Interannual variations in continental-scale net carbon exchange and sensitivity to observing networks estimated from atmospheric CO2 inversions for the period 1980 to 2005

Authors


Abstract

[1] Interannually varying net carbon exchange fluxes from the TransCom 3 Level 2 Atmospheric Inversion Intercomparison Experiment are presented for the 1980 to 2005 time period. The fluxes represent the model mean, net carbon exchange for 11 land and 11 ocean regions after subtraction of fossil fuel CO2 emissions. Both aggregated regional totals and the individual regional estimates are accompanied by a model uncertainty and model spread. We find that interannual variability is larger on the land than the ocean, with total land exchange correlated to the timing of both El Niño/Southern Oscillation (ENSO) as well as the eruption of Mt. Pinatubo. The post-Pinatubo negative flux anomaly is evident across much of the tropical and northern extratropical land regions. In the oceans, the tropics tend to exhibit the greatest level of interannual variability, while on land, the interannual variability is slightly greater in the tropics and northern extratropics. The interannual variation in carbon flux estimates aggregated by land and ocean across latitudinal bands remains consistent across eight different CO2 observing networks. The interannual variation in carbon flux estimates for individual flux regions remains mostly consistent across the individual observing networks. At all scales, there is considerable consistency in the interannual variations among the 13 participating model groups. Finally, consistent with other studies using different techniques, we find a considerable positive net carbon flux anomaly in the tropical land during the period of the large ENSO in 1997/1998 which is evident in the Tropical Asia, Temperate Asia, Northern African, and Southern Africa land regions. Negative anomalies are estimated for the East Pacific Ocean and South Pacific Ocean regions. Earlier ENSO events of the 1980s are most evident in southern land positive flux anomalies.

1. Introduction

[2] The fate of CO2 emitted into the atmosphere from the combustion of fossil fuels, industrial processes, changes in vegetation, and ocean processes remains a topic of active scientific research, not to mention being of great significance to the climate change policy community. Studies based on observations of atmospheric CO2 combined with estimates of global fossil/industrial and deforestation emissions have concluded that the land and oceans are removing roughly half of the anthropogenic CO2 input [Prentice et al., 2001]. Also evident is the considerable variation in this uptake from 1 year to the next [Baker et al., 2006]. Though a number of hypotheses have been suggested to explain the long-term and interannual variations, direct confirmation of these mechanisms at the regional to global scale remains elusive.

[3] The importance of an improved understanding of carbon exchange has emerged as a key research priority. This has been underscored by research showing a large spread of projected future warming when global carbon cycle models are coupled to simulations of climate change [Friedlingstein et al., 2006]. While some simulations show the ocean and land continuing as a carbon sink, others show the biosphere undergoing dramatic changes. For example, coupled simulations from the Hadley Center in the UK suggest a future in which the American tropical forests undergo significant dieback from increased water stress in addition to widespread oxidation of soil carbon [Cox et al., 2000].

[4] Efforts to understand the current exchange of carbon between the land, oceans, and atmosphere have included a number of observing techniques and simulation approaches. One approach that has met with increasing use is the atmospheric inversion approach, in which the spatial and temporal pattern of atmospheric CO2 observations is used in combination with simulated atmospheric transport to infer sources and sinks of CO2 at the surface [Enting, 2002].

[5] In the last two decades, a number of research groups around the world have carried out atmospheric inversion experiments in an effort to quantify CO2 sources and sinks at a variety of spatial and temporal scales. These began with efforts to estimate the long-term mean exchange at large scales, and have more recently focused on interannual variability and attempts to estimate fluxes at smaller scales [Enting et al., 1995; Fan et al., 1998; Bousquet et al., 2000; Rayner et al., 1999a; Gurney et al., 2002; Rödenbeck et al., 2003b; Peylin et al., 2005; Baker et al., 2006]. Though there is some agreement regarding the hemispheric estimates of carbon exchange, quantification at the continental and regional scale remains particularly uncertain. The spread in the estimates of continental-scale carbon exchange can be traced to a few aspects of the inverse approach: the paucity of CO2 observational data and methods for addressing this, the transport simulation employed, and the inversion conditioning and setup [Engelen et al., 2002].

[6] In the 1990s an experiment was constructed to inter-compare the many transport models and data used in atmospheric CO2 inversion studies in an effort to better understand the sources of uncertainty and potentially arrive at a consensus view of net carbon uptake in the land and oceans. The TransCom experiment contained a number of incremental steps, starting with forward simulations of fossil fuel CO2, net biosphere exchange, and sulfur hexafluoride. The most recent phase of the TransCom experiment is the TransCom 3 effort. TransCom 3 gathered the forward CO2 sensitivities of the participating modeling groups and explored the uncertainties arising in the inversion process from the transport, the data and the inversion set-up itself [Gurney et al., 2000].

[7] A number of papers have been written highlighting analysis of the TransCom 3 results, including an annual mean inversion [Gurney et al., 2002, 2003], a cyclo-stationary or seasonal inversion [Gurney et al., 2004] and an interannual inversion [Baker et al., 2006]. In addition, a number of sensitivity studies have been presented, examining particular aspects of the TransCom atmospheric inversion including data uncertainties [Law et al., 2003], uncertainties in the fossil fuel CO2 data [Gurney et al., 2005], extension of the observing network [Patra et al., 2003], an examination of ocean-only observing networks [Patra et al., 2006], an exploration of the error terms [Engelen et al., 2002], and a comparison to vertical profile observations [Stephens et al., 2007].

[8] The current study builds from the work of Baker et al. [2006] in which an interannual inversion was performed for the 1988 to 2003 time period using all of the TransCom 3, level 2 models. The focus of that study was on the presentation of a “control” or “typical” inversion setup and the flux results and the associated model-dependent uncertainties. Here, we extend the time period of interest to cover 1980 to 2005, allowing for examination of two prominent ENSO events in the 1980s. Most importantly, we build a series of CO2 observing networks for different subsets of our 27-year time span that progressively contain larger numbers of observing stations, allowing for constant observational constraints in the inversion for spans of different lengths [Rödenbeck et al., 2003b; Peylin et al., 2005]. The distinct contribution of this work is the ability to test the sensitivity of station network choices, while accounting for transport uncertainties using a 13 model portfolio, in determining the net carbon variability of 1 year to the next.

[9] Section 2 of this paper provides an explanation of our methods and how the current study differs methodologically from the work of Baker et al. [2006]. In section 3, we present the results of the interannual inversion, using on the mean of the thirteen participating model groups. We present the interannual variability by detrending and deseasonalizing the net carbon exchange in each of the 22 regions, and in grouping the results in a series of progressively larger regions for which the estimation confidence is greater. Both types of estimation error generated in the TransCom experiment are included in all results. This section also presents the estimated carbon exchange given using measurements from different observing networks spanning different sub-periods in the 27-year timespan. In section 4 we discuss the large, coherent cross-network variations, emphasizing the relationship to the El Niño/Southern Oscillation (ENSO). In section 5 we summarize these results and point toward further analysis and the next steps that the TransCom community is taking to further refine and explore atmospheric CO2 inversion research.

2. Methods

2.1. Formalism and Design

[10] The inversion approach used in this study follows the Bayesian synthesis method [Tarantola, 1987; Enting, 2002]. The goal of the atmospheric inversion process is to find the optimal combination of regional surface net carbon fluxes, equation image, that best matches observed CO2, equation image, after those fluxes have been transported through a model atmosphere represented by the operator, M. This is most succinctly represented by the minimization of cost function, J:

equation image

where C(equation image) is the covariance matrix assumed to represent mismatches between the modeled and observed concentrations, and C(equation image) is the covariance matrix representing uncertainties in the prior fluxes. A detailed description of the formalism employed and references to source material are given in previous work for the annual mean [Gurney et al., 2003] and interannual “control” [Baker et al., 2006] TransCom 3 inversions.

[11] Details of the experimental design can be similarly found in the TransCom 3 experimental protocol [Gurney et al., 2000] and the previous TransCom 3 results [Baker et al., 2006; Gurney et al., 2003]. The results presented here follow the inversion set-up outlined in the control interannual inversion [Baker et al., 2006]. The following is a brief summary.

[12] Thirteen transport models (or model variants) ran a series of forward CO2 tracer simulations in order to construct model-specific response functions of atmospheric CO2. The thirteen transports models are described in Baker et al. [2006] Table 2.

[13] For the interannual inversion experiment presented here, the forward simulations were run as Green's functions. A total of 268 tracers were simulated by each model, four of which were “background” global fluxes and 264 of which were region/month fluxes representing a combination of 12 months and the 22 land and ocean regions described in the annual mean inversion experiment [Gurney et al., 2002]. The prespecified background flux patterns were emitted for a single year, then discontinued, allowing the CO2 concentration field to decay for the following 2 years of simulation. The prespecified region/month flux patterns were emitted for a single month then discontinued for the remainder of the 3-year simulation.

[14] In order to lower the computational burden for the participating modeling groups, interannually varying transport was not required. Participants chose a variety of different annually repeating transport products encompassing reanalyzed winds and GCM transport simulations. The lack of interannual transport in the forward simulations is an important consideration in the flux estimation. Four studies have tested the influence of interannual transport versus repeated annual transport fields and the results remain somewhat inconclusive. [Dargaville et al., 2000; Rödenbeck et al., 2003a; P. Patra, personal communication, April 2005; Peylin et al., 2005]. When examining latitudinal aggregates (90–30°S, 30–0°S, 0–30°N, 30–90°N) for fixed versus interannually-varying winds, Dargaville et al. [2000] found that the predominant difference was in the flux amplitude as opposed to the phasing of the interannual variations. By contrast, Rödenbeck et al. [2003a] found both amplitude and phase flux differences when comparing interannually varying winds to a series of fixed-year transport winds, though the phasing differences were found mostly in the northern temperate zone (15–50°N). In this region, flux differences occasionally exceeded 1.0 GtC/year. Finally, the work of Peylin et al. [2005] indicates that the main impact of interannually varying winds is on the amplitude as opposed to the phasing of the carbon flux and these effects are evident at the regional scale only. As described in Baker et al. [2006], recent unpublished inverse-estimated fluxes show small standard errors of roughly 0.2 GtC/year result from computations run with fixed versus interannually varying winds [P. Patra, personal communication, April 2005]. However, it is important to note that though each group used a single year of winds which were repeated to build the CO2 response functions, many of the modeling groups employed different repeating calendar years in their simulations. Thus the spread in the model results not only reflects different transport algorithms but a mix of the individual year chosen for the recycled winds. Nevertheless, the lack of interannually varying transport must be considered when interpreting the interannual flux variations.

[15] The four background carbon fluxes consisted of 1990 and 1995 fossil fuel emission fields, an annually-balanced, seasonal biosphere exchange and air-sea gas exchange [Andres et al., 1996; Brenkert, 1998; Randerson et al., 1997; Takahashi et al., 1999]. These fluxes are included in the inversion with a small prior uncertainty (C(equation image) in equation (1)) so that their magnitude is effectively fixed. Examination and implication of these background fluxes have been explored in previous work [Gurney et al., 2003, 2005]. The 264 region/month fluxes estimated by the inversion are deviations from these global background fluxes for each month in each year and are given prior flux uncertainties of a much larger magnitude than the background fluxes, allowing for significant adjustment given the dictates of the CO2 observations. The background fossil fuel emission fluxes were prescribed without seasonality. The neutral terrestrial fluxes were purely seasonal (i.e., they integrate to zero across the full year), and the background ocean fluxes were prescribed with both seasonal variations and annual mean uptake.

[16] The same seasonally-varying prior fluxes used in the cyclostationary inversion study [Gurney et al., 2004] were used here. The prior flux uncertainty is important for keeping the estimated fluxes within biogeochemically realistic bounds, though the inclusion of prior fluxes can strongly influence the estimated flux where observations are limited. The extent of this influence depends on the prior flux uncertainty, which has been explored in previous TransCom work. The prior flux uncertainties used in the main diagonal of the prior flux covariance matrix are identical to those used in the cyclostationary control inversion [Gurney et al., 2004] for land, while the regional ocean prior flux uncertainties for each month are a constant value defined as

equation image

where C(SL1,ocn) are the annual mean prior flux uncertainties used in the annual mean control inversion [Gurney et al., 2002].

[17] The CO2 observational data were derived from the GLOBALVIEW-2006 data set [GLOBALVIEW-CO2, 2006]. GLOBALVIEW is a data product that interpolates flask and continuous CO2 measurements to give 48 synchronous values per year. Gaps in the data are filled by extrapolation from marine boundary layer measurements. Because many of the stations within the observational data set have been initiated/retired at various times over the last quarter century and significant gaps exist, a number of different station networks were constructed in this experiment and used to create independent interannual flux estimates. Including or retiring observing sites in a continuous fashion with one network construction can introduce false variability at the time of station introduction, a phenomenon avoided with the batch inversion method employed in this study.

[18] Three criteria were employed to arrive at the observing network makeup. First, in order to be included in a network, a station could not have more than a certain number of contiguous years in which the proportion of real to total data falls below 50%. For networks spanning more than 15 years, a station could not have more than 3 contiguous years under this criteria. For networks spanning lengths of 10 to 15 years, this number falls to no more than two. For networks spanning less than 10 years, this falls to no more than one. Second, the multi-year mean percentage of non-missing data in the years that qualified under the first criteria had to exceed 67%. Finally, no networks were begun after 1995 because of the limitation that would place on the length of available CO2 observations and because a network starting at a later date posed few differences from those begun in 1995.

[19] The combination of these three criteria naturally led to the networks represented in Table 1. The smallest network spanning the beginning of 1979 to the end of 2005 time period contained 24 stations while the largest, spanning the beginning of 1995 to the end of 2001 time period, contained 102. The 24 station and 102 station networks are shown in Figure 1.

Figure 1.

(a) Deseasonalized monthly net carbon exchange estimates Dr(t) (GtC/year) for the Europe land region for all of the 13 participating TransCom 3 Level 2 models; (b) as in (a) but with each individual model's long-term mean (1980–2005) mean subtracted off; (c) the 13-model mean of the monthly flux estimates from (b), bounded by the 1σ within uncertainty (outer interval) and the adjusted model spread (inner hatched interval).

Table 1. Site Number, Time Period, Station Count, and Individual Station Code Namesa for the Eight Networks Considered in This Studyb
Site #80-0580-9780-9090-0590-9993-0595-0595-01
24273057758995102
  • a

    See GlobalView [2006] for a complete description of the individual station codes.

  • a

    The site names contain the following elements: [3 character site name] [data grouping]_[lab #] [sampling strategy] [sampling platform] where the “data grouping” provides a height or lat/lon; the “lab #” refers to the specific measurement laboratory; the “sampling strategy” refers to discrete (D), continuous/quasi-continuous (C), or integrated sampling (I); the “sampling platform” refers to fixed (0), ship (1), aircraft (2), tower (3), kite (4), balloon (5), or Firn/ice core (6).

  • b

    CO2 observations at Darwin, Australia were excluded from consideration. See Gurney et al. [2004] and Law et al. [2003] for a discussion.

1alt_06D0alt_06D0alt_06D0alt_01D0aia005_02D2alt_01D0alt_01D0alt_01D0
2ams_11C0ams_11C0ams_01D0alt_02D0aia015_02D2alt_02D0alt_02D0alt_02D0
3asc_01D0asc_01D0ams_11C0alt_06C0aia025_02D2alt_06C0alt_06C0alt_04D0
4azr_01D0azr_01D0asc_01D0alt_06D0aia045_02D2alt_06D0alt_06D0alt_06C0
5brw_01C0bhd_15C0avi_01D0ams_11C0alt_01D0ams_11C0ams_11C0alt_06D0
6brw_01D0brw_01C0azr_01D0asc_01D0alt_02D0asc_01D0asc_01D0ams_11C0
7cba_01D0brw_01D0bhd_15C0bme_01D0alt_04D0azr_01D0ask_01D0asc_01D0
8cmn_17C0cba_01D0brw_01C0bmw_01D0alt_06C0bal_01D1azr_01D0ask_01D0
9gmi_01D0cmn_17C0brw_01D0brw_01C0alt_06D0bme_01D0bal_01D1azr_01D0
10key_01D0cmo_01D0cba_01D0brw_01D0ams_11C0bmw_01D0bme_01D0bal_01D1
11kum_01D0gmi_01D0cmn_17C0cfa_02D0asc_01D0brw_01C0bmw_01D0bhd_15C0
12mlo_01C0key_01D0cmo_01D0cgo_01D0bhd_15C0brw_01D0brw_01C0bme_01D0
13mlo_01D0kum_01D0gmi_01D0cgo_02D0bme_01D0bsc_01D0brw_01D0bmw_01D0
14nwr_01D0mbc_01D0key_01D0cmn_17C0bmw_01D0car030_01D2bsc_01D0brw_01C0
15psa_01D0mlo_01C0kum_01D0crz_01D0brw_01C0car040_01D2car030_01D2brw_01D0
16sch_23C0mlo_01D0mbc_01D0esp_06D0brw_01D0car050_01D2car040_01D2bsc_01D0
17sey_01D0nwr_01D0mlo_01C0frd040_06C3cba_01D0car060_01D2car050_01D2car030_01D2
18smo_01C0psa_01D0mlo_01D0gmi_01D0cfa_02D0cfa_02D0car060_01D2car040_01D2
19smo_01D0sch_23C0nwr_01D0gsn_24D0cgo_01D0cgo_01D0cfa_02D0car050_01D2
20spo_01C0sey_01D0psa_01D0hba_01D0cgo_02D0cgo_02D0cgo_01D0car060_01D2
21spo_01D0smo_01C0sbl_06D0izo_01D0cgo_04D0cmn_17C0cgo_02D0cba_04D0
22stm_01D0smo_01D0sch_23C0key_01D0cmn_17C0cpt_36C0cmn_17C0cfa_02D0
23stmebc_01D0spo_01C0sey_01D0kum_01D0crz_01D0crz_01D0coi_20C0cgo_01D0
24wes_23C0spo_01D0smo_01C0maa_02D0frd040_06C3eic_01D0cpt_36C0cgo_02D0
25 stm_01D0smo_01D0mhd_01D0gmi_01D0esp_06D0crz_01D0cgo_04D0
26 stmebc_01D0spo_01C0mid_01D0gsn_24D0frd040_06C3eic_01D0cmn_17C0
27 wes_23C0spo_01D0mlo_01C0hba_01D0gmi_01D0esp_06D0coi_20C0
28  stm_01D0mlo_01D0itn051_01C3gsn_24D0frd040_06C3cpt_36C0
29  stmebc_01D0mlo_02D0itn496_01C3hat_20C0gmi_01D0cri_02D0
30  wes_23C0mqa_02D0izo_01D0hba_01D0gsn_24D0crz_01D0
31   nwr_01D0izo_27C0hun_01D0hat_20C0eic_01D0
32   poc000_01D1key_01D0hun082_35C3hba_01D0esp_02D0
33   pocn05_01D1kum_01D0hun115_35C3hun_01D0esp_06D0
34   pocn15_01D1ljo_04D0ice_01D0hun082_35C3gmi_01D0
35   pocn20_01D1maa_02D0izo_01D0hun115_35C3hat_20C0
36   pocn30_01D1mhd_01D0jbn_29C0ice_01D0hba_01D0
37   pocs05_01D1mid_01D0key_01D0izo_01D0hun_01D0
38   pocs10_01D1mlo_01C0kum_01D0jbn_29C0hun115_35C3
39   pocs15_01D1mlo_01D0lef_01D0key_01D0ice_01D0
40   psa_01D0mlo_02D0maa_02D0kum_01D0izo_01D0
41   rpb_01D0mlo_04D0mhd_01D0lef_01D0izo_27C0
42   ryo_19C0mqa_02D0mhdcbc_11C0lmp_28D0jbn_29C0
43   sch_23C0nwr_01D0mhdrbc_11C0maa_02D0key_01D0
44   sey_01D0poc000_01D1mid_01D0mhd_01D0kum_01D0
45   shm_01D0pocn05_01D1mlo_01C0mhdcbc_11C0kum_04D0
46   smo_01C0pocn10_01D1mlo_01D0mhdrbc_11C0lef011_01C3
47   smo_01D0pocn15_01D1mlo_02D0mid_01D0lef_01D0
48   spo_01C0pocn20_01D1mnm_19C0mlo_01C0lef030_01C3
49   spo_01D0pocn25_01D1mqa_02D0mlo_01D0lef076_01C3
50   spo_02D0pocn30_01D1nwr_01D0mlo_02D0lef244_01C3
51   stm_01D0pocs05_01D1poc000_01D1mnm_19C0lef396_01C3
52   stmebc_01D0pocs10_01D1pocn15_01D1mqa_02D0ljo_04D0
53   syo_01D0pocs15_01D1pocn20_01D1nwr_01D0lmp_28D0
54   syo_09C0pocs25_01D1pocs05_01D1poc000_01D1maa_02D0
55   tap_01D0pocs30_01D1pocs15_01D1pocn05_01D1mhd_01D0
56   uum_01D0pocs35_01D1prs_21C0pocn15_01D1mhdcbc_11C0
57   wlg_01D0prs_21D0psa_01D0pocn20_01D1mhdrbc_11C0
58    psa_01D0rpb_01D0pocs05_01D1mid_01D0
59    rpb_01D0ryo_19C0pocs15_01D1mlo_01C0
60    ryo_19C0sch_23C0prs_21C0mlo_01D0
61    sch_23C0sey_01D0psa_01D0mlo_02D0
62    sey_01D0shm_01D0rpb_01D0mlo_04D0
63    shm_01D0sis_02D0ryo_19C0mnm_19C0
64    smo_01C0smo_01C0sch_23C0mqa_02D0
65    smo_01D0smo_01D0sey_01D0nwr_01D0
66    spo_01C0spo_01C0shm_01D0prs_21C0
67    spo_01D0spo_01D0sis_02D0psa_01D0
68    spo_02D0spo_02D0smo_01C0rpb_01D0
69    stm_01D0stm_01D0smo_01D0ryo_19C0
70    stmebc_01D0stmebc_01D0spo_01C0sch_23C0
71    syo_01D0syo_01D0spo_01D0sey_01D0
72    syo_09C0syo_09C0spo_02D0shm_01D0
73    tap_01D0tap_01D0stm_01D0smo_01C0
74    uum_01D0uta_01D0stmebc_01D0smo_01D0
75    wlg_01D0uum_01D0syo_01D0smo_04D0
76     wlg_01D0syo_09C0spo_01C0
77     wlg_33C0tap_01D0spo_01D0
78     wpo000_10D2uta_01D0spo_02D0
79     wpon05_10D2uum_01D0spo_04D0
80     wpon10_10D2wis_01D0stm_01D0
81     wpon15_10D2wlg_01D0stmebc_01D0
82     wpon20_10D2wlg_33C0syo_01D0
83     wpon25_10D2wpo000_10D2syo_09C0
84     wpon30_10D2wpon05_10D2tap_01D0
85     wpos05_10D2wpon10_10D2tdf_01D0
86     wpos10_10D2wpon15_10D2uta_01D0
87     wpos15_10D2wpon20_10D2uum_01D0
88     wpos20_10D2wpon25_10D2wes_23C0
89     zep_01D0wpon30_10D2wis_01D0
90      wpos05_10D2wpo000_10D2
91      wpos10_10D2wpon05_10D2
92      wpos15_10D2wpon10_10D2
93      wpos20_10D2wpon15_10D2
94      yon_19C0wpon20_10D2
95      zep_01D0wpon25_10D2
96       wpon30_10D2
97       wpos05_10D2
98       wpos10_10D2
99       wpos15_10D2
100       wpos20_10D2
101       wpos25_10D2
102       zep_01D0

[20] The construction of the prior observational data covariance matrix, C(D), for the eight observational networks and its scaling to account for model-data mismatch errors followed the methodology outlined in Baker et al. [2006] except that, in the current study, measurements considered coincident in space and time were retained and their uncertainty was increased by the square root of the number of coincident measurements. Observing locations were considered coincident provided their latitude and longitude were within 4° of each other and they were within 900 meters in the vertical.

[21] Because of the extended time period examined, updated fossil fuel emissions were used in this study and are derived from the work of Andres et al. [1996], Brenkert [1998] and Marland et al. [2006]. The global fossil fuel CO2 emissions for 2004 and 2005 were estimated using linear extrapolation of the 1998 to 2003 time period. This approach was chosen because these years exhibited consistent growth at a rate greater than previous years which, combined with the continued recent global economic expansion, were considered a better representation of emissions in the 2004/2005 timeframe.

2.2. Reduction of Inversion Results

[22] In order to analyze the interannual variability from the TransCom 3 inversion results, the neutral biosphere and background ocean fluxes were added back into the 11 land and 11 ocean residual fluxes returned by the inversion. This total flux represents all surface exchange other than the background fossil fuel CO2 emissions. The individual model monthly flux estimates are then deseasonalized using a compact 13 month trapezoidal running mean as follows:

equation image

where Dr(t) is the deseasonalized flux for a particular region, r, for a particular month, t, and Fr(t) is the original flux. Though the interannual variability exhibits some model-to-model variation, the dominant difference is due to a long-term mean offset. This was shown in Figure 2 of Baker et al. [2006]. These offsets can be removed by computing the individual model long-term mean and removing this from every monthly flux for that model. Finally, the model mean flux estimate can be calculated in an effort to quantify the mean flux for a given month and flux region across all transport realizations in the TransCom 3 experiment.

[23] Uncertainty associated with the model mean month flux estimate is represented by the “within” uncertainty and “model spread” described in equations 1 and 2 of Gurney et al. [2004]. In order to compute the deseasonalized within model uncertainty, the posterior covariance matrix must be operated on by a deseasonalization operator. This is done as follows:

equation image

where Q represents the deseasonalized posterior covariance matrix for an individual region and model, O represents the matrix form of the 13 month trapezoidal running mean, and P is the posterior covariance matrix for an individual region and model. An RMS operator is then applied to the diagonal of Q so that the within uncertainty can now be written as,

equation image

[24] The model spread of the deseasonalized fluxes is computed as in Gurney et al. [2004] except the flux, in the current case, is the deseasonalized rather than the fully seasonal flux. This can be expressed as,

equation image

As was shown in Baker et al. [2006], much of the spread across the models is due to long-term mean offsets and hence a model spread that includes a correction for the long-term mean offsets has been employed here. This was performed by computing an individual long-term mean estimated flux (spanning entire network time period) for each model/network relative to the ensemble model-mean estimate flux for that network and subtracting that offset from each monthly estimated flux for the particular model/network. This expression of model spread focuses on differences in variability rather than long-term means and, as such, is often a more appropriate measure of model-to-model variations. It is referred to throughout this study as the “adjusted model spread”. An example of this is shown in Figure 1 for the Europe land region.

3. Results

[25] Figure 2 reveals how sparse the CO2 observations can be, particularly across the tropical latitudes and in some of the TransCom land regions, such as Africa and South America. Though the networks that span the later time periods have greater station density over land, large gaps remain. This forms the core motivation for adding the prior constraint (S0 in equation (1)) in the inversion [Enting, 2002]. Aggregating the regional fluxes in space and time, post-inversion, is one way to lower the uncertainty due to the limited observational constraint [Kaminski et al., 2001; Engelen et al., 2002]. Figure 3 shows the result of aggregating the basis function regions for total land, ocean and latitudinal totals.

Figure 2.

The locations of the observing stations for the largest (102 stations spanning 1995 to 2001—open square) and the smallest (24 stations spanning 1980 to 2005—closed circle) networks considered in this study. The 11 land and 11 ocean regions of the TransCom 3 experiment are also shown.

Figure 3.

TransCom 3 Level 2 model mean deseasonalized net carbon exchange estimated using the 24-station 1980 to 2005 observational network. (a) Total Land (thick red line), Total Ocean (thick blue line) and Global Total (thick black line) fluxes. (b) Northern Land (solid black line; sum of Boreal N America, Temperate N America, Boreal Asia, Temperate Asia, Europe), Tropical Land (solid red line; sum of Tropical America, Northern Africa, Tropical Asia), and Southern Land (solid blue line; sum of South America, Southern Africa, Australasia) fluxes. (c) Northern Oceans (solid black line; sum of Northern Pacific, Northern Ocean, North Atlantic), Tropical Oceans (solid red line; sum of East Pacific, West Pacific, Tropical Atlantic, Tropical Indian), and Southern Oceans (solid blue line; sum of South Pacific, Southern Ocean, South Atlantic, South Indian) fluxes. Within uncertainty interval (1σ) is shown by thin solid lines and the adjusted model spread by the hatched interval. Note the two y-axes in each plot and the shifted scales for each. Also shown in (a) (thick green line) is the Multivariate El Niño/Southern Oscillation (ENSO) Index (MEI) [Wolter and Timlin, 1993, 1998], as well as the timing of the Mt. Pinatubo eruption.

3.1. Aggregated Results

[26] Figure 3a shows the model mean Total Land, Total Ocean and Global Total fluxes for the longest of the observational networks constructed in this study (1980 to 2005—24 stations). Both the within uncertainty and adjusted model spread are included. Also included in the figure is the Multivariate ENSO Index (MEI) [Wolter and Timlin, 1993, 1998]. The MEI is based on the six main observed variables over the Tropical Pacific and includes sea-level pressure, zonal and meridional surface winds, sea surface temperature, surface air temperature, and total cloudiness fraction. Positive values reflect the warm phase (El Niño) and are associated with warm surface water and air temperatures in the Eastern Tropical Pacific [Trenberth and Tepaniak, 2001].

[27] The Global Total carbon exchange variability is driven primarily by terrestrial exchange and shows a strong correlation to the ENSO timing, as has been noted by other investigators [Bacastow, 1976; Keeling et al., 1995; Bousquet et al., 2000; Rayner et al., 1999b; Feely et al., 1999; Rödenbeck et al., 2003b; Peylin et al., 2005 among others]. The Total Land carbon exchange tends to lag somewhat behind the peak of the MEI index and responds by turning from sink to source or a lessened sink. Furthermore, the influence of the June 1991 Pinatubo eruptions can be seen as a gradual strengthening of the global sink in the early 1990s, confounding the ENSO events of the early 1990s [Bousquet et al., 2000].

[28] As found in Baker et al. [2006], the relative magnitude of the Total Land and Total Ocean interannual variability is partly due to the tighter prior flux uncertainty of the oceans compared to the land. However, experiments in which the ocean prior flux uncertainty was scaled to closely match the average level over the land regions show that the relative amounts of interannual variability are not completely dependent upon the prior uncertainty, but reflect constraints imposed by the observed CO2. The ratio of Total Ocean to Total Land 1σ variability in the 1980–2005 observational network with commensurate prior flux uncertainties was 0.8 (compared to 0.31 for the “non-loosened” prior uncertainties) while the same ratio, when the inversion is performed with the 1995 to 2001 observational network, is 0.68 (compared to 0.40 in the “non-loosened” case). The tighter ocean prior flux uncertainties have an influence on the relative land/ocean variability, but that influence lessens as more observing stations are included in the inverse setup.

[29] Figures 3b and 3c show the ocean and land regions divided into north, tropical and southern domains. On land, the Southern Land and Tropical Land aggregated regions exhibit slightly more interannual variability than the Northern Land region. In this study, the Southern Land includes the Southern Africa region and this accounts for much of the variability in the Southern Land region. In the ocean, the Tropical Ocean aggregate region exhibits the greatest variability. In both the land and ocean aggregated regions, the within uncertainty in the tropics is larger than the adjusted model spread. This, in turn, is driven by the limited number of observations in the tropical regions, particularly over land, and the fact that rapid vertical mixing limits the constraint imposed by surface observations [Baker et al., 2006]. As will be shown later, this is improved somewhat by networks that include observing stations added in the 1990s and the first part of the 21st century.

[30] The adjusted model spread in the flux estimation across the models is much smaller than the within uncertainty in all of the aggregated regions—this is primarily due to the fact that the spread has been constructed by removing the individual model long-term mean offsets. The dominance of the within error is a persistent theme throughout the analysis and suggests that the models agree on the interannual variations to a much greater extent than found with the estimation of long-term means. Other studies have arrived at similar conclusions, whether considering a transport model suite as in Baker et al. [2006] or a range of inversion parameters [Bousquet et al., 2000; Peylin et al., 2005].

3.2. Observing Network Influence

[31] The influence of using different observing networks can be seen in Figure 4, which presents the global total and latitudinal aggregates for the entire time period and with all observing networks. For the Total Land and Total Ocean, both the long-term means and the interannual variations show considerable agreement. As noted before, the larger (and hence shorter timespan) networks differ in size (number of stations) from one another to a greater degree than the smaller networks. Furthermore, these stations contain a higher proportion of continental versus ocean sites and hence pose greater challenges to model transport [Patra et al., 2006]. Not surprisingly, inter-network differences are much greater in the 1990s and the early years of the 21st century.

Figure 4.

TransCom 3 Level 2 model mean deseasonalized net carbon exchange estimated for latitudinal aggregates (as defined in caption to Figure 3) using all eight observational networks (see Table 1): 1980–2005; 24 stations (black solid), 1980–1997; 27 stations (red solid), 1980–1990; 30 stations (blue solid), 1990–2005; 57 stations (green solid), 1990–1999; 75 stations (gold solid), 1993–2005; 89 stations (black dashed), 1995–2001; 102 stations (green with diamond symbol), 1995–2005; 95 stations (lavender with square symbol). (a) Total Land; (b) Total Ocean; (c) Northern Land; (d) Northern Oceans; (e) Tropical Land; (f) Tropical Oceans; (g) Southern Land; (h) Southern Oceans.

[32] The agreement between the observing networks in the latitudinal aggregates is much less but is primarily composed of constant long-term offsets and amplitude differences. Much of the phasing of the variability remains coherent across the networks. Exceptions to this are the 1990 to 1995 period in the Tropical Land region where the networks that begin in 1990 (57 and 75 stations) show features nearly opposite in sign to those in the longer networks (24 and 27). The Northern Ocean shows a similar feature in the 2000 to 2002 time period. Other notable phase mismatches occur in the Northern Land and the Southern Ocean, 1997 to 1999.

[33] Figure 5 presents the model mean, deseasonalized carbon exchange for all of the individual land and ocean regions and for all observing networks. The within uncertainty and the adjusted model spread associated with the longest-running observing network (1980 to 2005) are included in the figure. It is worth noting that the adjusted model spread in the longest-running observing network reflects the model-to-model differences with long-term means removed but the model mean flux for the remaining networks contain long-term mean offsets.

Figure 5.

TransCom 3 Level 2 net carbon exchange estimated for individual regions using all eight observational networks (as defined in caption to Figure 4). (a) Boreal N America. (b) Temperate N America. (c) Tropical America. (d) South America. (e) Northern Africa. (f) Southern Africa. (g) Boreal Asia. (h) Temperate Asia. i) Tropical Asia. (j) Australasia. (k) Europe. (l) North Pacific. (m) West Pacific. (n) East Pacific. (o) South Pacific. (p) Northern Ocean. (q) North Atlantic. (r) Tropical Atlantic. (s) South Atlantic. (t) Southern Ocean. (u) Tropical Indian. (v) South Indian. Within uncertainty interval (1σ) is shown by thin black solid lines and the adjusted model spread by the hatched interval for the longest observational network only (1980–2005; 24 stations). Note that the vertical scale is different for the land versus ocean regions.

[34] Many regions show significant differences in both the long-term mean flux and flux variations across the different station networks. Long-term mean flux offsets are notable in Tropical America and Northern Africa, with the smaller networks showing net emission of over 1.0 GtC/year versus the larger networks which show a near-neutral net exchange. Similarly, the shorter networks show net uptake in Temperate Asia across the decades of the 1990s and 2000s whereas the longer networks show near neutral exchange. The smaller networks show a somewhat weaker uptake over the 1990s and 2000s in Europe relative to the longer observing networks, which estimate uptake of roughly 1.5 GtC/year. In the ocean regions, the South Pacific ocean shows large long-term mean flux discrepancies (>1 GtC/year) from one network to the other in the 1990s and 2000s. The South Atlantic ocean and Southern Ocean show discrepancies in the mid- to late-1990s while the North Atlantic ocean shows some discrepancies in the 1980s.

[35] Though network dependence does occur, the phasing of the flux variability is often consistent across station networks even when the amplitude is not. Examples of this include Temperate North America, Southern Africa, Europe, the Pacific and Indian oceans.

[36] On land, the interannual variability is consistently greatest in the Temperate North America and Tropical America land regions. In the oceans, the South Pacific ocean and Tropical Indian ocean tend to dominate the interannual variability. The Tropical America land region exhibits shifts of almost 2 GtC/year over the course of 2 to 3 years, while the Tropical Indian ocean region exhibits variations as large as 1 GtC/year over a 2-year timespan.

[37] The two uncertainties vary in a relative sense from region to region. Overall, the 1σ within uncertainties are larger than the 1σ model spread (true for both original and adjusted form). Exceptions to this occur in Europe and particular time periods in Temperate North America and the Northern Ocean regions. In these few cases, flux differences among the individual models are as large as the average of individual model error. Furthermore, the magnitude of the model spread often varies over time, suggesting that disagreement between model estimated fluxes is more pronounced at times.

[38] Another way of examining the dependence of the uncertainty estimates on the network size is presented in Figure 6. For each of the eight networks, the within uncertainty and model spread were computed based on the first 5 years of error covariances and fluxes. As the station network increases in size, the overall observational constraint increases, causing the within uncertainty to decline. However, the increase in network size has the opposing influence on the model spread, causing it to increase. The decline in the within uncertainty reflects the increasing observational constraint placed on the flux estimates. The rise in the model spread, however, reflects the fact that as stations are added to the observing network, the differences between the models are systematically exposed to a greater degree. The model spread is also consistently larger over the land regions than over the ocean. This is due to the limited number of observing sites in tropical land regions and the fact that CO2 gradients over land are larger and more dependent upon differences in transport out of the planetary boundary layer [Gurney et al., 2003].

Figure 6.

Network within uncertainty and model spread for the aggregate Tropical Land region (sum of Tropical America, Northern Africa, Tropical Asia) based on the first 5 years of net carbon exchange covariances (within uncertainty) and fluxes (model spread) for each of the eight observing networks (as defined in caption to Figure 4). The within uncertainty calculation was computed from the full posterior covariance matrices according to equation (5).

4. Discussion

[39] As expected, the interannual variations in the net terrestrial carbon exchange at the global scale are consistent with a number of other inverse studies that quantified interannual variations [Bousquet et al., 2000; Rödenbeck et al., 2003b; Patra et al., 2005; Rayner et al., 2005; Peylin et al., 2005; Baker et al., 2006]. All show correspondence between land fluxes and the El Niño/Southern Oscillation. In addition, all these studies show a reduction in the ENSO-flux correlation for the period following the Mt. Pinatubo eruption in the Philippines, in which large amounts of aerosol were emitted into the stratosphere and subsequently transported around the planet. Beyond cancellation of the positive carbon flux anomaly associated with the ENSO timing, the period following the Pinatubo eruption is marked by a gradual increase in uptake on land. This negative flux anomaly is most readily seen in the Boreal Asia, Temperate North America, Tropical Asia, and Europe regions (see Figure 5). Negative anomalies are also evident in the Tropical America and Temperate Asia regions for the shorter networks suggesting that the post-Pinatubo impact was felt widely across both the tropical and northern extratropical land regions. This has been hypothesized as NPP stimulation from the increase in diffuse to direct radiation resulting from the aerosol loading in the atmosphere [Gu et al., 2003; Farquhar and Roderick, 2003]. Recent modeling research suggests that, rather than an NPP stimulation, the decline in the CO2 atmospheric growth rate results from a combination of greater ocean uptake, reduced heterotrophic respiration, and reduced biomass burning [Angert et al., 2004].

[40] Results from Peylin et al. [2005] show the inverse result placing most of the negative anomaly in North America but the biogeochemical model simulations also performed in that study show a somewhat broader anomaly across the northern hemisphere. Rödenbeck et al. [2003b] place the negative anomaly in Tropical America and the eastern portion of Temperate North America.

[41] The most obvious example of the relationship between ENSO and net carbon exchange can be seen for the very strong ENSO event of 1997/1998. Large biomass fires in Tropical Asia were related to this ENSO event; one study estimated the carbon loss during this period to range from 0.8 to 2.6 GtC/year, while another had a larger upper bound estimating losses at 0.8 to 3.7 GtC/year [Page et al., 2002; Langenfelds et al., 2002; van der Werf et al., 2006]. All are consistent with the TransCom results, which show a positive flux anomaly during this period [Schimel and Baker, 2002; Baker et al., 2006].

[42] The magnitude of this global land flux anomaly evident in Figure 3 is roughly 4.0 GtC/year. This anomaly is seen in the aggregated Tropical Land region estimate, totaling roughly 2 to 2.5 GtC/year, and in the Southern Land region total at approximately 1.0 to 1.5 GtC/year. Though the eight networks show mean offsets, the anomalous flux is consistent in terms of phasing and magnitude, as shown by the normalization of the Tropical Land flux anomalies in Figure 7.

Figure 7.

TransCom 3 Level 2 model mean Tropical Land net carbon exchange anomalies for all eight observational networks (as defined in caption to Figure 4). Fluxes have been deseasonalized, detrended (monthly fluxes adjusted by the long-term mean slope to achieve a long-term mean slope of zero), standardized (anomalies are in units of standard deviation), and long-term means are centered on zero flux.

[43] Recent work by van der Werf et al. [2006] on fire events shows significant consistency with some of the larger flux anomalies in the Tropical Land and Southern Land regions. Roughly two thirds of the 1997 global carbon emissions due to fire were attributed in that study to a region similar to the Tropical Asia region in this study (Figure 5i). The remainder of fire-derived carbon emissions in 1997 were attributed to regions closely aligned with the Southern Africa and South America regions denoted here (Figures 5f and 5d, respectively). The 1997 ENSO anomaly is also evident in the inverse estimated fluxes for the Temperate Asia region, though this flux anomaly disappears for networks composed of greater than 30 stations. These same larger networks show a larger positive flux anomaly for the Tropical Asia region then, raising the possibility that the smaller networks have insufficient information to separate the fluxes in these two neighboring flux regions. No significant fire emissions are evident in van der Werf et al. [2006] for the Temperate Asia region.

[44] The station networks that extend back to 1980 allow for an analysis of previous ENSO events. The Southern Land aggregate region shows positive flux anomalies in both the 1982/1983 and 1987 ENSO events, while the Tropical Land shows a response for the 1987 ENSO only, driven solely by Tropical America (though somewhat lagged). Unlike the strong 1997/1998 ENSO event, the Tropical Asia region shows little response to the earlier ENSO events. South America and Southern Africa exhibit positive flux anomalies coincident with all three large ENSO events among the southern land regions. Temperate Asia also shows a positive anomaly during these earlier ENSO events though the fact that the larger networks eliminated the response in 97/98 raises questions about how well the sparser networks are able to separately estimate fluxes at the continental scale. Rödenbeck et al. [2003b] also found little response in Tropical Asia to the ENSO events of the 1980s, while showing a large positive flux response to the 1997/1998 event.

[45] The strong ENSO event of 1997/1998 was also observed to cause lessened outgassing of CO2 from the Eastern Tropical Pacific and has been estimated in some of the inverse studies [Feely et al., 2002; Patra et al., 2005; Peylin et al., 2005]. Like the study of Baker et al. [2006], this study finds an anomalous uptake in the Eastern Tropical Pacific (∼0.2 GtC/year) and a larger anomalous uptake in the South Pacific region (∼0.6 GtC/year).

[46] Consistent among most of the interannual inverse studies is the conclusion that the interannual variability in the global exchange is dominated by the land versus the ocean [Bousquet et al., 2000; Rödenbeck et al., 2003b; Patra et al., 2005; Baker et al., 2006]; Peylin et al., 2005]. Furthermore, the oceanic interannual variability is greatest in the tropics, though large variability can be found in the Southern Ocean and South Pacific. Tropical ocean variability is dominated by the Tropical Indian ocean. As mentioned previously, the lesser ocean versus land interannual variability can partly be explained by the tighter ocean prior flux uncertainty, though as the observational constraint increases with increasing observing network size, the lesser ocean variability appears to be a reflection of the observational data constraint.

5. Conclusions

[47] The TransCom 3 Level 2 experiment is an atmospheric CO2 inversion in which 13 participating modeling groups submitted atmospheric CO2 concentration flux responses to which the inverse method has been applied. In the current study, the TransCom 3 interannual inversion has been extended to span the 1980 to 2005 time period. In addition to examining the estimation uncertainty and spread of results from the participating modeling groups, this study has also examined the sensitivity of the results to eight different observing networks. Constructing different networks allows us to examine earlier time periods in which fewer sites were taking data, while avoiding potential time-dependent biases associated with varying the sites used inside a single inversion.

[48] The results of this research confirm that the land is the dominant contributor to the interannual variability in the net carbon exchange: the inversions for all the station networks exhibit this broad feature. In both the ocean and land, the tropics tend to account for the majority of the interannual variability.

[49] As has been found in previous work, the total land net carbon exchange shows considerable correlation with the El Niño/Southern Oscillation, with positive terrestrial carbon flux anomalies coinciding with the peak of the ENSO warm phase. The large 1997/1998 ENSO event is easily discerned and is seen most clearly in the carbon flux estimates for the Tropical Asia, Temperate Asia, Northern Africa and Southern Africa land regions. The carbon emission during this period is consistent with the observation of large fires in Tropical Asia. Earlier ENSO events of the 1980s are most evident in southern land positive flux anomalies, particularly evident in South America and Southern Africa.

[50] The slowing of the global CO2 growth rate after the Pinatubo eruption is also evident in the Total Land net carbon exchange and this appears to be evident at the regional scale over much of the tropical and northern extratropical land regions. The within uncertainty remains greatest in the tropics but lessens as additional observing stations are included in the station networks. The model spread, however, increases with increasing station density, owing to the greater observational constraint illuminating the model-to-model transport differences.

[51] Consideration of the interannual variability shows that the phasing of the flux variability is relatively consistent across model transport and the eight networks, particularly for the latitudinally aggregated land and ocean regions. This consistency is somewhat less when the 22 individual basis function regions are considered. However, even at the individual basis function region, inconsistencies are for particular isolated timespans only.

[52] The consistency of the results among the station networks in terms of the interannual variability suggests that these time-dependent inversions could be used to explore the relationships between net carbon exchange and climate indices or climate variability itself. Such an exploration may hold insights into biogeochemical mechanisms and is an important consideration for future research.

[53] Though the participating models in the TransCom 3 experiment show considerable agreement on the interannual variations in net carbon exchange, further exploration into the model-to-model differences remains a top priority for research, as transport is a significant contributor to the uncertainty in inverse results. Furthermore, the potential for bias in the transport algorithms of the models involved must be examined and used to improve transport simulations. This is particularly true for long-term mean flux estimation, for which uncertainties appear much greater than estimates of interannual variation.

[54] Other aspects of this work deserving of more attention include an exploration of individual station sensitivity. It is likely that a small subset of stations are responsible for most of the inconsistencies seen between the networks.

[55] Lastly, the time-dependant inversions performed here should be updated with the newly available vertical CO2 profile measurements in order to explore how these measurements might alter both the long-term mean and the interannually-varying carbon flux estimates as has been suggested in recent research [Stephens et al., 2007].

Acknowledgments

[56] We thank Ken Masarie, Pieter Tans, ESRL personnel and all contributors to the GLOBALVIEW-CO2 data product for the timely and open availability of the atmospheric CO2 observations. We would also like to thank all the TransCom 3 L2 modelers for providing the response functions used in this study.

Ancillary