. Therefore Up is the repeatability at AMY since we propagate the standard scale through the same analyzer and setup for the atmospheric monitoring.

In the future, quoted uncertainties could be greater due to the inclusion of more error sources. Repeatability and reproducibility may become more precise with improvements in technologies and methods.

The L1 hourly data, L2 daily data, and smoothed curves fitted to L2 daily data are shown in Fig. 3. Episodes of elevated CO2 were often observed at AMY, with a mean difference between maximum and minimum L1 hourly values in a year of ∼102.1±12.1 ppm; for the other sites, maximum minus minimum values were ∼62.5±9.2 ppm at JGS and ∼55.1±9.6 ppm in ULD. The enhancement relative to the local background mole fraction helps evaluate local additions of CO2, with the excess signal defined as CO2XS=CO2OBS-CO2BG, where CO2OBS is L1 hourly data and CO2BG indicates regional background at the site, determined from the smoothed curve fitted to L2 daily data (Sect. 2.3.3). When we roughly analyzed the footprints for hourly CO2XS at three stations, the potential source region was considered not only to be the Korean Peninsula but also northeastern China (KMA, 2014). This happens due to the synoptic system in which the developing low pressure over the source regions causes the pollution to uplift into the free troposphere and makes it descend to the downwind area (Tohjima et al., 2010, 2014; Lee et al., 2016).

Monthly mean CO2XS at AMY was 4.3±3.3 ppm, with 1.7±1.3 ppm at JGS and 1.0±1.9 ppm at ULD during 2012 to 2016. As described in Sect. 2.1, since there are a lot of local activities around AMY, the mean value is larger than at other stations. It was assumed that CO2XS is greater in winter compared to other seasons since photosynthesis is not active and respiration is diminished, while anthropogenic sources such as residential sectors would dominate. However, all three stations showed the highest CO2XS in summer (JJA); it was 6.3±4.9 ppm at AMY, 2.8±1.4 ppm at JGS, and 1.6±2.7 ppm at ULD. Meanwhile the smallest CO2XS was during spring (MAM) at AMY with 2.8±1.5 ppm and during winter (DJF) at JGS and ULD with 0.9±0.5 ppm and 0.4±0.4 ppm, respectively. Even though the selected data, which agree with the conditions given in Sect. 2.3.3, accounted for 55 % to 60 % of total data, the percentages are different according to the seasons. For example, during summer they decreased to 46 % at AMY, 43 % at JGS, and 34 % at ULD; meanwhile they account for 61 %–75 % at all stations during winter. This means that since the Korean Peninsula is affected by the Siberian high from winter to spring with a strong westerly wind, CO2OBS was measured in well-mixed air relative to summer. Also, the wind speed decreased and diurnal variation increased during summer, so CO2OBS might reflect local/regional sources and sink more than other seasons. We also discuss this issue in Sect. 3.3 and 3.4.

To understand the influence of local surface wind on observed CO2, bivariate polar plots were used. These plots are expressed by dependence of all hourly CO2 mole fractions (L1 data) on wind direction and speed in 2016 (Figs. 4–6). The wind data are derived from the AWS which was described in Sect. 2.1.

At AMY, lower CO2 from autumn to winter occurred when winds mainly come from 315 to 360∘. In spring, lower CO2 started to include winds from 180 to 225∘ and the dominant wind direction shifted to the south (180 to 225∘) in summer, indicating that lower CO2 is linked to air masses from the sea (Yellow Sea). However, when wind speed is less than 5 m s-1, CO2 is elevated in all seasons and even in the Yellow Sea. Especially in summer, this condition (wind speed < 5 m s-1) accounts for 80 % of total data, which might enhance CO2XS, as indicated in Sect. 3.2. This also suggests that the high CO2 can be observed in the air mass transported from not only from mainland South Korea but also from west regions from western parts of the Yellow Sea.

JGS observed the strongest winds among the three stations for all seasons, with wind speed > 7 m s-1 occurring almost 36 % of the time and a maximum speed of up to ∼40 m s-1. Lower CO2 was observed with winds from 315 to 340∘ (Yellow Sea) and 120 to 160∘ (East China Sea), with wind speed > 5 m s-1 regardless of seasons. In contrast, JGS is contaminated with local CO2 emissions when wind comes from 45 to 135∘ with wind speed ≤5 m s-1. Since the National Geopark is east of the station, JGS could be affected by tourist activities such as transportation. The station is surrounded by farmlands, so it also could be affected by farming activities such as burning trash and fields. High CO2 was also observed even with strong wind, especially on the side of the Yellow Sea.

For ULD, the main wind directions are quite clearly from 0 to 90∘ (30 %) and from 180 to 270∘ (33 %), and wind speed less than 5 m s-1 occurs 72 % of the time. Normally lower CO2 is monitored regardless of wind direction and wind speed. High CO2 episodes were mainly observed when the wind sector was between 180 to 225∘, presumably affected by the industry complex located in the southeast of the Korean Peninsula and the brickyard 200 m from the station. This wind direction is very dominant in summer with lower wind speed than other seasons.

Overall, both stations on the west side of South Korea, AMY and JGS, might be more affected by continental air mass, so their observations contain information about its sources and sinks, while they are also affected by local activities. Our eastern station, ULD, reflects lower CO2 than other two stations with limited local activities. And it was also suggested that data from regional GAW stations have complex information, so it is necessary to develop a selection method for baseline conditions to better understand regional characteristics.

Diurnal CO2 variations, calculated as the average departure from the daily mean, in April, August, November, and January, are used to represent the average diurnal variations in spring, summer, autumn, and winter over 5 years in Fig. 7. The standard deviations of the hourly means are ∼16, ∼7, and ∼5 ppm in AMY, JGS, and ULD in January, April, and November, but increased in August to ∼20, ∼10, and ∼8 ppm at AMY, JGS, and ULD, respectively.

Prior studies described that diurnal variations can be influenced by the atmospheric rectifier effect that is derived from the covariance between terrestrial ecosystem metabolism, such as the intensity of photosynthesis and density of vegetation, and vertical atmospheric transport (Denning et al, 1999; Chan et al., 2008). Generally, rapid growth of turbulence at the surface after sunrise results in a high boundary layer and leads to decreased CO2 measured at the station during daytime, while CO2 accumulates in a stable nocturnal boundary layer created by a temperature inversion due to surface radiative cooling during the night (Higuchi et al., 2003). Also, the diurnal cycle in summer is the result of a combination of several factors, including active photosynthesis.

AMY and JGS showed those typical characteristics during all seasons, even though the differences between minimum and maximum CO2 values significantly varied by month. However, ULD only showed this trend in summer, while very steady values were shown for other seasons throughout the day.

At AMY, the differences between maximum and minimum values were 13.5 and 6.9 ppm in August and November, respectively, while these values were around 3 ppm in other seasons. This trend is very typical, as mentioned above. For JGS, these values were observed in the order of 9.6 > 3.3 >2.8 > 0.88 ppm in August, April, November, and January, respectively. During summer, both AMY and JGS show an afternoon plateau in CO2 from around mid-afternoon due to the combination of changes in the photosynthetic rate and increased boundary layer before sunset. In the evening CO2 increases again when respiration dominates and the boundary layer becomes neutral or stable. Those two stations also show the clear wind pattern such as land–sea breeze which might enhance the CO2 diurnal cycle in summer. In contrast, at ULD, an average diurnal cycle was only obvious in August (peak to peak value of 3.9 ppm), and CO2 increased monotonically during the afternoon. In other seasons, diurnal variations were 0.5–1 ppm.

For ULD the wind has no diurnal pattern different from the other two stations; however, wind comes from certain sectors regardless of time, which we mentioned in Sects. 2.1 and 3.3. ULD, at 221 m, is higher than AMY and JGS, so it is less affected by local activities. Those geographical characteristics lead to steady values at ULD except for summer when the photosynthesis is most active.

Seasonal variations from KMA's three stations and two other stations, WLG and RYO, in East Asia, are compared in Fig. 8. WLG flask-air data from NOAA/ESRL/GMD and quasi-continuous measurements at RYO by the Japan Meteorological Agency, which were downloaded from the World Data Centre for Greenhouse Gases (WDCGG), were fitted with smoothed curves and compared to KMA observations. It is known that the seasonal cycle of atmospheric CO2 at surface observation stations in the Northern Hemisphere is driven primarily by net ecosystem production fluxes from terrestrial ecosystems (Tucker et al., 1986; Fung et al., 1987; Keeling et al., 1989). The averaged seasonal amplitude from 2012 to 2016 was smallest at WLG with 12.2±0.9 ppm and largest at AMY with 15.4±3.3 ppm. For JGS and RYO, peak to peak amplitudes were similar at 13.2±1.7 ppm and 13.5±1.6 ppm, whereas the peak to peak amplitude was 14.2±3.1 ppm at ULD (Table 3).

Normally, maximum CO2 appears from 4.8 ppm at JGS to 5.8 ppm at AMY in April, while the minimum appears in August between -6.8 ppm at WLG and -9.6 ppm at AMY according to the station. The highest maximum and lowest minimum mean value appeared at AMY, indicating that even though AMY is located at a similar latitude to these other stations, it seems to capture photosynthetic uptake and respiration release of CO2 by terrestrial ecosystems more than others. Also atmospheric CO2 at AMY includes added anthropogenic emissions transported through the Yellow Sea from the Asian continent as explained in Sects. 3.2 and 3.3. Meanwhile WLG is hardly affected by vegetation due to its altitude (Table 1).

The annual growth rate of CO2, which was computed by the increase in annual means of de-seasonal trends from one year to the next at KMA sites, was quite similar to other East Asian stations and to the global growth rate from WMO (Fig. 8b). From 2012 to 2016, the average annual increase observed at all stations in East Asia was between 2.4±0.7 and 2.6±0.9 ppm yr-1. This mean value is similar to the global increase of 2.21 ppm yr-1 from 2007 to 2016 reported by WMO. (This value is determined by the absolute differences from previous year.) The large increase in 2016 and 2015 was due to increased natural emissions of CO2 related to the most recent El Niño event (Betts et al., 2016). Averaged annual CO2 was highest at AMY and lowest at WLG among East Asian stations listed in Table 3, which shows that their differences are 8.5±0.7 ppm. The low growth rate in 2014 at ULD might be caused by unusually low CO2 in July–August 2014, resulting in no significant annual differences between 2013 and 2014, although the reasons are still unclear. Further studies are necessary to fully understand these results.

Since CO2 is a long-lived atmospheric species, the growth rate should be similar between the stations in the same region, even if they are subject to different combinations of anthropogenic and biogenic fluxes. However, our long-term trend comparison showed that measurement and environmental changes also affected its growth rate.

The long-term trends of CO2 mole fractions at AMY, WLG, and RYO from 2002 to 2016, which were extracted using the method of Thoning et al. (1989), are shown in Fig. 9. The trends of CO2 at WLG and RYO increased in parallel, whereas AMY increased with a similar slope but with larger fluctuations than the other stations. Especially the negative growth rate, which was only observed at northern high latitudes in 1992 due to the Mount Pinatubo eruption, was recorded in 2004 and 2006 at AMY, while a high growth rate was recorded in 2012 without the El Niño–Southern Oscillation (Stenchikov et al., 2002; Heimann and Reichetein, 2008; WDCGG, 2017).

In July 2004, the inlet height at AMY was changed from 20 to 40 m above ground (Table 1); observed CO2 mole fractions before moving the inlet height reflected more influence from local activities that affected the long-term trend (Song et al., 2005). According to the logbook, in 2005 AMY was under construction to expand the space with a new building, and the instrument showed strong and highly localized signals during the period.

The measurement system, such as the instruments, the drying systems, and the standard scale, was changed in 2012, as described in Sects. 2.2 and 2.3.1. It was proved that CRDS provides higher precision measurements than the NDIR, and there were CO2 offsets in a comparison between the two instruments (Chen et al., 2010; Zellweger et al., 2016). Maintaining the traceability of standard gas to the primary standard with the same scale under the GAW Programme would be more of an incentive to assure the long-term consistency (WMO, 2017). This result suggests that factors not only related to local sources and sinks, but that environmental changes around stations and level of technical skill are also very important for the monitoring of regional background CO2 in the long term. On the other hand, ongoing comparisons of measurements at co-located sites and for the same species, such as between discrete samples and continuous measurement (Masarie et al., 2001), are valuable means to maintain data quality and identify sampling issues rapidly. After 2012, long-term trends increased in parallel, with AMY 5.5±0.3 ppm greater than RYO, and RYO 2.9±0.3 ppm greater than WLG.