Samuel Takele KENEA,Young-Suk OH,Jae-Sang RHEE,Tae-Young GOO,Young-Hwa BYUN,Shanlan LI,Lev D.LABZOVSKII,Haeyoung LEE,and Robert F.BANKS

1Climate Research Division,National Institute of Meteorological Sciences,33,Seohobuk-ro,Seogwipo-si,Jeju-do,63568,Republic of Korea

2Environmental Meteorology Research Division,National Institute of Meteorological Sciences,33,Seohobuk-ro,Seogwipo-si,Jeju-do,63568,Republic of Korea

3Meteorology Department,Delta Air Lines,Inc.,Atlanta GA,United States

ABSTRACT

Key words: model evaluation,in-situ observations,CarbonTracker,East Asia

1. Introduction

Atmospheric carbon dioxide (CO2) is a key greenhouse gas that causes global warming (IPCC, 2013). The increase of the CO2concentration, on average 2 ppm yr-1globally,is signif icantly related with contribution from human activities. In fact, the relative contribution of CO2to the atmosphere varies greatly with region. East Asia is an important region that emits a large amount of CO2into the atmosphere(Tian et al., 2016). On the other hand, the sink of CO2is predominantly controlled by terrestrial vegetation uptake via photosynthesis and oceanic uptake (Keeling et al., 1989;Hansen et al.,2007;Watson et al.,2011).

Accurate estimates of CO2sources and sinks are of great importance for validating carbon emissions reduction eあorts and reducing the uncertainties of carbon cycle—climate feedbacks.Eあorts have been undertaken to reduce the uncertainty of model estimations of f luxes through the assimilation of accurate observations of CO2data from the global network(Tans et al., 1990; Gurney et al., 2002). However, as previous studies point out, there are uncertainties in terms of the localization of sources and sinks on regional scales(e.g.Baker et al., 2006). Owing to less spatial coverage of accurate observations of CO2over East Asia, the regional CO2f luxes estimated from atmospheric inversions are still uncertain(e.g.,Swathi et al.,2013).In addition,uncertainties in the planetary boundary layer height, or in horizontal winds, affect the modeled near-surface CO2concentrations(Gurney et al.,2002;Lin and Gerbig,2005;Gerbig et al.,2008;Prather et al., 2008; Ahmadov et al., 2009). Therefore, continuous assessment and evaluation of model simulations of CO2concentrations against accurate in-situ observations is vital.

Recently, the nested-grid CarbonTracker (CT)-Asia model was run using versions CT2013B(Cheng et al.,2013)and CT2016 (documented at http://carbontracker.noaa.gov,Peters et al.,2007)by the National Institute of Meteorological Sciences(NIMS),Republic of Korea. In this simulation,in-situ continuous hourly CO2concentrations(including daytime and nighttime) at Ryori, Minamitorishima, Yonagunijima,and Tae-ahn Peninsula,which are close to Anmyeondo station, and CONTRAIL (Comprehensive Observation Network for Trace gases by Airliner)data were assimilated. This result could be used as an indicator for where the model needs improvement in order to estimate accurate f luxes through accurate estimates of f lux error in the assimilation process.

In the present study, we investigated the spatiotemporal distribution of simulated near-surface CO2concentrations along with CO2f luxes from fossil fuel emissions and the biosphere. Then, we statistically evaluated the model's performance in reproducing diurnal and seasonal variations of nearsurface CO2concentrations through comparison with in-situ observations over target stations. We also examined the wind speed and directions over the target stations during daytime and nighttime for both the winter and summer seasons.

2. Sites

The stations used to evaluate the model's performance in simulating near-surface CO2concentrations were as follows(suきcient information on the stations is provided in the literature, some examples of which are cited): the coastal stations of Anmyeondo, Gosan, and Ryori (Sasaki, 2006); the remote stations of Yonagunijima(Fukuyama,2013)and Minamitorishima;the mountain sites of Mt. Waliguan(Zhou et al., 2004, 2006) and Lulin (Qu et al., 2013); and the inland stations of Kisai and Shangdianzi(Cheng et al.,2018).These sites are located under diあerent vegetation types,climate features, and economic zones(e.g., Fang et al.,2014; Cheng et al.,2018).Figure 1 displays a map of the stations overlaid on a spatial plot of the terrain height.

3. Data and methods

3.1. CT inverse model

CT is an inverse atmospheric model developed by the National Oceanic and Atmospheric Administration Earth System Research Laboratory Global Monitoring Division(http://www.esrl.noaa.gov/gmd/ccgg/carbontracker). Here,we adopt a nested-grid CT inverse model with a horizontal resolution of 1°×1°for the simulation of atmospheric CO2concentrations over East Asia. Model simulations were run by NIMS based on the CT2016 version. The model uses Transport Model 5(Krol et al.,2005),forced by meteorological f ields from ERA-Interim (Dee et al., 2011). The model gives four components of CO2signals, which respectively derive from fossil fuel emissions, air—sea CO2exchange,and terrestrial f luxes from wildf ire emissions and non-f ire net ecosystem exchange. The model uses the a priori information for surface CO2f lux data for each module. The biosphere CO2f lux is supplied by a biosphere model,CASA(the Carnegie—Ames Stanford Approach).In the f ire module,CO2released by f ire is taken from the Global Fire Emission Database,version 4.1s,at a three-hourly temporal resolution.We used two diあerent fossil fuel CO2emissions datasets—namely, “Miller” and ODIAC (Open Source Data Inventory for Anthropogenic CO2)—which were used to help assess the uncertainty in the mapping process.In the ocean module,prior estimates of air—sea CO2f lux were determined from the Ocean Inversion Fluxes scheme,and the updated version of the Takahashi et al. (2009) pCO2 climatology. Further details are provided in the CT document (CT2016 release,https:/www.esrl.noaa.gov/gmd/ccgg/carbontracker/CT2016 doc.php). While obtaining CO2signals from oceanic and terrestrial biospheric surface f luxes, we used data from Taeahn Peninsula, Minamitorishima, Yonagunjima, and Ryori for optimizing those f luxes in the assimilation process. The model uses an ensemble Kalman f ilter to estimate the surface CO2f lux with atmospheric CO2measurements as a constraint(Peters et al.,2007,2010).

3.2. In-situ measurements

In-situ measurements of CO2concentrations were taken using non-dispersive infrared(NDIR)absorption sensors and cavity ring-down analyzers (CRDS). Measurements derived from both instruments were used for evaluating the model's ability to simulate diurnal and seasonal variations of nearsurface CO2concentrations. The accuracy of CO2measurements from those systems is typically better than 0.1 ppm(Andrews et al.,2014).All sites used in this study only have one air intake height.More information concerning the sites can be found at http://www.esrl.noaa.gov/gmd/ccgg/insitu/.The in-situ data were obtained from the World Data Centre for Greenhouse Gases (WDCGG) (https://ds.data.jma.go.jp/gmd/wdcgg/cgi-bin/wdcgg/catalogue.cgi), which operates under the framework of the WMO's Global Atmosphere Watch. Simultaneous measurements of meteorological parameters, such as atmospheric temperature, wind speed and direction, and relative humidity, were also provided by the WDCGG.In this work, surface sampling stations were chosen at inland,coastal,mountainous,and remote sites(see Fig.1 and Tables 1 and 2 for more information).

3.3. Comparison method

Fig.1.Terrain height(units: km)with an overlay map of the network of in-situ measurement stations,where the three-letter abbreviations mean the following:WAL,Mt.Waliguan;SDZ,Shangdianzi;AMY,Anmyeondo;RYO,Ryori;KIS,Kisai; GSN, Gosan; YON, Yonagunijima; LUL, Lulin. The data are obtained from CT model.

Table 1.Details of the instruments at selected stations.

To evaluate the performance of the model's ability in simulating CO2concentrations through comparison with in-situ observations, we f irst applied temporal and spatial coincidence criteria over all selected stations.Model outputs were sampled at the nearest grid point and the model vertical level that corresponded to the in-situ inlet height of the stations.In fact, there were other methods/techniques that we could have applied for the spatial coincidence criteria, such as the nearest grid point method, interpolation with concentration slope or linear interpolation, and grid-averaging, but they were found to aあect the correlations between the simulated and observed results, particularly at coastal and inland stations. This f inding is consistent with Patra et al. (2008). In evaluating the model's performance at seasonal time scales,we considered daytime(1330—1630 LST),nighttime,and allhourly averaged data. In addition, we investigated the amplitude and phase of the seasonal cycle, and quantitatively determined the Pearson's correlation coeきcient(R)between simulations and observations. Linear regression analysis was applied to examine how the data were spread over the linear f itting. We also assessed the model's ability to reproduce the diurnal variations of the observations by comparing the phase and amplitude of the diurnal cycle.The bias was expressed as the diあerence between simulation and observation.

4. Results and discussion

We compared the downscaled CT model-simulated atmospheric CO2concentration based on CT2016 versions with in-situ observations over East Asia. Here, we discuss the comparison of CO2at diurnal and seasonal time scales.

4.1. Spatiotemporal distribution of CO2 concentrations

Here, we highlight the spatiotemporal variations of the simulated near-surface CO2concentrations and f luxes during the period 2009—13. The distribution of the multi-year seasonal mean near-surface CO2concentrations in East Asia shows signif icant spatial heterogeneity. The spatial and temporal variations of near-surface CO2concentrations were predominantly driven by the anthropogenic emissions, and by the variations of the biospheric CO2, resulting from the seasonal phenomena of growth and decay of land vegetation,as well as atmospheric transport.The surface-level simulated CO2concentration was maximum in winter and minimum in summer. Figure 2 displays the simulated near-surface atmospheric CO2concentrations for each season over East Asia in the period 2009—13, while Fig. 3 shows the CO2f luxes derived from fossil fuels and biogenic emissions in diあerent seasons.The seasonal spatial distribution of the simulated CO2concentrations had a consistent pattern with that of the spatial distribution of CO2f luxes from fossil fuel emissions and the biosphere over East Asia (20°—51°N, 90°—150°E).The hot spots of higher fossil-fuel f luxes (see Fig. 3, top panel)were located over a region encompassing the megacities of Korea,Japan,and eastern China,which is a recognized region of high CO2emissions(e.g.,Ballav et al.,2012;Shim et al.,2013).

4.2. Diurnal cycle of CO2

Validating the amplitude of the diurnal variability of the model simulation through comparison with in-situ nearsurface observations is vital,as the diurnal variability of nearsurface CO2represents the sources, sinks, and related surface processes (e.g., Bakwin et al., 1998). The phenomena of photosynthesis and respiration,boundary layer dynamics,and pollution transport are key factors for the observed diurnal cycles of CO2concentrations near the surface (Bakwin et al., 1998). Law et al. (2008) examined model simulations of CO2concentrations along with observational data and found that the diurnal amplitude errors were contributed by sampling choice in the vertical and horizontal directions,the model resolutions,and the land surface f lux.

Figures 4 and 5 provide a comparison of the in-situ and model-simulated diurnal variations of atmospheric CO2concentrations during winter and summer over the course of the study period for the following four stations: Anmyeondo,Gosan, Ryori, and Kisai. The overall patterns of the diurnal cycles of the simulated CO2concentrations followed the observations during daytime,but with large discrepancy during nighttime. Pronounced diurnal cycles of CO2were present in summertime,with peaks at night and troughs in the afternoon.The bias,as shown in Table 2,was less than 6.3 ppm on the all-hourly mean basis,and this bias was further reduced to a maximum of 4.6 ppm(see Table 3)when considering only daytime.

Looking specif ically at Anmyeondo and Gosan stations,the model result agreed well with observations in that both showed no distinct diurnal cycle. At Anmyeondo, there was an estimated small positive bias of the model simulation(0.2 ppm)against the in-situ observations during winter.In summertime, the bias was 3.6 ppm. Representation error is subject to f lux gradients and the direction of land or sea breezes(Tolk et al.,2008).Although the f lux gradient was high over Anmyeondo in winter(see Fig.3),a small bias was estimated.As evident in the wind rose plot in Fig.S1 in the Electronic Supplementary Material(ESM),there was a strong inf luence of sea breezes bringing ocean airmasses containing depleted CO2.We can infer that,to a large extent,such a phenomenon was captured well by the model.

Fig.2. Model-simulated seasonal mean near-surface CO2 concentrations (units: ppm) during 2009—13. Plus signs denote the in-situ observation stations used in the study; DJF, MAM, JJA and SON denote the winter(December—February),spring(March—May),summer(June—August)and autumn(September—November)seasons,respectively.

Fig.3.Model-simulated seasonal mean CO2 f lux(units: μmol m-2 s-1)(fossil fuels f lux in the top panel and biospheric f lux in the bottom four panels)during the period 2009—13.DJF,MAM,JJA and SON denote the winter (December—February), spring (March—May), summer (June—August) and autumn (September—November)seasons,respectively.

Fig.4.Mean diurnal cycle of CO2 in winter during 2009—13(except for Gosan,which is in the period 2009—11).The blue curve shows the observed result and the black curve the model result. The 1-σ standard deviation is shown by the blue shading for the observations and by the black vertical lines for the model outputs. The red line represents the model-simulated CO2 f lux. Note: the scale of the y-axis is diあerent.

Fig.5.As in Fig.4 but for the summer season.

Table 2.Mean diurnal CO2 concentrations(units: ppm)of the CTAsia model, in-situ observations, and bias during winter and summer in the study period.

Table 3. Daytime (1330—1630 LST) mean CO2 concentrations(units: ppm) of the CT-Asia model, in-situ observations, and bias during winter and summer in the study period.

On the other hand, the relative inf luence of local and regional emissions might have been strong in summer,since the prevailing wind directions were southwest and northwest,which brought airmasses rich with CO2from the land (see right-hand panels of Fig. S1 in the ESM). Furthermore, the standard deviation of the in-situ observations was higher than the model results,which ref lects the inf luence of local emissions and sinks, typically playing a more important role in observed CO2concentrations at low wind speeds (Figs. S1,S2, and S3 in the ESM). Moreover, there was a strong terrestrial biospheric f lux gradient around the station,where the signal was high compared to the corresponding nearest grid point.The bias was estimated to be 3.6 ppm in this particular season,as shown in Table 2.

Regarding Gosan station, located at the tip of the west coast of Jeju Island,it is too small to be captured by the resolution of the regional inverse model. This could also be a factor contributing to the diあerences between the simulation and observation.In summer,the in-situ observations depicted a large amplitude (peak-to-peak CO2concentrations) on the order of 8.3 ppm, while the model only produced an amplitude of 3.3 ppm, which was less than half that of the in-situ observations(see Fig.5b).

Looking at Ryori and Kisai, the model exhibited a pronounced diurnal cycle in both winter and summer.The model exhibited good agreement in the afternoon at both stations,but a large discrepancy occurred at nighttime. This indicates that the model failed to simulate nocturnal CO2accumulations in both seasons. We can see almost comparable magnitudes of total simulated CO2f lux(Figs.5c and d)between 0000 and 0500 LST during summer at these stations. However,Kisai had a relatively elevated CO2concentration compared to Ryori.This might have been caused by underestimation of the nighttime boundary layer height and/or advection of CO2containing airmasses within this layer (Figs.S4 and S5 in the ESM). Some studies have examined the impact of planetary boundary layer height uncertainties on the modeled near-surface CO2concentrations during daytime, and estimated a bias of ～3 ppm(e.g.,Kretschmer et al.,2012,2014),while uncertainties in the horizontal winds can induce a total CO2transport uncertainty of ～6 ppm (Lin and Gerbig,2005). At night, the representation error due to unresolved topography will be amplif ied within the nocturnal boundary layer(Tolk et al.,2008).

4.3. Seasonal variations of CO2

Here, the ability of the model to re produce the seasonal variations of CO2is evaluated through comparison with the in-situ observations at selected stations. The seasonal cycle amplitude,phase and bias were investigated. Examining the time oあset of the seasonal cycle has important implications for f lux estimates(Keppel-Aleks et al.,2012).

Figure 6 compares the seasonal cycle of the simulated CO2with the in-situ observations at selected stations. Daytime, nighttime, and all-hourly averaged data were considered separately. The in-situ data from Ryori, Yonagunijima,and Minamitorishima were assimilated in the CT2016 version, while the rest of the sites were independent of the CT2016 data assimilation system. Overall,the results for the seasonal cycle of the simulated CO2concentrations broadly agreed with the observations,with concentrations peaking in April and a minimum occurring in August/September. This result was found for all stations, when considering the daytime data.However,large discrepancies were identif ied when applying the nighttime data,in particular at Ryori,Kisai,and Shangdianzi. At Kisai and Shangdianzi, we quantif ied how well the model reproduced the seasonal variations when all daytime and nighttime data were incorporated; the correlation coeきcients were 0.55 and 0.26(Table 4 and Fig.7),respectively. However, the result improved when considering only the daytime data, with correlation coeきcients of about 0.84 and 0.71,and the biases were estimated to be-1.6±3.2 and 5.3±5.7(1-σ)ppm(see Table 5 and Fig.6).This suggests that the inclusion of nighttime and early morning data aあected the seasonal comparison.

Despite the overall agreement between the simulation and observations at Anmyeondo,it is evident that the model slightly overestimated the CO2concentrations during summertime.As noted in section 4.2,the diurnal variations of the simulated CO2concentrations at Anmyeondo during summer were estimated to be higher than the in-situ measurements. A mean bias of 1.2 ppm was found,with a corresponding standard deviation of 4.0(1-σ)ppm.The mismatch in the CO2seasonal amplitude indicates that the simulated CO2surface f luxes cannot capture the peak of the terrestrial carbon exchange(Yang et al.,2007).

Fig.6.Monthly time series of mean CO2 concentrations for daytime, nighttime, and both day-and nighttime, during 2009—13 (except for Gosan, which is during 2009—11, and Shangdianzi, which is during 2010—13), as observed over selected sites in East Asia. Note that the in-situ monthly mean time series are not depicted specif ically for daytime and nighttime at Lulin,Mt. Waliguan,and Shangdianzi.

Fig.7.Model-simulated versus in-situ observed monthly mean(includes all hourly values)CO2 concentrations during 2009—13(except for Gosan,which is during 2009—11,and Shangdianzi,which is during 2010—13),as observed at nine selected sites in East Asia. The blue dashed line is the 1:1 f itting line. See Table 4 for the statistical results.

Over the mountain stations,the seasonal cycles(Figs.6g and h) indicated that the model overestimated the observations, with a pronounced phase diあerence. When applying the comparison method, the f irst vertical level in the model did not match with the in-situ inlet height,because the model resolution could not resolve the topography of such complex terrain accurately (see Fig. 1). Consequently, we selected the model vertical level that approximately represented the in-situ inlet height above sea level (i.e. the f ifth vertical level for Mt. Waliguan and the eighth for Lulin). The slightmismatch in the model sampling vertical level with the insitu inlet height may have had an impact on the phase diあerence and bias in the seasonal cycle, probably because of the timing diあerence in the transport process. For example, at Mt. Waliguan,the peak often occurred earlier than observed,whereas the minimum occurred later. Fang et al. (2014) noticed the occurrence of seasonal CO2maximum periods f luctuate considerably, ranging from December at Longfengshan and Lin'an,to March for Shangdianzi and May for Mt.Waliguan. This diあerence is believed to be driven not only by regionally diあerent terrestrial ecosystems and human activities,but also by local meteorological conditions(Zhang et al.,2008).At Mt. Waliguan and Lulin,we found a good level of agreement in capturing the variability of the observations,with correlation coeきcients of 0.89 and 0.87, respectively.Both stations exhibited positive biases,with estimated values of 3.7 (Mt. Waliguan) and 3.9 ppm (Lulin). At remote stations like Minamitorishima and Yonagunijima, a very good level agreement was obtained, with a correlation coeきcient of 0.99 and an RMSE less than 0.79.

Table 4.Statistical comparison of atmospheric CO2 concentrations(units:ppm)between the in-situ observations and CT-Asia model on a monthly mean basis(includes all hourly values)during the study period.

Table 5.As in Table 4 but for daytime(1330—1630 LST).

Fig.8. Correlation coeきcient (model versus in-situ observations) for the CO2 concentration in each season at Anmyeondo, Ryori, Kisai, Lulin, and Mt.Waliguan during 2009—13.

We also determined the correlation coeきcients for Anmyeondo, Ryori, Kisai, Lulin, and Mt. Waliguan, for each season(Fig.8). Kisai revealed a poor correlation(R=0.29)in summer, whereas a good correlation was found in winter(0.84). Ryori showed similar results, with a correlation of 0.52 in summer and 0.96 in winter. This suggests that the model's performance in capturing the monthly variations for each season varies, with better representation of winter than other seasons. The model's ability in capturing the eあect of terrestrial vegetation,particularly in the peak growing season,needs further improvement.

5. Conclusions

The performance of a nested-grid CT model in simulating atmospheric CO2concentrations was assessed through comparison with in-situ observations over nine selected stations in East Asia during the period 2009—13.The evaluation was conducted in terms of diurnal and seasonal variations,in which the amplitude,phase diあerences and bias were examined. The diurnal cycles of terrestrial biospheric f luxes and the planetary boundary layer were the most likely deriving factors for the variations of surface level CO2concentrations.

Large discrepancies existed with regard to the diurnal cycle at night when comparing the observations and model results in winter and summer,as inferred from data at four stations(Anmyeondo,Gosan,Ryori,and Kisai).In general,the model's ability to reproduce the CO2diurnal cycle remains challenging. On the other hand, the model exhibited a very good level of agreement with observations in daytime at those target stations. Overall,biases were less than 6.3 ppm on an all-hourly mean basis,and this was further reduced to a maximum of 4.6 ppm when considering only daytime. For example, at Anmyeondo, a small bias was obtained in winter, on the order of 0.2 ppm, whereas in summertime the bias was higher.The observed large discrepancy at nighttime might be reduced by increasing the horizontal resolution and vertical levels.

In terms of seasonal variation,the level of agreement was judged by considering daytime,nighttime,and all-hourly averaged data. The overall performance of the model in reproducing the observed seasonal variations of CO2all-hourly averaged basis was good at most sites,with the exception of Ryori, Kisai, and Shangdianzi stations. The level of agreement was further improved at almost all stations when considering the comparison during daytime, with a correlation coeff icient ranging from 0.70 to 0.99; however, the model performed poorly in capturing the peak drawdown of CO2during the summer season at Shangdianzi. Choosing the model sampling level that corresponded to the in-situ inlet height may have led to a small overestimation or underestimation of CO2; however, this particular eあect was observed mainly over mountainous areas.

The f indings of this study highlight the strengths and weaknesses of the model in reproducing diurnal and seasonal variations of near-surface CO2concentrations. We recommend that continued eあorts are made to evaluate the model's performance, particularly in capturing nighttime observations. Also, a series of model experiments are required in the future in order to explicitly quantify the biases in the simulated near-surface CO2concentrations on the diurnal time scale in relation to meteorological parameters(e.g.,boundary layer height,wind speed and direction,relative humidity).

Acknowledgements.This work was supported by the Korea Meteorological Administration Research and Development Program “Research and Development for KMA Weather, and Earth system Services-Development and Assessment of AR6 Climate Change Scenarios” under Grant (KMA2018-00321). We gratefully acknowledge those who provided the access to the in-situ data at the WDCGG (https://ds.data.jma.go.jp/gmd/wdcgg/cgi-bin/wdcgg/catalogue.cgi).

Electronic supplementary material.Supplementary material is available in the online version of this article at https://doi.org/10.1007/s00376-019-8150-x.