Global Biogeochemical Cycles

Role of biomass burning and climate anomalies for land-atmosphere carbon fluxes based on inverse modeling of atmospheric CO2



[1] A Time-dependent inverse (TDI) model is used to estimate carbon dioxide (CO2) fluxes for 64 regions of the globe from atmospheric measurements in the period January 1994 to December 2001. The global land anomalies agree fairly well with earlier results. Large variability in CO2 fluxes are recorded from the land regions, which are typically controlled by the available water for photosynthesis, and air temperature and soil moisture dependent heterotrophic respiration. For example, the anomalous CO2 emissions during the 1997/1998 El Niño period are estimated to be about 1.27 ± 0.22, 2.06 ± 0.37, and 1.17 ± 0.20 Pg-C yr−1 from tropical regions in Asia, South America, and Africa, respectively. The CO2 flux anomalies for boreal Asia region are estimated to be 0.83 ± 0.19 and 0.45 ± 0.14 Pg-C yr−1 of CO2 during 1996 and 1998, respectively. Comparison of inversion results with biogeochemical model simulations provide strong evidence that biomass burning (natural and anthropogenic) constitutes the major component in land-atmosphere carbon flux anomalies. The net biosphere-atmosphere carbon exchanges based on the biogeochemical model used in this study are generally lower than those estimated from TDI model results, by about 1.0 Pg-C yr−1 for the periods and regions of intense fire. The correlation and principal component analyses suggest that changes in meteorology (i.e., rainfall and air temperature) associated with the El Niño Southern Oscillation are the most dominant controlling factors of CO2 flux anomaly in the tropics, followed by the Indian Ocean Dipole Oscillation. Our results indicate that the Arctic and North Atlantic Oscillations are closely linked with CO2 flux variability in the temperate and high-latitude regions.

1. Introduction

[2] Carbon dioxide is the most important anthropogenic greenhouse gas in the Earth's atmosphere. However, the regional distribution and quantification of carbon sources and sinks still have large uncertainties, and also the interannual variability in CO2 emission and sequestration lack process-based understanding [e.g., Schimel et al., 2001]. Two approaches are generally applied to determine regional distribution of carbon sources and sinks in the global carbon cycle: (1) bottom-up integration technique where fluxes associated with different processes are estimated individually using observed data that control fluxes and processed based modeling and (2) using atmospheric CO2 observations together with atmospheric transport models to infer the net exchange between surface and the atmosphere. The former requires continuous monitoring of changes in the landcover and ecosystems types [e.g., Lucht et al., 2002; Page et al., 2002; Janssens et al., 2003], as well as partial pressure differences of CO2 between the atmosphere and the ocean [see Takahashi et al., 2002, and references therein]. The first approach often suffers from insufficient data coverage, but in recent times, satellite observations are being used to fill some of the data gaps [Nemani et al., 2003; van der Werf et al., 2004]. The latter (inverse or top-down) technique uses background measurements of atmospheric CO2, which contain information of carbon exchanges at regional scale [Bolin and Keeling, 1963; Tans et al., 1990; Enting et al., 1995]. These estimations can be utilized to study the large-scale variability of land/sea-air exchanges of CO2, given observed climate variations, observations of volcanic eruptions, biomass burning, and also possibly not yet identified processes [e.g., Bousquet et al., 2000; Gu et al., 2003; Rödenbeck et al., 2003b] (see also P. K. Patra et al., Interannual and decadal changes in the sea-air CO2 flux from atmospheric CO2 inverse modeling, submitted to Global Biogeochemical Cycles, 2005]) (hereinafter referred to as Patra et al., submitted manuscript, 2005).

[3] The focus of this article is to understand the variability in CO2 fluxes from land regions during the 1990s. Some earlier studies have already indicated that biomass burning sources of CO2 can be traced by inverse models for some regions [Schimel and Baker, 2002; Rödenbeck et al., 2003b]. We have attempted here to understand the link between CO2 emission from land areas and climate conditions at different sizes of the regions. In addition, the results are compared with the bottom-up approaches, when available, and ecosystem model results.

2. Materials and Methods

2.1. Inverse and Transport Models

[4] Monthly mean CO2 fluxes from 64 regions (Figure 1) of the globe are derived from atmospheric observations worldwide using a time-dependent inverse (TDI) model based on Bayesian approach. The TDI model was originally developed by Rayner et al. [1999], and has been widely used in the transport model comparison study TransCom-3 to solve for fluxes from 22 regions [Gurney et al., 2004]. The TransCom-3 TDI model has been modified for higher spatial resolution recently (Patra et al., submitted manuscript, 2005). In Bayesian (or synthesis) inversion the cost function

equation image

is minimized with the solution

equation image

and posterior flux error covariance

equation image

Here G represent atmospheric transport operator, the diagonal elements in S are the optimized fluxes of CO2 from 64 regions of the inverse model, and the elements of CD and image are the error variance-covariance matrix of the data and the prior fluxes [e.g., Rodgers, 2000]. The values of D and CD and S0 and image are chosen the same as those described by Patra et al. (submitted manuscript, 2005). The abundances of atmospheric CO2 and associated uncertainties at 87 stations are taken from the GlobalView data set [GlobalView-CO2, 2002], and the method is described in more detail elsewhere (Patra et al., submitted manuscript, 2005). The NIES/FRSGC global transport model [Maksyutov and Inoue, 2000], driven by 6-hourly NCEP/NCAR reanalysis meteorology [Kistler et al., 2001], has been used to simulate concentrations from the flux distributions. To determine the robustness of our results, we have performed sensitivity experiments with respect to CO2 sampling network, initial conditions, and prior uncertainties which are documented by Patra et al. (submitted manuscript, 2005). That study suggested that the CO2 flux anomalies at various regional scales are identical in phase and vary only in amplitude under this inverse model framework using an 87-stations data network.

Figure 1.

Inverse model regions (42 land and 22 ocean) and surface observation network (87 stations) used in this study. Four yellowish shades are used to identify the four partitions (except two for tropical Asia) of the large subcontinental scale land regions (names are shown as used in the discussion), and greenish shades are used for the ocean regions.

2.2. Biogeochemical Model

[5] To analyze the inverse model results, we compare them with the predictions of the Biome-BGC (BioGeochemical Cycles) ecosystem model [Running and Coughlan, 1988]. Biome-BGC simulates net primary production (NPP) and heterotrophic respiration (HR) of the land biosphere. The model uses daily climate data from NCEP/NCAR reanalysis, vegetation type as defined by a set of ecophysiological characteristics and site conditions (e.g., plant functional types, soil properties, terrain) to estimate fluxes of carbon, nitrogen, and water through ecosystems (see Fujita et al. [2003] and references therein for details on the present setup). In this model, NPP estimates depend primarily on the amount of photosynthetically active radiation (PAR) and water availability (through rainfall). The interannual variability in PAR and its effect on global carbon cycle are not analyzed in this study. Heterotrophic respiration of litter and soil organic carbon depends on the size of the litter and soil carbon pools and their decomposition rates. The rate constants are a function of soil moisture, which is affected by rainfall and temperature. Decomposition rates depend also on the availability of soil mineral N (nitrogen) for those steps which are immobilizing N [Thornton et al., 2002]. The Biome-BGC model version used for this study does not model individual fire events, but includes a parameterization of fire frequency and intensity via fire mortality fraction of plant carbon pools [Thornton et al., 2002]. In this scheme, some fraction of the ecosystem is subjected to fire each year, at a rate commensurate with the long-term fire frequency. Net ecosystem exchange (NEE), defined as (NPP-HR-Fire), will be utilized in the discussion here. If NPP is greater than (HR+Fire), the land biosphere acts as a carbon sink which corresponds in our convention to a negative CO2 flux anomaly.

2.3. Correlation Calculation and Significance Test

[6] Covariability of CO2 flux anomaly with several climate indices and meteorological parameters are studied to indicate the key controlling factors for interannual variations in CO2 sources and sinks. For this purpose we have calculated correlation coefficients (R) between CO2 flux anomaly for the land regions and climate oscillation indices as well as rainfall and temperature anomalies. We have used time lags up to 6 months for CO2 flux anomalies. Three-month moving window averages (MWAs) are taken for all the parameters between 1994 and 2001 to reduce the serial correlation in the time series (sample size = 32). This is in contrast to running averages taken in diagrams for reducing high-frequency variability in the monthly mean fields. Since no monthly data are used more than once in MWA, the average values are less correlated serially compared to the running averages. The adopted method for correlation calculation in this work cannot be used for estimating statistical significance of correlations by the standard procedure if the anomaly time series or climate indices are auto-correlated [Ebisuzaki, 1997]. We have also tested the correlations using 6-month MWAs (i.e., at half degrees of freedom) and observed that the standard significance test (e.g., one-tailed student's t-test) remain consistent with results obtained using 3-month MWAs. For example, if an estimated R of 0.35 corresponds to 97.5% significance level for N = 32 (the 3-month MWA case) that R increases to more than 0.49 for N = 16 (in 6-month MWA case) to satisfy the 97.5% significance level.

3. Results and Discussions

[7] Carbon dioxide fluxes are estimated for the period 1994–2001 from 64 regions of the TDI model using CO2 observations at 87 stations. Though we have retrieved CO2 fluxes for the period from January 1988 to December 2001, the period analyzed here covers the period from January 1994 to December 2001 only. This is because CO2 measurements are available at 81 stations in 1994, about 25 stations more compared to the measurement network in 1992). The global land and ocean fluxes of CO2 are estimated to be −1.15 ± 0.74 and −1.88 ± 0.53 Pg-C yr−1, respectively. The spatial distribution of oceanic fluxes is fairly well constrained in this inverse model framework (for further details see Patra et al., submitted manuscript, 2005). However, since the observations of CO2 at the land are far more variable and available over fewer locations compared to oceanic region, the flux estimates for many land regions are only poorly constrained (not shown here).

3.1. Interannual Variability in Total Land Fluxes

[8] The CO2 flux anomalies are retrieved with lesser uncertainty from atmospheric CO2 observations by the inverse models at both global (Figure 2) and regional scales compared to the absolute magnitude of fluxes. The absolute flux determination could be biased owing to the biases in forward modeling of CO2 concentrations at specific sites (see the last term in equation (2)). We report here results from a control inversion and four sensitivity inversions for which initial covariance matrices of CO2 data (CD) and fluxes image are varied. The control inversions use TransCom-3 a priori CO2 fluxes and uncertainties scaled to agree with TransCom regions when aggregating subregions. Four sensitivity inversions are carried out with (1) doubled prior flux uncertainty image for ocean regions only, (2) doubled image for all the regions, (3) doubled CO2 data uncertainty (CD), and (4) doubled both image and CD. To put our results in perspective we compare them with previous estimates of Rayner et al. [1999] (results updated), who used similar inverse modeling frameworks but different number of inverse model regions and CO2 observation networks. Rayner et al. [1999] used δ13C and O2/N2 measurements at Cape Grim to separate global land and ocean interannual variability in addition to CO2 data. The results of Bousquet et al. [2000] are also obtained using a similar Bayesian inversion technique from atmospheric measurements at 67 sites. Keeling et al. [1995] have applied a double deconvolution method to atmospheric CO2 and δ13C observations for estimating the land and oceanic CO2 flux variabilities (results updated). The estimation by LeQuéré et al. [2003] is based on an ocean model simulation. The results of Rödenbeck et al. [2003b] are obtained using an adjoint of the tracer transport model TM3 [Heimann, 1996] and CO2 data from 35 CMDL flask sampling stations only. For global land region, the CO2 flux anomalies obtained by different groups and estimation techniques agree fairly well and overall the variabilities are captured consistently. Though all models produced an anomalous CO2 release from the land to atmosphere during the 1997/1998 El Niño period (seen in the background), our results indicate a much larger anomaly of ∼5.63 Pg-C yr−1. This CO2 flux anomaly is as large as 84% of the global fossil fuel emission (6.65 Pg-C yr−1) from all the industrial sources during 1997–1998 [Marland et al., 2003]. The regions of largest variability and potentially associated processes are discussed in details below.

Figure 2.

Variability in global land CO2 fluxes as estimated by several groups using different modeling approaches. Average seasonal cycles for the analysis period are subtracted from the monthly fluxes to calculate the monthly anomalies. The TDI model results (this work) are shown as 6-month running averages. The a posteriori flux estimate uncertainties are shown as vertical bars for each month. The x axis major tickmarks represent January of each year, and positive CO2 flux anomaly indicates a net exchange of carbon from the land to the atmosphere. These conventions will be used throughout this article.

3.2. Variability in Regional Land Fluxes and Meteorology

[9] Figure 3 shows the time series of CO2 flux anomalies derived for a few TDI model regions, and regional scale anomalies in two basic meteorological parameters: air temperature at 2 m height and accumulated monthly rainfall. For tropical Asia and South America (Figures 3a and 3b) the comparison suggests that meteorological conditions play a dominant role in controlling the CO2 flux variability. In general for the tropics, positive temperature and negative rainfall anomalies coincide with positive phases of CO2 flux anomaly, i.e., net fluxes from land to the atmosphere. A possible explanation is that during periods of warmer air temperature and negative rainfall anomaly (drought-like condition), heterotrophic respiration supersedes the net primary production. The reverse is true during the negative temperature and positive rainfall anomaly periods. This may in particular be true for the Amazon basin, which contains about 50% of the world's undisturbed tropical evergreen forest and large areas of tropical savannah. This interpretation is consistent with the results obtained using a process-based biogeochemical model of terrestrial ecosystem [Tian et al., 1998] that suggest water availability reduces carbon uptake in drier and warmer El Niño periods by limiting tree growth. Drought leads to enhanced biomass burning in the tropics due to the favorable dry condition for naturally and anthropogenically caused fire. The response of tropical Asia region to a drier meteorological condition is similar to the Amazon region (see Figure 3a), and the change in rainfall above the southern part seems to be more critical compared to the air temperature for CO2 flux variation. Statistically significant correlation coefficients (R) between CO2 flux anomaly and rainfall (temperature) anomaly in the period 1994–2001 are found to be −0.47 (0.57) and −0.63 (0.41) for tropical South America and tropical Asia, respectively. The correlation between CO2 and rainfall anomalies is highest for a time lag of the fluxes by 4–5 months (see Table 1 for maximum R and associated time lags of other regions). Typically, maximum rainfall anomalies are seen in the months of maximum seasonal rainfall over the tropical regions. Hence we suggest that this lag reflects the time period between biomass growth during the wet season and subsequent drying, and finally CO2 release to the atmosphere because of burning. In boreal Asia the basic atmospheric parameters do not show any clear covariability with CO2 flux anomaly.

Figure 3.

Anomaly in regional CO2 flux (in Pg-C yr−1; control run), air temperature (at 2 m, in °C) from NCEP/NCAR reanalysis, and NOAA Climate Prediction Center (CPC) merged analysis of precipitation (cm month−1) for (a) tropical Asia, (b) tropical South America, and (c) boreal Asia. The anomalies are plotted for individual TDI model regions (Figure 1) and after aggregating to larger TransCom-3 regions. Six-month running means are taken for all the time series to reduce high-frequency noise. To calculate the monthly mean anomaly of air temperature and rainfall over TDI model regions, area averaged values are estimated first, and then the average seasonal cycles for respective regions are subtracted from the monthly values. The TDI model estimated a posteriori flux uncertainties are about ±1.1, ±1.0, and ±0.85 Pg-C yr−1 for tropical Asia, tropical South America, and boreal Asia, respectively.

Table 1. Auto-Correlation Coefficients (R) Between the CO2 Flux Anomaly or PCs of Land Fluxes With the Climate Indices or Regional Anomaly in Two Basic Meteorological Parameters (Rainfall and Air Temperature at 2 m) for the Period 1994–2001a
Flux Anomaly Region/PCClimate Oscillation IndicesMeteorology
  • a

    Rime lags in CO2 flux anomaly to obtain maximum correlations are given in parentheses (unit: month). The following climate indices are considered here: ENSO, El Niño Southern Oscillation; NAO, North Atlantic Oscillation; PDO, Pacific Decadal Oscillation; IOD, Indian Ocean Dipole; and AO, Arctic Oscillation. The larger region flux anomalies are calculated after aggregating the fluxes from four smaller-size TDI model regions (except tropical Asia, which has two regions). The values of R higher than 0.4 are shown in boldface. The sample size in our correlation calculation is 32; effective number of degrees of freedom is not calculated by taking into account any serial correlation in the time series. However, if we neglect serial correlation in the MWA time series, R = 0.4 corresponds to 99% significance level (under one-tailed test; 32 degrees of freedom; see, Significance of Correlations link).

Boreal North America0.15 (1)0.13 (5)0.11 (2)0.31 (0)0.40 (5)−0.24 (0)0.25 (1)
Temperate North America0.41 (1)0.47 (2)−0.13 (3)0.43 (4)−0.34 (1)−0.32 (3)0.33 (4)
Tropical South America0.48 (0)−0.20 (0)0.26 (3)0.59 (3)−0.24 (1)0.47 (4)0.57 (0)
Temperate South America−0.36 (5)0.15 (5)−0.13 (0)−0.19 (5)0.39 (2)−0.18 (4)−0.39 (0)
Tropical Africa0.71 (3)−0.17 (0)0.40 (0)0.52 (5)−0.25 (0)0.42 (4)0.36 (0)
South Africa0.64 (2)−0.25 (2)0.43 (5)0.57 (2)−0.12 (0)−0.14 (0)−0.20 (5)
Boreal Asia0.21 (5)−0.36 (5)0.35 (2)−0.13 (0)0.44 (5)0.17 (5)0.23 (0)
Temperate Asia−0.29 (3)−0.11 (0)−0.13 (5)0.14 (0)−0.22 (3)0.14 (5)−0.20 (1)
Tropical Asia0.61 (2)−0.18 (0)0.34 (3)0.67 (5)−0.23 (3)0.63 (5)0.41 (2)
Australia0.31 (0)−0.31 (2)0.20 (2)0.40 (3)−0.33 (0)−0.22 (5)0.18 (0)
Europe−0.18 (0)0.45 (5)−0.38 (5)0.14 (0)0.38 (3)0.18 (5)0.47 (1)
PC - 10.88 (1)0.14 (0)0.56 (1)0.80 (2)−0.16 (3)  
PC - 20.59 (4)0.22 (0)0.44 (1)0.55 (5)0.54 (0)  
PC - 30.20 (3)−0.20 (1)0.45 (0)0.22 (5)0.41 (1)  

3.3. Seasonality in Regional Fluxes and Fire Counts

[10] There are strong indications that biomass burning is a main reason for positive anomalies in land-atmosphere fluxes [e.g., Langenfelds et al., 2002; Page et al., 2002; van der Werf et al., 2004]. Figure 4 shows derived CO2 flux anomaly estimates for several inverse model regions which are relatively well constrained by CO2 observations in the downstream of synoptic wind regimes for two distinctly different years. During the first period from July 1997 to June 1998, the CO2 flux anomaly for tropical Asia was estimated to be very large, ∼1.27 ± 0.22 Pg-C yr−1, by the inverse model. The uncertainty in flux anomaly is calculated as standard deviation of flux anomalies based on the control and four sensitivity runs of the TDI model. All the TDI model estimates capture the temporal variation in the anomalous CO2 emission fairly consistently (sometimes only the amplitude varies; not shown here). As seen from the ATSR World Fire Atlas (European Space Agency–ESA/ESRIN, Frascati, Italy,, biomass burning started in the southern part of tropical Asia in late 1997 and then fire propagated to the northern part by early 1998. This is captured in the CO2 flux anomaly quite well; the flux anomaly showed higher values in the south part during 1997, and during 1998 both the parts exhibited high positive CO2 flux anomalies (see Figure 4a). The comparison indicates that the location and timing of CO2 flux anomalies associated with forest fires are captured fairly consistently for several other regions of our TDI model (tropical Africa, southern Africa, Australia). Only two other regions, namely, tropical South America (Figure 4b) and boreal Asia (Figure 4c) are depicted here. Probably fires in Amazon and Mexico basin (Figure 4b) and ecosystem disturbances due to climate variation jointly led to a release of anomalous CO2 amount of about 2.06 ± 0.37 Pg-C yr−1 in the July 1997/June 1998 period. Anomalous CO2 sources from boreal Asia are estimated to be 0.83 ± 0.19 and 0.45 ± 0.14 Pg-C yr−1, respectively, during the years 1996 and 1998. Apparently, the timing and location of ATSR detected forest fire corresponds fairly well with the temporal and spatial distribution of CO2 flux anomaly. Similar covariability in fire counts and CO2 flux anomaly are also discussed by Rödenbeck et al. [2003a]. The CO2 flux anomalies from tropical and South Africa are also observed to be high, estimated at 1.17 ± 0.20 (a posteriori uncertainty = 0.9) and 1.48 ± 0.27 (a posteriori uncertainty = 1.1) Pg-C yr−1, respectively.

Figure 4.

Spatial distribution of CO2 flux anomaly (in Pg-C yr−1) during two different years (first and third column) and ATSR fire count anomaly (second and fourth column; in count/month) for three geographic areas: (a) tropical Asia, (b) tropical South America, and (c) boreal Asia. Note that the time periods of averaging (given as heading) as well as the shading intervals (see color bar) are different for each plot. The ATSR World Fire Atlas is generated at 1 × 1 degree horizontal resolution from nighttime radiometric data only, and we have used the fire maps derived using Algorithm 1 (hot spot if: 3.7 μm is saturated, i.e., >312 Kelvins) that is designed to produce no false alarm (see for more information).

3.4. Comparison of Regional Fluxes With Other Estimates

[11] Figure 5 shows the variability in annual mean CO2 flux anomalies estimated using TDI and the Biome-BGC ecosystem model. The bottom-up estimates of CO2 fluxes due to Indonesian fire in 1997 and fires in the eastern Russia during 1996–2001 as well as ATSR fire count anomalies (annual mean) are presented for comparison. In general, the ATSR fire count anomalies (annual mean) are also shown as the circumstantial evidence for enhanced biomass burning location and periods. In general, it is observed that the ecosystem model flux anomalies are much smaller in amplitude compared to those obtained from the TDI model or suggested by the other estimates [e.g., Page et al., 2002]. The best match between TDI and Biome-BGC flux estimations is seen for tropical South America. It should be reminded here that this version of Biome-BGC model does not have explicit treatment of biomass burning. However, on the basis of the fire mortality ratio (that varies with ecosystem types and the climate), some carbon is released to the atmosphere [Thornton et al., 2002]. However, this not sufficient to simulate intense CO2 release events due to forest fires, for example, 0.45 ± 0.31 Pg-C yr−1 from central and northern South America, and 1.34 ± 0.67 Pg-C yr−1 from Southeast Asia according to estimates of van der Werf et al. [2004]. Apparently, the mismatches between TDI and Biome-BGC modeled fluxes in our study are of similar magnitudes, about 0.75 Pg-C yr−1 for tropical South America and 1.09 Pg-C yr−1 for tropical Asia, considering the differences in region selection and uncertainties in both estimations (this study and that of van der Werf et al. [2004]).

Figure 5.

Time series of several independent estimates of annual mean CO2 flux anomaly (circle, TDI estimate; triangle, Biome-BGC model) and regional total ATSR fire count anomaly (rotated square). The fire count anomalies are divided by 1000 to fit to a common y axis scale. In addition, the bottom-up estimate of anomalous carbon emission due to Indonesian forest fire is shown in Figure 5a (vertical bar), and that due to forest fire in Boreal Asia is shown in Figure 5c (squares). Note here that the annual means for the tropical regions are calculated for July–June period (value at the beginning of a year) and that for the boreal region is done using January–December period (value at the middle of a year). The ATSR fire counts for 1996 are available only from the month of July onward.

[12] CO2 flux anomalies as estimated by TDI show a significant correlation with ATSR fire count anomaly; R values are 0.96 and 0.92 for tropical Asia and tropical South America, respectively, and 0.56 for boreal Asia. A less significant correlation for the boreal Asia region is obtained. This may either indicate that fires are only partially responsible for carbon release or a reflection of the heterogeneity in biomass distribution. For example, the negative CO2 flux anomaly in 1999 was possibly caused by the below normal occurrence forest fire in boreal Asia as a whole and the eastern region in particular [Sukhinin et al., 2003]; furthermore, according to ATSR fire count data, fires were less widespread in the year 2000 than the average. The bottom-up estimate for carbon release during Indonesian peat fire in 1997 of Page et al. [2002] is within the range of our estimate but for a larger region (tropical Asia) and longer period (July 1997 to June 1998). During the enhanced fire years the TDI CO2 flux anomalies for boreal Asia compare well with those obtained using forest inventories and fire counts for eastern Russia [Kasischke and Bruhwiler, 2003]. For the 1996 CO2 flux anomaly (+0.83 Pg-C yr−1), contribution of ecosystem response (simulated by Biome-BGC and assuming negligible fire occurrence in this model) mainly due to climate anomalies was about 0.24 Pg-C yr−1. This would imply CO2 fire fluxes in 1996 and 1998 of a similar magnitude of 0.59 and 0.48 (−0.03 from ecosystem flux) Pg-C yr−1, respectively. As per the Global Fire Monitoring Center report on Russian Federation forest fire statistics [Sukhinin et al., 2003], the number of fires was larger in 1996 (7103 counts) compared to 1998 (6307 counts), whereas the total burned area was the opposite, 6.05 × 104 km2 in 1996 and 9.53 × 104 km2 in 1998. Thus anomalous CO2 release due to forest fires from the boreal Asia region in those 2 years may be realistic. On the basis of a similar analysis, CO2 emissions from fire in 2001 are estimated at 0.44 Pg-C yr−1 due to about 7.56 × 104 km2 fire burned area (7095 counts). In 2000 both the fire burned area and counts were also larger; however, our TDI model does not capture any related CO2 flux signal.

3.5. Correlations Between Flux Anomalies and Climate Oscillation Indices

[13] We have calculated the correlation coefficients (R) between regional CO2 flux anomalies and several climate oscillation indices, namely, the ENSO, PDO, NAO, AO, and IOD. The El Niño Southern Oscillation (ENSO) is defined as the sea surface temperature anomaly (SSTA) in the equatorial East Pacific [Diaz and Markgraf, 2000]. In its positive phase (referred to as El Niño), below normal rainfall and warmer surface air temperature (SAT) is observed in the tropical land regions, while an opposite weather condition prevails over southwestern United States. During the positive phase of Pacific Decadal Oscillation (PDO), SSTA are positive (southerly wind stress) on the North American coast, which produces enhanced rainfall in the adjacent land region [Mantua et al., 1997]. A positive Arctic Oscillation (AO) is defined as a period of below normal Arctic sea level pressure (SLP), which is associated with warmer and wetter than normal conditions in northern Europe and eastern North America [Thompson and Wallace, 2001]. The effect of North Atlantic Oscillation (NAO) on eastern Europe and western North America is similar to that of the AO [Hurrell et al., 2003]. The Indian Ocean Dipole (IOD) is defined as the SSTA between eastern and western parts of the equatorial Indian Ocean [Saji and Yamagata, 2003]. Its positive phase (warmer western Indian Ocean) is associated with above normal rainfall over continental Africa and India with an opposite effect on the Southeast Asia region.

[14] CO2 flux anomalies in tropical regions are strongly correlated with ENSO related climate variations while temperate and boreal regions are only weakly correlated. The correlation coefficients between CO2 flux anomaly for subcontinental size regions (aggregated) and climate indices or meteorological parameters are given in Table 1. The climate index second in importance after ENSO is the IOD index; its relation on CO2 flux anomaly is apparent (high correlation) in the neighboring land regions of the Indian Ocean, particularly Africa, tropical Asia, and Australia, and extends as far as tropical South America. In the northern high latitudes, we observe the NAO/AO to be the dominant controlling factor for the European CO2 flux anomaly, likely through the temperature variation because of strong anticorrelation of flux anomaly with SAT for Europe (Table 1). The warmer weather in these regions may lead to negative CO2 flux anomaly since that condition stretches the growing season length [see Dye and Tucker, 2003]. It should be pointed out here that the climate indices or other anomaly time series used for our analysis have not been detrended before calculating the correlations.

3.6. Principal Component Analysis of Flux Anomaly

[15] To analyze the factors controlling CO2 fluxes further, we have performed a principal component analysis (PCA) of CO2 flux anomalies. The PCA procedure is briefly described here (see Wilks [1995] for further details). The eigenvectors (em) of the correlation matrix (S of dimension 64 times 96) of CO2 flux anomaly (x′) for 64 TDI regions are called the empirical orthogonal functions (EOFs) since they form a complete set of orthogonal basis to represent the CO2 flux anomaly. The EOFs are plotted as maps by associating each component with its corresponding spatial location as shown in Figure 6. If the eigenvalue corresponding to the mth eigenmode is λm, then the percentage variance associated with that mode is given by (λm/equation image λi) × 100, which are shown at the top of each panel. The PCs for each modes of variability (um) can be calculated as um = equation imageeimxi.

Figure 6.

EOF distributions of first three most dominant modes of land CO2 flux variabilities. The percentage of flux variability captured by each mode is given in titles of each panel.

[16] The three most dominant EOF distributions are shown in Figure 6, which confirms that the tropical regions in South America, Africa, and Asia are most variable in terms of CO2 flux anomaly (Figures 6a and 6b). Larger absolute values suggest higher level of interannual variation in the fluxes (x′) over that location, and the signs denote the relative phase of variation in one region with respect to the other. The first two modes jointly capture about 27% of the global flux variance. The correlation coefficients of first PC (most dominant mode) with the ENSO and IOD are 0.88 and 0.80, respectively, while that for the second PC with ENSO and AO exhibited highest values of R(= −0.59 and 0.54) (Table 1). Note that the correlations of AO with CO2 flux anomaly for temperate South America were also estimated high (Table 1), and that region is indicated as the location of largest variation in CO2 fluxes (see EOF-2). Thus, as seen from the EOF distributions (Figures 6a and 6b) of first two dominant PC modes, the regions of maximum variability are located in the tropics. The PC-3 showed high correlations with AO and PDO, which captures 7.1% of CO2 flux variability. PC-3 and AO correlation increased further (R = 0.85) when the winter AO indices are used (for DJFM months only). The DJFM NAO index and PC-3 correlation is also estimated high, R = 0.89. The EOF distribution corresponds well with the common understanding of AO/NAO influences on Europe, eastern part of North America, and some areas in Asia (Figure 6c). During the positive phases of AO/NAO, northern Europe and eastern USA observe wetter and warmer winter [Thompson and Wallace, 2001; Hurrell et al., 2003], while east Asian summer monsoon rainfall weakens [Gong et al., 2001].

4. Conclusions

[17] The results obtained for CO2 fluxes from 42 land regions, using a time-dependent inverse model and surface CO2 observations at 87 stations around the globe, have been discussed for the period 1994–2001. Our results are overall in good agreement with previous estimates of global land-atmosphere flux variability. The global anomalous CO2 flux from land biosphere to atmosphere during the 1997/1998 El Niño period is as large as 84% of the total CO2 emission from all industrial sources in that period according to our inversion. The regional CO2 flux anomaly estimates exhibit significant correlation, with the air temperature and rainfall anomalies as the major drivers of heterotrophic respiration and photosynthesis, respectively. However, the largest factor for enhanced CO2 emission over a short period from the land regions (tropical, temperate, and boreal) seems to be related to the widespread biomass burning or forest fire events. The Biome-BGC ecosystem model predicts less CO2 flux variability than the TDI model. This may be due to the lack of adequate processes to initiate biomass burning. The detailed correlation study and principal component analysis suggest that the CO2 flux anomaly in tropical regions (e.g., tropical South America, Africa, tropical Asia, and part of Australia) are controlled dominantly by climate variability related to ENSO in agreement with numerous previous studies, followed by the IOD influences. For Europe and temperate South America, we found a significant relation between AO and CO2 flux anomalies.


[18] ATSR World Fire Atlas is obtained from European Space Agency–ESA/ESRIN via Galileo Galilei, CP 64, 00044 Frascati, Italy. We appreciate the support of Hajime Akimoto for this research. The transport and inverse model simulations are performed using the Earth Simulator and SX-5 systems (NEC), respectively. We wish to thank the reviewers for constructive comments and suggestions as well as helping us with the language to improve clarity.