Estimating Asian terrestrial carbon fluxes from CONTRAIL aircraft and surface CO 2 observations for the period 2006 to 2010

H. F. Zhang, B. Z. Chen, I. T. van der Laan−Luijkx, T. Machida, H. Matsueda, Y. Sawa, Y. Fukuyama, R. Langenfelds, M. van der 5 Schoot, G. Xu, J. W. Yan, M.L. Cheng, L. X. Zhou, P. P. Tans, W. Peters 1 State Key Laboratory of Resources and Environment Information System, Institute of geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China 10 2 University of Chinese Academy of Sciences, Beijing 100049, China 3 Department of Meteorology and Air Quality (MAQ), Wageningen University, Droevendaalsesteeg 3a, NL−6700 PB, Wageningen, The Netherlands 4 Center for Global Environmental Research, National Institute for Environmental Studies, Tsukuba, Japan 15 5 Geochemical Research Department, Meteorological Research Institute, Tsukuba, Japan 6 Atmospheric Environment Division, Global Environment and Marine Department, Japan Meteorological Agency 7 Centre for Australian Weather and Climate Research/CSIRO Marine and 20 Atmospheric Research, Aspendale, Victoria, Australia 8 Key Laboratory for Atmospheric Chemistry of China Meteorological Administration, Research Institute of Atmospheric Composition of Chinese Academy of Meteorological Sciences, Beijing 100081, China 9 Earth System Research Laboratory, National Oceanographic and Atmospheric 25 Administration, Boulder, Colorado 80305, USA 10 Centre for Isotope Research, Groningen, The Netherlands


Introduction
The concentration of carbon dioxide (CO 2 ) has been increasing steadily in the atmosphere since the industrial revolution, which is considered very likely to be responsible for the largest contribution of the climate warming (Huber and Knutti, 2011;Peters et al., 2011). Knowledge of the terrestrial carbon sources and sinks is critically important for understanding and projecting the future atmospheric CO 2 levels and climate change. The global terrestrial ecosystems absorbed about 1-3 Pg carbon every year during the 2000s, with obvious interannual variations, offsetting 10-40 % of the anthropogenic emissions (Le Quéré et al., 2009;Maki et al., 2010;Saeki et al., 2013). However, estimates of the terrestrial carbon balance vary considerably when considering continental scales and smaller, as well as when estimating the CO 2 seasonal and inter-annual variability (Houghton, 2007;Peylin et al., 2013).
Asia, as one of the biggest Northern Hemisphere terrestrial carbon sinks, has a significant impact on the global carbon budget (Jiang et al., 2013;Patra et al., 2012;Piao et al., 2009Piao et al., , 2012Peylin et al., 2013;Yu et al., 2013). It is estimated that Asian ecosystems contribute over 50 % of the global net terrestrial ecosystem exchange (Maksyutov et al., 2003) and their future balance is thought to be a great source of uncertainty in the global carbon budget (Ichii et al., 2013;Oikawa and Ito, 2001). Even though the importance of the Asian ecosystems is increasingly recognized and many efforts have been carried out to estimate the Asian terrestrial carbon sources and sinks, they still remain poorly quantified (Ito, 2008;Patra et al., 2012Patra et al., , 2013Piao et al., 2011). One reason is that a steep rise of fossil fuel emissions in most Asian countries has imposed large influences on the Asian CO 2 balance and leads to an increased variability of the regional carbon cycle (Francey et al., 2013;Le Quere et al., 2009;Patra et al., 2011Patra et al., , 2013Raupach et al., 2007). In addition, rapid land-use change and climate change have likely increased the variability in the Asian terrestrial carbon balance (Cao et al., 2003;Patra et al., 2011;Yu et al., 2013). This makes it challenging to accurately estimate CO 2 fluxes of the Asia ecosystems.
Currently two approaches are commonly used to estimate CO 2 fluxes at regional to global scales: the so-called "bottom-up" and "top-down" methods. The bottom-up approach is based on local data or field measurements to retrieve the carbon fluxes, including direct measurements (Chen et al., 2012;Clark et al., 2001;Fang et al., 2001;Mi-zoguchi et al., 2009;Takahashi et al., 1999) and ecosystem modeling Fan et al., 2012;Randerson et al., 1997;Sellers et al., 1986Sellers et al., , 1996. The top-down method uses atmospheric mole fraction data to derive the CO 2 sink/source information. As one of the important "top-down" approaches, atmospheric inverse modeling has been well developed and widely applied (Baker et al., 2006;Chevallier and O'Dell, 2013;Deng et al., 2007;Gurney et al., 2003;Gurney et al., 2004), and has shown to be particularly successful in estimating regional carbon flux for regions rich in atmospheric CO 2 observations like North America and Europe (Broquet et al., 2013;Deng et al., 2007;Peters et al., 2007Peters et al., , 2010Peylin et al., 2005Peylin et al., , 2013Rivier et al., 2011Rivier et al., , 2010. However, estimating Asian CO 2 surface fluxes with inverse modeling remains challenging, and the inverted Asian CO 2 fluxes still exhibit a large uncertainty partly because of a lack of surface CO 2 observations. For example, in the TransCom3 annual mean control inversion, Gurney et al. (2003) used a set of 17 models to estimate the carbon fluxes and obtained different results for the Asian biospheric CO 2 budget, ranging from a large CO 2 source of +1.00 ± 0.61 Pg C yr −1 to a large sink of −1.50 ± 0.67 Pg C yr −1 for the year 1992-1996. In the REC-CAP (REgional Carbon Cycle Assessment and Processes) project, Piao et al. (2012) presented the carbon balance of terrestrial ecosystems in East Asia from eight inversions during the period 1990-2009. The results from these eight inversion models also show disagreement. Six models estimate a net CO 2 uptake with the highest net carbon sink of −0.997 Pg C yr −1 , while two models show a net CO 2 source with the largest net carbon emission of +0.416 Pg C yr −1 in East Asia. The important role of the sparse observational network was demonstrated by Maki et al. (2010), who reported a large Asian land sink of −1.17 ± 0.50 Pg C yr −1 or much smaller sink of −0.65 ± 0.49 Pg C yr −1 over the Asian region depending on which set of observations was included in the same inversion system. This situation suggests that a more accurate estimate of the surface CO 2 flux is urgently required in Asia, and the ability to base it on as much observational data as possible is key.
To expand the number of CO 2 observations, the aircraft project CONTRAIL has measured CO 2 mole fractions onboard passenger flights since 2005, and has produced a large coverage of in situ CO 2 data ranging over various latitudes, longitudes, and altitudes . CONTRAIL observations have also already successfully been used to constrain surface flux estimates (Niwa et al., , 2012Patra et al., 2011). Patra et al. (2011) reported the added value of CONTRAIL data to inform on tropical Asian carbon fluxes, as their signals are transported rapidly to the free troposphere over the west Pacific.
In this study, we also used the CONTRAIL CO 2 observations (http://www.cger.nies.go.jp/contrail/) together with a global network of surface observations to estimate the Asian weekly net ecosystem exchange of CO 2 (NEE) during the  (Gurney et al., 2002;Gurney et al., 2003). These divided regions are presented in the small inset in the bottom left corner (same as thereafter).
period 2006-2010. Our inversion model is the state-of-theart CO 2 data assimilation system CTDAS (CarbonTracker Data Assimilation Shell, http://carbontracker.eu/ctdas/). Our work complements previous inverse modeling studies as it (1) presents the inverted CO 2 results of Asian weekly net ecosystem exchange not shown previously; (2) uses surface observations not available in earlier top-down estimates; (3) assimilates the continuous CO 2 observation from a number of Asian continental sites for the first time; (4) includes extra free tropospheric CO 2 observations to further constrain the estimate; (5) uses a two-way atmospheric transport model TM5 (Krol et al., 2005) with higher horizontal resolution than previous global CO 2 data assimilation studies that focused on Asia (this study uses a 1 • × 1 • grid over Asia while globally a 2 × 3 • resolution, see Fig. 1b).
This paper is organized as follows. Methods and materials are described in Sect. 2, the inferred Asian land flux and its temporal-spatial variations are presented in Sect. 3. To examine the impact of CONTRAIL data on Asian flux estimates, we also compared inverse results with and without CONTRAIL data during the period 2006-2010. In Sect. 4, we compare our inverted Asian surface fluxes with previous findings and discuss our uncertainty estimates and future directions. Note that "Asia" refers to lands as far west as the Urals, and it is further divided into boreal Eurasia, temperate Eurasia and tropical Asia based on TransCom regions (Gurney et al., 2002(Gurney et al., , 2003) (see small inset in the bottom left corner of Fig. 1).

The atmospheric inversion model (CTDAS)
The atmospheric inverse model CTDAS was developed by NOAA-ESRL (National Oceanic and Atmospheric Administration's Earth System Research Laboratory) and Wageningen University, the Netherlands. Previous versions of the system have been applied successfully in North America and Europe (Masarie et al., 2011;Peters et al., 2007Peters et al., , 2010. CTDAS was designed to estimate net CO 2 terrestrial and oceanic surface fluxes by integrating atmospheric CO 2 concentration measurements, a global transport model, and a Bayesian synthesis technique that minimizes the difference between the simulated and observed CO 2 concentrations. The first step is the forecast of the atmospheric CO 2 concentrations using the transport model TM5 (Krol et al., 2005) with a global resolution of 3 • × 2 • and 1 • × 1 • over Asia (Fig. 1b). The TM5 transport model is driven by meteorological data of the ERA-interim analysis of the European Centre for Medium-Range Weather Forecasts (ECMWF), and propagates four separate sets of bottom-up fluxes (details are presented in Sect. 2.2). The forecasted four-dimensional (4-D) concentrations (x, y, z, t) are sampled at the location and time of the observed atmospheric CO 2 mole fractions, and subsequently compared. The difference between the observed and simulated CO 2 concentrations is minimized. This minimization of the mole fraction differences in CTDAS is done by tuning a set of linear scaling factors which are applied to find the set of sources and sinks that most closely match the observed CO 2 concentration in the atmosphere. As described in Peters et al. (2007), four a priori and imposed CO 2 fluxes integrate in CTDAS to instantaneous CO 2 fluxes F (x, y, t) as follows: F (x, y, t) = λ r F bio (x, y, t) + λ r F oce (x, y, t) where F bio and F oce are 3-hourly, 1 • × 1 • a priori terrestrial biosphere and ocean fluxes, respectively; F ff and F fire are monthly 1 • × 1 • prescribed fossil fuel and fire emissions, and λ r is a set of weekly scaling factors, and each scaling factor is associated with a particular region of the global domain that is divided into 11 land and 30 ocean regions according to climate zone and continent. Nineteen ecosystem types (Olson et al., 1985) (Fig. 1a) have been considered in each of the 11 global land areas (Gurney et al., 2002), dividing the globe into 239 regions (239 = 11 land × 19 ecosystem types + 30 ocean regions). The actual region number assimilated in this system is 156, after excluding 83 regions which are associated with a non-existing ecosystem (such as "snowy conifers" in Africa). The corresponding scaling factors have been estimated as the final product of CTDAS, and have been applied to obtain the terrestrial biosphere and ocean fluxes at the ecosystem and ocean basin scale by multiplying them with the a priori fluxes. The adjusted fluxes are then put into the transport model to produce an optimized 4-D CO 2 mole fraction distribution.

A priori CO 2 flux data set
In CTDAS, four types of CO 2 surface fluxes are considered as follows: (1) the a priori estimates of the oceanic CO 2 exchange are based on the air-sea CO 2 partial pressure differences from ocean inversions results . These air-sea partial pressure differences are combined with a gas transfer velocity computed from wind speeds in the atmospheric transport model to compute fluxes of carbon dioxide across the sea surface every 3 h; (2) the a priori terrestrial biosphere CO 2 fluxes are from GFED2 (Global Fire Emissions Database version 2), which is derived from the Carnegie-Ames Stanford Approach (CASA) biogeochemical modeling system ( . A monthly varying NEE flux (NEE = R e − GPP) was constructed from the following two flux components: gross primary production (GPP) and ecosystem respiration (R e ), and interpolated to 3-hourly net land surface fluxes using a simple temperature Q 10 relationship assuming a global Q 10 value of 1.5 for respiration, and a linear scaling of photosynthesis with solar radiation. (3) The imposed fossil fuel emission estimates from the global total fossil fuel emission of the CDIAC (Carbon Dioxide Information and Analysis Center) (Marland et al., 2003) were spatially and temporally interpolated following the EDGAR (Emission Database for Global Atmospheric Research) database Commission, 2009;Olivier and Berdowski, 2001;Thoning et al., 1989); (4) the biomass-burning emissions are from GFED2, which combines monthly burned area information observed from satellites  with the CASA biogeochemical model. Fire emissions in GFED2 are available only up to 2008, so for 2009 and 2010 we use a climatology of monthly averages of the previous decade. Note that GFED3 (and now even GFED4) is available for quite a few years, and offers higher spatial resolutions in biomass-burning emissions that are attractive for model simulation. But it uses a different product for the satellite observed NDVI (Normalized Difference Vegetation Index) and FPAR (the Fraction of Photosynthetically Active Radiation) (MODIS (the MODerate resolution Imaging Spectroradiometer) instead of AVHRR (Advanced Very High Resolution Radiometer)) which causes a different seasonality in the biosphere fluxes which are calculated alongside the fire emissions in GFED, with a less realistic amplitude. Since this amplitude of the seasonal biosphere is important to us, we did not update to this new GFED3 product. We also tested the GFED4 data with SIBCASA (Simple Biosphere/Carnegie-Ames-Stanford Approach) to make a new data set of fire estimates but our analyses showed that the impact of using GFED4 versus GFED2 on estimated Asia fluxes is very weak.

Atmospheric CO 2 observations
In this study, two sets of atmospheric CO 2 observation data were assimilated as follows: (1) surface CO 2 observations distributed by NOAA-ESRL (http://www.esrl.noaa. gov/gmd/ccgg/obspack/, data version 1.0.2) and by the WD-CGG (World Data Centre for Greenhouse Gases, http://ds. data.jma.go.jp/gmd/wdcgg/) for the period 2006-2010 (the Asian surface site information is summarized in Fig. 1a and the global surface sites in Table S1 of the Supplement). Individual time series in this surface set were provided by many individual PIs (Principal Investigators) which we kindly acknowledge; (2) for the free tropospheric CO 2 observations, we use the aircraft measurements from the CONTRAIL project for the period 2006-2010 (see Fig. 1b).
A summary of Asian surface sites used in this study is shown in Table 1 and Fig. 1a for reference. There are fourteen surface sites with over 7957 observations located in Asia, including ten surface flask stations and four surface continuous sites. The surface CO 2 mole fraction data used in this study are all calibrated against the same CO 2 standard (WMO-X2007) (The World Meteorological Organization CO 2 mole fraction scale for 2007). For most of the continuous sampling sites at the surface, we derived an averaged afternoon CO 2 concentration (12:00-16:00, local time) for each day from the time series, while at mountain-top sites we constructed an average based on nighttime hours (00:00-04:00, local time) to reduce local influence and compare modeled with observed values only for well-mixed conditions. We note that from the CONTRAIL program Matsueda et al., 2008), stratospheric CO 2 data were not included into CTDAS because the stratospheric observations had a seasonal phase shifting and its smaller amplitude was difficult to compare to the tropospheric measurements . A summary of the CON-TRAIL aircraft measurements is presented in Table 2 and Fig. 1b. The CONTRAIL aircraft data are reported on the NIES (the National Institute for Environmental Studies) 09 CO 2 scale, which are lower than the WMO−X2007 CO 2 scale by 0.07 ppm at around 360 ppm and consistent in the range between 380 and 400 ppm . Thus the CONTRAIL CO 2 data sets are comparable to surface data. We follow the method from Niwa et al. (2012) to divide the data into four vertical bins (575-625, 465-525, 375-425, 225-275 hPa) from ascending and descending profiles and one vertical bin (225-275 hPa) from level cruising. We also divide CONTRAIL data into 42 horizontal bins/regions (Fig. 1b), which amounts to a total of 65 bins. Before daily averaging the CONTRAIL measurements for each 65 regional/vertical bins, we pre-process the aircraft data to obtain free troposphere CO 2 values by filtering out the stratospheric CO 2 data using a threshold of potential vorticity (PV) > 2 PVU (Potential Vorticity Unit, 1 PVU = 10 −6 m 2 s −1 K kg −1 ), in which PV is calculated from TM5 (using ECMWF temperature, pressure and wind fields ) . A total number of 10 467 CO 2 aircraft observations over Asia have been used during the period from January 2006 to December 2010 in our inversion.

Sensitivity experiments and uncertainty estimation
Because the Gaussian uncertainties strongly depend on choices of prior errors in CTDAS, the formal covariance estimates for each week of optimization only reflect the random component of the inversion problem rather than a characterization of the true uncertainties of the estimated CO 2 flux. As an alternative, we performed a set of sensitivity experiments to obtain a more representative spread in the flux estimates and complement the formal Gaussian uncertainty estimates. We take different plausible alternative settings in CTDAS to design a more comprehensive sensitivity test, and use the minimum and maximum flux inferred in these experiments to present the range of the true flux. The following six inversions were performed to investigate the uncertainty span in this study: Case 1: prior flux as in Sect. 2.2 + observations as in Sect. 2.3 + TM5 transport model runs at global 3 • × 2 • and a 1 • × 1 • nested grid over Asia. This is the base simulation (quoted as surface-CONTRAIL) which is used to analyze the 5 year carbon balance in this study.
Case 2: same as Case 1, but excluding CONTRAIL observations. We use these results (quoted as surface-only) to examine the impact of CONTRAIL data on Asian flux estimates by comparison with Case 1.
Case 3: like Case 1, but CTDAS runs with the updated fossil fuel emissions based on Wang et al. (2012) over China. is a value assigned to a given site that is meant to quantify our expected ability to simulate observations and used to calculate the innovation X 2 (Inn. X 2 ) statistics. N denotes the number available in CTDAS. Flagged observations mean a model-minus-observation difference that exceeds 3 times the model-data mismatch, these data are therefore excluded from assimilation. The bias is the average of the posterior residuals (assimilated values-measured values), while the modeled bias is the average of prior residuals (modeled valuesmeasured values). Different from fossil fuel data in Case 1, the data of Wang et al. (2012) calculated carbon emissions from energy consumption, transportation, household energy consumption, commercial energy consumption, industrial processes and waste. And the seasonal variations between the two data sets are different. the fossil fuel emissions in Case 1 had the largest carbon emission in January and the smallest carbon source in July every year, while data of Wang et al. (2012) had the smallest fossil-fuel CO 2 emissions in February or March. This simulation is meant to partly address the impact of uncertainty in fossil fuel emissions over the region as suggested by Francey et al. (2013). Case 4: like Case 1, but CTDAS runs based on 110 % of prior biosphere flux derived from CASA-GFED2; Case 5: like Case 2, but the TM5 transport model is used at global 6 • × 4 • without nested grids. This tests the impact of model resolution; Case 6: like Case 2, but replacing the underlying land use map with MODIS data (Friedl et al., 2002) and keeping the number of ecoregions unchanged. The MODIS land use maps can be found in Fig. S1 in the Supplement.
The Cases 1 and 2 span the period 2006-2010 (the period 2004-2005 was discarded as spin-up), while the other sensitivity experiments were done from 2008 to 2010 only when the observational coverage was best. In general, these six sensitivity tests investigate most variations in the components of the assimilation framework. These variations are prior fluxes, observations available, the ecoregion map, the fossil fuel emissions, and transport. They also give alternative choices for the main components of the system. The sensitivity results are summarized in Table 3 and further discussed in the next section.

Results
We will from here on refer to carbon sinks with a negative sign, sources are positive, and will include the sign also when discussing anomalies (positive = less uptake or larger source, negative = more uptake or smaller source). We describe the results mainly over Asia (global flux estimates can be found in Table S2 in the Supplement), where we expected the CON-TRAIL data to provide the additional constraints. Note that the results of Case 1 are analyzed as the best assimilation for the period of 2006-2010 in this study.

CO 2 concentration simulations
First we checked the accuracy of the model simulation using the surface CO 2 concentration observations and CONTRAIL aircraft CO 2 measurements. Figure 2a shows the comparison of modeled (both prior and posterior) CO 2 concentration with measurements at the discrete surface site of Mt. Waliguan (WLG, located at 36.29 • N, 100.90 • E). Note that the prior CO 2 concentrations here are not really based on a priori fluxes only, as they are a forecast started from the CO 2 mixing ratio field that contains all the already optimized fluxes (1, ..., n−1) that occurred before the current cycle of the data assimilation system (n). So these prior mole fractions only contain five weeks of recent un-optimized fluxes and constitute our "first-guess" of atmospheric CO 2 for each site. For the WLG site, the comparison of the surface CO 2 time series shows that the modeled (both prior and posterior) CO 2 concentration is in general agreement with observed data during the period 2006-2010 (correlation coefficient R = 0.87), although the modeled result still could not adequately reproduce all the observed CO 2 seasonal variations. The posterior annual model-observation mismatch of this distribution is −0.10 ± 1.25 ppm, with 0.07 ± 1.50 ppm bias for the summer period (June-July-August) and 0.02 ± 0.80 ppm bias for the winter period (December-January-February). The model-observation mismatch is a little larger in Case 2 without CONTRAIL data (model-observation mismatch: −0.13 ± 1.26 ppm), suggesting that the surface fluxes derived with CONTRAIL agree with the surface CO 2 mixing ratios at WLG station. Over the full study period, the WLG modeled mole fractions exhibit good agreement with the observed CO 2 time series and the changes in inferred mixing ratios/flux are within the specified uncertainties in our inversion system, an important prerequisite for a good flux estimate.
We also checked the inversion performance in the free troposphere in addition to the surface CO 2 .   of CONTRAIL measurements have been selected and added into the system, only three vertical bin observations have really been assimilated as sparse measurements associated with the 575-625 hPa in CONTRAIL data. Note that the prior CO 2 concentrations here are not really based on a priori fluxes only, as they are a forecast started from the CO 2 mixing ratio field that contains all the already optimized fluxes (1, ..., n − 1) that occurred before the current cycle of the data assimilation system (n). So these prior mole fractions only contain five weeks (five weeks are the lag windows in our system) of recent un-optimized fluxes and constitute our "first guess" of atmospheric CO 2 for each site. , indicating the transport model can reasonably produce the vertical structure of observations. We found that the observed CO 2 concentration profiles were modeled better after assimilation than before (modeled − observed = 0.05 ± 1.25 ppm for a priori and −0.01 ± 1.18 ppm for posterior), although our inverted (posterior) mole fractions still could not adequately reproduce the high values in winter (December-January-February) and the low values in summer (June-July-August). This mismatch of CO 2 seasonal amplitude suggests that our inverted (posterior) CO 2 surface fluxes do not catch the peak of terrestrial carbon exchange well. Previous studies have also found this seasonal mismatch, which may correlate with atmospheric transport, and has already been identified as a shortcoming in most inversions (Peylin et al., 2013;Saeki et al., 2013;Stephens et al., 2007;Yang et al., 2007). In addition, we found that the optimized CO 2 mole fractions seem better captured at low altitude with smaller standard deviations of the model-observation mismatch (±1.12, ±1.18 and ±1.26 ppm for 475-525, 375-425, 225-275 hPa) and higher correlation coefficient at 475-525 hPa. This suggests that the near surface layers are comparatively well constrained in CTDAS. Overall, the agreement between the model and measurements is fairly good and consistent with previously known behavior in the CarbonTracker systems, derived mostly from North American and European continuous sites. Note that all model-observation mismatch of Asian surface sites and CONTRAIL data have been included in Tables 1 and 2 (see column of "Bias (modeled)").

Five-year mean
During the period from 2006-2010, we found a mean net terrestrial land carbon uptake (a posteriori) in Asia of −1.56 Pg C yr −1 , consisting of −2.02 Pg C yr −1 uptake by the terrestrial biosphere and +0.47 Pg C yr −1 release by biomass-burning (fire) emissions (Table 6). This terrestrial uptake compensates 38 % of the estimated +4.15 Pg C yr −1 CO 2 emissions from fossil fuel burning and cement manufacturing in Asia. An uncertainty analysis for the Asian terrestrial CO 2 uptake derived from a set of sensitivity experiments has been conducted and put the estimated sink in a range from −1.07 to −1.80 Pg C yr −1 (Table 3), while the 1-sigma of the formal Gaussian uncertainty estimate is ±1.18 Pg C yr −1 ( Table 6). The estimated Asian net terrestrial CO 2 sink is further partitioned into a −1.02 Pg C yr −1 carbon sink in boreal Eurasia and a −0.68 Pg C yr −1 carbon sink in temperate Eurasia, with a +0.15 Pg C yr −1 CO 2 source in tropical Asia.
The annual mean spatial distribution of net terrestrial carbon uptake over Asia is shown in Fig. 3. Note that the estimated fluxes include terrestrial fluxes and biomass-burning sources but exclude fossil fuel emissions. Most Asian regions were natural carbon sinks over the studied period, with the strongest carbon uptake in the middle and high latitudes of the Northern Hemispheric part of Asia, while the low-latitude region releases CO 2 to the atmosphere. This flux distribution pattern is quite consistent with previous findings that northern temperate and high latitude ecosystems were large sinks  Tables 4 and 5 and Fig. 4 (see Case 1). The majority of the carbon sink was found in the regions dominated by forests, crops and grass/shrubs. The largest uptake is by the forests with a mean sink of −0.77 Pg C yr −1 , 83 % of which (−0.64 Pg C yr −1 ) was taken up by conifer forests and 18 % of which (−0.14 Pg C yr −1 ) by mixed forest, whereas the tropical forests released CO 2 (+0.08 Pg C yr −1 ). The estimated flux by CTDAS in Asian cropland ecosystems was −0.20 Pg C yr −1 , with the largest crop carbon sink located in temperate Eurasia (−0.17 Pg C yr −1 ). The grass/shrub lands in Asia absorbed −0.44 Pg C yr −1 , with most of these grass/shrub sinks located in temperate Eurasia (−0.36 Pg C yr −1 ). Other land-cover types (e.g., wetland, semi tundra and so on) sequestered about −0.15 Pg C yr −1 (10 % of total) over Asian regions. This suggests that according to our model, many ecosystems contributed to Asian CO 2 sinks, highlighting the complexity of the total northern hemispheric sinks.
Also, we note that the detailed CO 2 flux partitioning in our assimilation system highly relies on the prior model description of the ecosystem-by-ecosystem flux patterns. To evaluate the Gaussian errors of the CO 2 flux estimate for a related ecosystem type, we calculated the posterior/prior Gaussian errors (1-sigma) as well as the error reduction for individual ecosystem types during the period 2006-2010 (Table 5). As shown in Table 5, the uncertainty reduction rates are 24.30 %, 23.81 % and 23.81 % for forestlands, Grass/Shrub ecosystems and croplands, respectively. This error reduction suggests that the inferred carbon sink partitioning for individual Atmos. Chem. Phys., 14, 5807-5824, 2014 www.atmos-chem-phys.net/14/5807/2014/ ecosystem types are to some extent constrained by the assimilation system. However, a large uncertainty still exists in the posterior carbon sink for most ecosystem types.We can make the assumption that the correlation between two inverted ecosystem-related fluxes indicates how well the ecosystem-related estimation of carbon fluxes is being constrained by the observations (lower correlation, stronger constrained; while higher correlation, weaker constrained), to further explore the optimized carbon fluxes during the period 2006-2010 (data shown in Table 4). As shown in Fig. 5, the absolute values of posterior correlation coefficients are less than 0.5 (most in the range of −0.3 to 0.5), while they started uncorrelated (0.0). This confirms that ecoregion fluxes have not been fully independently retrieved. Figure 6 shows the prior and posterior seasonal cycles of CO 2 fluxes for the Asia region and its three sub-regions as well as their Gaussian uncertainties. The seasonal amplitude in boreal Eurasia as shown in Fig. 6b proves to be the major contributor to the seasonal signal in Asia (Fig. 6a). The large uptake of boreal Eurasia occurs in summer and the large differences between the prior and the posterior fluxes are also found in the summer growing season, indicating the surface observation network and CONTRAIL data largely affect the estimated fluxes. Our monthly variability is very close to changes in boreal Eurasia presented by Gurney et al. (2004). In Fig. 6c, the seasonal pattern for the temperate Eurasia region shows a comparable pattern to boreal Eurasia but with a smaller seasonal magnitude. And the adjustments of the prior flux in spring and summer are also smaller. The largest CO 2 uptake in temperate Eurasia subregion, however, is shifted from July to August compared to boreal Eurasia, suggesting that a phase shift in the growing season occurred here with the highest CO 2 sink occurring later in the year. This seasonal cycle is slightly different from that reported by Gurney et al. (2004), but shows a nice agreement with the seasonal dynamics of Niwa et al. (2012) in the Southern temperate Asia region, and of Patra et al. (2011) in the Northwest Asia region. In tropical Asia (Fig. 6d), the seasonal variation is very different from other Asian subregions characterized by a weak CO 2 uptake peak in August-October and much smaller carbon release in May-July. Overall, the posterior uncertainty reduction for the period 2006-2010 was about 25 % in Asia, with the largest uncertainty remaining in the summer, suggesting that our model may not fully capture the biosphere sink signal in the growing season. Figure 7 shows the estimated annual cumulative net ecosystem exchange in Asia during the period from 2006-2010 as well as its anomaly with weekly intervals. Here, the biomassburning and fossil fuel emissions are excluded, and only the sum of fluxes from respiration and photosynthesis are shown, because biomass-burning emissions have large interannual variability, especially for tropical Asia. The coefficient of IAV (IAV = standard deviation/mean) in Asian land carbon flux is 0.12, with a peak-to-peak amplitude of 0.57 Pg C yr −1 (amplitude = smallest -largest CO 2 sink), ranging from the smallest carbon uptake of −1.71 Pg C yr −1 in 2010 and the largest CO 2 sink of −2.28 Pg C yr −1 in 2009. As has been noted in many other studies (Gurney et al., 2004(Gurney et al., , 2008Mohammat et al., 2012;Patra et al., 2011;Peters et al., , 2010Yu et al., 2013), the IAV of the carbon flux strongly correlates with climate factors, such as air temperature, precipitation and moisture.

Interannual variability (IAV)
The year 2010 stands out as a particularly low uptake year in Asia, with a reduction of terrestrial uptake of 0.31 Pg C yr −1 compared to the five-year mean. This reduction mainly appeared in temperate Eurasia and ropical Asia, leading to +0.25 Pg C yr −1 (35 % sink reduction) and +0.04 Pg C yr −1 flux anomalies (24 % sink reduction) in their corresponding regions. In 2010, Asia experienced a set of anomalous climate events. For example, temperate Eurasia experienced a severe spring/autumn drought, and a heavy summer flood and a heat wave occurred in 2010 (National Climate Center, 2011). From Fig. 7b, we can see that 2010 did not show large anomalies until after the spring growing season. As anomalous climate appeared, the summer flood and autumn drought were identified as dominant climatic factors controlling vegetation growth and exhibiting a significant correlation with the land carbon sink, particularly in the croplands, grasslands and forests of temperate Eurasia. In the end, 2010 only showed −1.71 Pg C yr −1 biospheric CO 2 uptakes (excluding fires) by the end of the year.
In contrast to 2010, the year 2009 had the strongest carbon sink for the study period, with much stronger uptake in temperate Eurasia (−0.20 Pg C yr −1 anomaly, 28 % increase in CO 2 uptake) as well as in boreal Eurasia (−0.05 Pg C yr −1 anomaly, 4 % uptake increase compared to the five-year mean). It can be seen that 2009 started with a lower-thanaverage release of carbon in the first 4 months (17 weeks) of the year amounting to +0.28 Pg C yr −1 compared to the five-year average of +0.45 Pg C yr −1 . This variation of the Asian terrestrial carbon sink in the spring vegetation growing season may partly relate to a higher spring temperature in 2009 which induced an earlier onset of the growing season and led to a high vegetation productivity by extending the growing season Richardson et al., 2009;Walther et al., 2002;Wang et al., 2011;Yu et al., 2013). From Fig. 7b, 2009 shows a very high carbon uptake in the summer growing season (June-August, weeks 22 to 32) concurrent with favorable temperature and abundant precipitation conditions. After this summer, the vegetation productivity returned back to normal and the total cumulative carbon sink added up to −2.28 Pg C yr −1 at the end of the year with −0.26 Pg C yr −1 extra uptake compared to the fiveyear mean.      respectively. The Asian CO 2 uptake thus ranges from −1.07 to −1.80 Pg C yr −1 across our sensitivity experiments, which complements the Gaussian error. Despite the small numbers of years included, this range suggests that the Asian terrestrial was a sizable sink, while a carbon source implied in previous studies by the 1-sigma Gaussian error of ±1.18 Pg C yr −1 on the estimated mean, is very unlikely. The largest sensitivity in inferred flux is to the change of prior terrestrial biosphere fluxes (Case 4, difference = Case 4 -Case 1). The inversions with different model resolutions (Case 5, difference = Case 5 -Case 2) and with different Chinese fossil fuel emissions (Case 3, difference = Case 4 -Case 1) also show large variations in the inverted CO 2 fluxes, while the sensitivity to the change of land cover types (Case 6, difference = Case 6 -Case 2) is generally modest. This highlights the current uncertainties in the Asian sink and the best method to estimate it from inverse modeling.

Impacts of the CONTRAIL data on inverted Asian CO 2 flux
We examined the impacts of the CONTRAIL data on Asian flux estimation by comparing results from Case 1 (surface-CONTRAIL) and Case 2 (surface-only) ( Table 6 and Fig. 8a). Note that the uncertainties shown in the Table 6 and Fig. 8b are now the Gaussian uncertainties as we did not repeat all sensitivity experiments. As shown in Table 6, inclusion of the CONTRAIL data induces an averaged extra CO 2 sink of about −0.47 Pg C yr −1 to Case 1 (0.47 = 1.56-1.09), with most addition to the grass/shrub ecosystem (Fig. 4). The spatial pattern of Asian fluxes also changed considerably (see Fig. 8a). For instance, a decrease in CO 2 uptake was found in the northern area of boreal Eurasia together with an increase in the south of boreal Eurasia, leading to almost identical total carbon sink strength in boreal Asia between with and without CONTRAIL data. Whereas the estimated flux  distribution in tropical Asia showed a small spatial change and a large increase in regional sink size with CONTRAIL observations included. Table 6 and Fig. 8b shows the reduction of the Gaussian error between Case 1 and Case 2. The error reduction rate (ER) is calculated as the following percentage: where σ surface−only and σ surface−CONTRAIL are Gaussian errors in Case 2 (surface-only) and Case 1 (surface-CONTRAIL), respectively. By including the additional CONTRAIL data into the inversion system, the uncertainty of the posterior flux over Asia is significantly reduced (> 10 %), especially for the southeast of boreal Eurasia, southeast of temperate . The more pronounced reduction was found in boreal Eurasia and tropical  Asia (reducing by 14 % and 15 %, respectively). This suggests that current surface CO 2 observations data alone do not sufficiently constrain these regional flux estimations (there are no observation sites in boreal Eurasia and only one in tropical Asia), and the additional CONTRAIL CO 2 observations impose an extra constraint that can help reduce uncertainty on inferred Asia CO 2 fluxes, especially for these two surface observation sparse regions.

Impact of CONTRAIL
Our modeling experiments reveal that the extra aircraft observations shift the inverted CO 2 flux estimates by imposing further constraints. This confirms the earlier findings by Saeki et al. (2003) and Maksyutov et al. (2013) that the inverted fluxes were sensitive to observation data used. For tropical Asia, inclusion of the CONTRAIL data notably reduced the uncertainties (about 15 % reduction). Compared with an inversion study with the CONTRAIL data for the tropical Asia region (Niwa et al., 2012) , the error reduction rate in land flux estimation in this study for the same region is smaller than that of Niwa et al. (34 %). This difference in uncertainty reduction likely results from the differences in inversion system design between these two studies, of which vertical mixing represented in transport model, and covariance assigned to prior fluxes are typically most important. We furthermore note that the set of observations used in these studies was not identical, we for instance included one tropical surface site (BKT, see Table 1 and Fig. 1a) to constrain the inferred flux estimation but Niwa, et al. (2012) did not. Our results share other features with the Niwa et al. (2012) study, for instance the largest impact on the least data constrained regions. As reported by Niwa et al. (2012), the inclusion of CONTRAIL measurements not only constrains the nearby fluxes, but also reduces inferred flux errors in the regions far from the CONTRAIL measurement locations. For instance, in boreal Eurasia, where no surface site exists and which is far from the CONTRAIL data locations (after preprocessing of horizontal/vertical bins and filter operation of stratospheric, there is no CONTRAIL observation available over this region), uncertainty reductions are large (14 % re-duction in uncertainty). Similar results were also presented by Niwa et al. (2012), with an 18 % error reduction in boreal Eurasia. These two studies consistently suggest that including the CONTRAIL measurements in inversion modeling systems will help to increase the NEE estimation accuracy over boreal Eurasia.
The CONTRAIL constraint on temperate Eurasia is generally modest, only having a 6 % error reduction. This may because temperate Eurasia has more surface observation sites than other regions in Asia. However, it is interesting that the difference in inverted NEE in this region between surfaceonly and surface-CONTRAIL is large (−0.35 Pg C yr −1 ), but inconsistent with Niwa et al. (2012). One cause of this is likely the sensitivity of these inverse systems to vertical transport (Stephens et al., 2007), as also suggested by Niwa et al. (2012). The uneven distribution of observations at the surface and free troposphere may also aggravate this discrepancy.

Comparison of the estimated Asian CO 2 flux with other studies
Our estimated Asian terrestrial carbon sink is about −1.56 Pg C yr −1 for the period 2006-2010. Most parts of Asian were estimated to be CO 2 sinks, with the largest carbon sink (−1.02 Pg C yr −1 ) in boreal Eurasia, a second large CO 2 sink (−0.68 Pg C yr −1 ) in temperate Eurasia, and a small source (+0.15 Pg C yr −1 ) in tropical Asia. This spatial distribution of estimated terrestrial CO 2 fluxes is overall comparable to the results for the period of 2000-2009 by Saeki et al. (2013), derived from an inversion approach focusing on Siberia with additional Siberian aircraft and tower CO 2 measurements, especially in the high latitude areas. Comparisons of our inverted CO 2 flux with previous studies are summarized in Table 7. In boreal Eurasia, our inferred land flux (−1.02 Pg C yr −1 ) is higher than Gurney et al. (2003)  and CT2011_oi (−1.00 Pg C yr −1 , downloaded from http: //carbontracker.noaa.gov). In Temperate Eurasia, our inverted flux is −0.68 Pg C yr −1 , which is well consistent with Gurney et al. (2003) (−0.60 Pg C yr −1 ), but higher than CTE2013 (−0.33 Pg C yr −1 ) and CT2011_oi  and CT2011_oi (+0.14 Pg C yr −1 ). The estimated total Asian terrestrial carbon sink is −1.56 Pg C yr −1 , which is close to the CTE2013 (−1.05 Pg C yr −1 ) and CT2011_oi (−1.27 Pg C yr −1 ). The IAVs comparison between the results from this study and from CTE2013 is also presented in Ta and this study, but the size of the carbon sink is inconsistent. Differences likely stems from the additions of Asian sites and CONTRAIL data in this study. Compared to previous findings, our updated estimation with these additional data seems to support a larger Asian carbon sink over the past decade.
The spatial patterns of NEE in Asia are complex because of large land surface heterogeneity, such as land cover, vegetation growth rates, soil types, and varying responses to climate variations. This makes accurately estimating NEE over Asia challenging. We believe this study is therefore useful to improve our understanding of the Asia regional terrestrial carbon cycle even though our estimation still has remaining uncertainties and biases in the inverted fluxes. By these comparisons, we can also conclude that our inferred Asia land surface CO 2 fluxes support a view that both large boreal and mid-latitude carbon sinks in Asia are balanced partly by a small tropical source. This would support the earlier suggestion that Asia is of key interest to better understand the global terrestrial carbon budget in the context of climate change.
The majority of the CO 2 sink was found in the areas dominated by forests, crops and grass/shrubs, although these were Atmos. Chem. Phys., 14, 5807-5824, 2014 www.atmos-chem-phys.net/14/5807/2014/ not all individually constrained by the observations. Asian forests were estimated to be a large sink (−0.77 Pg C yr −1 ) during the period 2006-2010, the sink size is slightly larger than the bottom-up derived results of Pan et al. (2011) (−0.62 Pg C yr −1 ) for the period 1990-2007. One cause of this discrepancy is likely due to that our estimate is presented at a coarse resolution (a 1 • × 1 • grid may contain other biomes with lower carbon uptake than forests). Another reason may be that about half of Temperate Eurasia was not included in the statistical analysis by Pan et al. (2011). Note that the carbon accumulation in wood products is not considered in our estimates and needs further analysis in future studies.
The croplands in Asia were identified to be an average sink of −0.20 Pg C yr −1 during the period 2006-2010. The uptake in croplands is likely associated with agricultural technique and crop management. Different from other natural ecosystems, crop ecosystems are usually under intensive farming cultivation, with regular fertilizing and irrigation of the crops. This increases crop production, and in return leads to high residues and root to the soil, which increases the carbon sink in cropland . However, the accumulation of crop carbon in most crop ecosystems is relatively low, and agricultural areas are even considered not to contribute to a long-term net sink (Fang et al., 2007;Piao et al., 2009;Tian et al., 2011). This is because the carbon accumulation in the crop biomass is harvested at least once per year and released back as CO 2 to the atmosphere after consumption. We should note that our estimate in the crop sink is different from the results of "crop no contribution " (Piao et al., 2009). Our atmospheric inversion system can well capture the crop's strong CO 2 uptake during the growing season, but the atmosphere locally does not reflect the emission of the harvested crops, which normally have been transported laterally and is consumed elsewhere. This harvested product is likely released from a region with high population density and hard to detect against high fossil fuel emissions, whereas the estimated crop flux remains a large net CO 2 uptake over the period considered even though the crop flux into the soil is relatively small. Thus the croplands' sink in this study might be overestimated due to the absence of harvesting in our modeling system. This issue was also raised by Peters et al. (2007Peters et al. ( , 2010. Grassland/Shrub ecosystems also play an important role in the global carbon cycle, accounting for about 20 % of total terrestrial production and could be a potential carbon sink in future (Scurlock and Hall, 1998). The grass/shrub lands in Asia absorbed a total of −0.44 Pg C yr −1 , accounting for about 25 % of the total Asian terrestrial CO 2 sink, which is close to the averaged global grassland sink percentage of 20 %. Compared to the bottom-up results that net ecosystem productivity was 10.18 g C m −2 yr −1 by Yu et al. (2013), our estimate of 34.32 g C m −2 yr −1 is much higher. This might be due to the fact that the areas in this study include shrubs, whereas other studies only consider grasslands.
The Supplement related to this article is available online at doi:10.5194/acp-14-5807-2014-supplement.