Interannual variability of carbon monoxide emission estimates over South America from 2006 to 2010



[1] We present the first inverse modeling study to estimate CO emissions constrained by both surface and satellite observations. Our 4D-Var system assimilates National Oceanic and Atmospheric Administration Earth System Research Laboratory (NOAA/ESRL) Global Monitoring Division (GMD) surface and Measurements Of Pollution In The Troposphere (MOPITT) satellite observations jointly by fitting a bias correction scheme. This approach leads to the identification of a positive bias of maximum 5 ppb in MOPITT column-averaged CO mixing ratios in the remote Southern Hemisphere (SH). The 4D-Var system is used to estimate CO emissions over South America in the period 2006–2010 and to analyze the interannual variability (IAV) of these emissions. We infer robust, high spatial resolution CO emission estimates that show slightly smaller IAV due to fires compared to the Global Fire Emissions Database (GFED3) prior emissions. South American dry season (August and September) biomass burning emission estimates amount to 60, 92, 42, 16 and 93 Tg CO/yr for 2006 to 2010, respectively. Moreover, CO emissions probably associated with pre-harvest burning of sugar cane plantations in São Paulo state are underestimated in current inventories by 50–100%. We conclude that climatic conditions (such as the widespread drought in 2010) seem the most likely cause for the IAV in biomass burning CO emissions. However, socio-economic factors (such as the growing global demand for soy, beef and sugar cane ethanol) and associated deforestation fires, are also likely as drivers for the IAV of CO emissions, but are difficult to link directly to CO emissions.

1. Introduction

[2] The Amazon is the largest tropical rain forest region in the world and hosts a large diversity of flora and fauna [Sala et al., 2000]. Intact forests are suggested to be a net carbon sink [Malhi et al., 2008] and appear much more resilient to drought conditions than damaged forests. However, large parts of the forest are currently at risk due to climate change and land-use changes. Tropical forest clearing (deforestation) is often associated with biomass burning leading to higher greenhouse gas emissions [Hoffmann et al., 2002; Malhi et al., 2008]. The deforestation process may also have important consequences for 1) the regulation of regional [Nobre et al., 1991] and global climate [Shukla et al., 2000; Bala et al., 2007], 2) biogeochemical cycles [Houghton, 2006] and 3) maintenance of soil fertility [Davidson et al., 2007].

[3] Currently deforestation mainly takes place in the arc-of-deforestation along the southern and eastern borders of the Amazon [Morton et al., 2006]. Intact forest is cleared and converted to agricultural lands for cattle grazing or crop production. 55% of Brazil is still covered by forest, 35% is used as pasturelands for ranching and only 7% for agriculture (soy and corn using half of this) [Rovere et al., 2011].

[4] Another potential thread to the Amazon is the expansion of sugar cane plantations. Sugar cane ethanol is used as a fuel (in combination with gasoline in so-called flex-fuel cars) in Brazil and this country is also the largest exporter of sugar cane ethanol in the world. Global demands for sugar cane ethanol are expected to rise enormously in the coming decades to meet lower carbon dioxide (CO2) emissions [Rovere et al., 2011; Walter et al., 2011; Cardoso et al., 2012]. However, this higher demand calls for more agricultural land and hence, expansion of sugar cane plantations might lead to additional deforestation.

[5] Fire is the main clearing tool in the deforestation process: in the states of Mato Grosso and Pará in Brazil, highly mechanized land clearing is done by large landholders [van der Werf et al., 2009]. The remaining biomass is piled up and ignited multiple times until the biomass is completely burned and large plantations can be started up.

[6] The biomass burning fires emit large amounts of carbon in the form of CO2, carbon monoxide (CO) and other gases and aerosols into the atmosphere. Enhanced trace gas concentrations can be observed by space-borne satellite instruments. Large interannual variability (IAV) in aerosol loading has been identified by Torres et al. [2010] for the last decade. Torres et al. [2010] found that the high aerosol loads correlated well with fire counts and were inversely related to precipitation. The link between fire activity and precipitation was also found by Aragão et al. [2008], who also reported the peak deforestation month to be 3 months before the peak fire month. Interannual variability in SCIAMACHY (SCanning Imaging Absorption spectroMeter for Atmospheric CartograpHY) CO was reported by Gloudemans et al. [2009] for the period 2003–2007. The years 2004, 2005 and 2007 showed the largest CO columns, likely associated with enhanced biomass burning emissions compared to the years 2003 and 2006.

[7] The IAV in carbon emissions from biomass burning have been estimated by van der Werf et al. [2010] for the period 1997 to 2009. Estimates of emission factors were used to convert carbon emissions to trace gas emissions for CO, CO2 and other gases. This database was released as the Global Fire Emission Database (GFED) version 3 inventory. Some recent studies reported that the previous GFED2 product was too low over South America. For example, Kopacz et al. [2010] used an inverse modeling approach and found that GFED2 was too low over South America by 50–100% for 2004. In contrast, Hooghiemstra et al. [2012], found that GFED3 (already lower than GFED2) was slightly too high to fit MOPITT observations in 2004 over this region. Hence, the uncertainty in the bottom-up biomass burning inventory remains large and a reliable top-down method to quantify the emissions is needed. CO is a perfect tracer of these pyrogenic emissions due to its relatively long lifetime in comparison with volatile organic compounds and aerosols emitted during fires and its short lifetime relative to CO2 and methane. With a lifetime of approximately two months, CO can be traced as it travels between continents [Gloudemans et al., 2006; Edwards et al., 2006].

[8] Earlier inverse modeling studies focussing on CO used only surface observations (mainly from the NOAA network) to constrain the emissions [Bergamaschi et al., 2000; Kasibhatla et al., 2002; Pétron et al., 2002; Pison et al., 2009; Hooghiemstra et al., 2011]. These measurements are performed with high accuracy and precision and the measurement time series go back to the early 1990's [Novelli et al., 1998, 2003]. However, the location of the surface stations is biased toward the Northern Hemisphere (NH) and particularly the Tropics are poorly constrained by these observations. Moreover, the temporal resolution of the NOAA flask observations is typically weekly and hence these observations might miss short term pollution events. Most recent inverse modeling studies therefore use satellite derived CO columns [e.g., Arellano et al., 2004; Pétron et al., 2004; Arellano et al., 2006; Stavrakou and Müller, 2006; Chevallier et al., 2009; Jones et al., 2009; Kopacz et al., 2009, 2010; Fortems-Cheiney et al., 2011]. Satellite observations have the great benefit that they provide global coverage, typically within a couple of days. Hence the much greater spatial and temporal resolution of these observations over the Tropics and the SH can be used to constrain emissions better in these regions. However, satellite derived CO column measurements come with substantial uncertainties and biases. As a result, a biased (satellite) observation may result in artificial emission increments in an inverse modeling exercise. For example, CO emissions as derived in current satellite-driven inverse modeling studies, show large discrepancies in the remote SH compared to NOAA surface flask observations [Arellano et al., 2006; Kopacz et al., 2010; Fortems-Cheiney et al., 2011; Hooghiemstra et al., 2012]. Ideally, one would like to use the surface observations to anchor the inversion and the satellite observations to constrain CO emissions on a high resolution.

[9] The goal of this study is to infer robust monthly total CO emission estimates for South America at a high spatial resolution of 1° × 1° for the period 2006–2010. These years are in particular interesting because the Amazonian drought of 2010 [Lewis et al., 2011] is included in this time window and because deforestation rates have decreased compared to the 2000–2005 period. Moreover, we present the first study in which surface and satellite observations of CO are jointly assimilated in a four-dimensional variational (4D-Var) data assimilation system to optimally constrain the CO emissions. To accomplish this, a bias correction scheme had to be implemented.

[10] The outline of this study is as follows: section 2 describes the inversion methodology, the atmospheric chemistry and transport model TM5, the observations and the bias correction scheme. In section 3 we describe our main results for the 2006–2010 optimized emissions over South America, which are further discussed in section 4. We conclude our findings in section 5.

2. Inverse Modeling

[11] The 4D-Var approach starts with the definition of the cost function inline image(x) (equation (1)) to be minimized.

display math

where inline image is the optimal state vector that, when simulated by the model H, fits the measurements yi (with observation error covariance matrix Ri) while staying close to the prior state xb (with prior error covariance matrix B). M is the number of measurements and ⊺ the transpose operator. The cost function is minimized iteratively using the CONGRAD method [Fisher and Courtier, 1995], which is based on the conjugate gradient method [Hestenes and Stiefel, 1952] and the Lanczos algorithm [Lanczos, 1950]. This algorithm produces the optimal posterior state as well as the N leading eigenvalues λi and eigenvectors νi of the Hessian of the cost function, which can be used to estimate the posterior error covariance matrix A [Fisher and Courtier, 1995]:

display math

where inline image As in our previous study, we use the TM5-4D-Var system for CO [Hooghiemstra et al., 2012]. The main changes applied to the system for this study are described below.

2.1. TM5 Model

[12] The atmospheric chemistry and transport model TM5 [Krol et al., 2005] and its adjoint [Meirink et al., 2008a] are used to simulate CO mixing ratios given a set of emissions and to calculate the sensitivity of the emissions to a set of model-data mismatches, respectively. In this study, the parallelized version of TM5 has been used which heavily reduced the runtime of the model (by a factor 4 with respect to our previous work). Model transport is driven by 3-hourly meteorological fields (6-hourly for 3-D fields) from the European Centre for Medium-Range Weather Forecasts (ECMWF). Here, the ERA-Interim meteorological fields have been used, on a subset of 25 of the originally 60 hybrid ECMWF layers. The TM5-4D-Var system uses a climatological OH field [Spivakovsky et al., 2000] which is scaled by a factor 0.92. This factor is based on methylchloroform simulations for the period 2000–2006 [Huijnen et al., 2010]. Although a coarse global model resolution (6° × 4°) is used, the nested zoom capability of TM5 is exploited over South America. Via an intermediate zoom region with a resolution of 3° × 2°, the model resolution is refined to 1° × 1° over a part of South America. The zoom region covers the Amazon forest and is bounded by the coordinates (30°S, 78°W) in the lower left corner and by (10°N, 36°W) in the upper right corner (see Figure 1).

Figure 1.

The nested grid of TM5. The gray lines represent the coarse 6° × 4° resolution, the red lines the intermediate 3° × 2° zoom region and the blue lines correspond to the finest zoom region with a resolution of 1° × 1°. The locations of the NOAA stations Ascension Island (ASC), Easter Island (EIC) and Tierra del Fuego (TDF) are also given as well as the validation station Arembepe, Bahia, Brazil (ARB).

2.2. Prior State and Uncertainties

[13] The 5 year inversion period (2006–2010) is split into five inversion periods. Since we are specifically interested in the South American biomass burning season (July October), each inversion starts on April 1 and ends on December 31. The first 3 months (April, May and June) are used as spin-up, the last two months (November and December) are used to constrain in particular the emissions in October. Hence, although the emissions for the spin-up and spin-down periods are optimized, these numbers should be interpreted with care since they may be influenced by for instance the starting CO field. The starting CO mixing ratio field on April 1 is constructed by a forward simulation for three months starting at January 1 with a climatological CO mixing ratio field from a previous multiyear forward simulation.

[14] In contrast with our previous work [Hooghiemstra et al., 2012], we only optimize the total CO emissions, on a monthly resolution. This is done because the posterior correlations in a three-category inversion are highly negatively correlated on monthly timescales. In particular the biomass burning and natural + NMVOC CO sources can hardly be distinguished by the system (resulting in a posterior correlation coefficient of −0.8 for the monthly emissions). In our previous inversions results were aggregated over a yearly timescale. This allowed for a clearer separation due to the differences in seasonal cycle of the two emission categories. Although we do not optimize biomass burning emissions separately, the biomass burning CO emissions are the dominant CO source in the dry season over South America and we assume that the IAV in total CO emissions can be largely attributed to biomass burning CO emissions.

[15] The prior emissions are the sum of a fossil fuel and biofuel component, a biomass burning component and a biogenic source (direct CO emissions from vegetation and the ocean, plus the production of CO from NMVOC oxidation). We have used the Emissions Database for Global Atmospheric Research (EDGARv4.1, compiled for the year 2004, EDGAR Project Team [2010]) for the fossil fuel and biofuel source (462 Tg CO/yr). For biomass burning the GFED3 [van der Werf et al., 2010] inventory was used (in which biomass burning CO emissions are 34, 86, 21, 7 and 109 Tg CO for both August and September for 2006 to 2010, respectively over the South American 1° × 1° zoom region). For the natural source we use the emissions as described in Hooghiemstra et al. [2012] (515 Tg CO/yr). Production of CO from methane oxidation is not optimized in our system. Instead we use optimized methane 3-D mixing ratio fields [Bergamaschi et al., 2009; S. Houweling et al., manuscript in preparation, 2012], which are oxidized in the model using a CO-yield of 1.0 and yielding another 865 Tg CO/yr on the global scale.

[16] Monthly emissions are optimized globally, that is, emissions in the South American zoom region are optimized on a 1° × 1° resolution and emissions in regions with a coarser resolution are optimized on either a 3° × 2° or a 6° × 4° resolution. However, since we specifically focus on South America in this study we only report the emissions for the 1° × 1° zoom region over South America.

[17] The prior grid-scale errors are defined as 50% of the corresponding grid-scale monthly emissions. This relatively tight error setting was chosen because small errors are needed to obtain a balanced cost function in a system that assimilates both surface and satellite observations. Although with a grid-scale error of 50% the background part of the cost function (first term in equation (1)) is still roughly one order of magnitude smaller compared to the observational part of the cost function (second term in equation (1)), the weight of the background part of the cost function has increased significantly compared to a grid-scale error of 250% that we used in our previous work. Especially for inversions in which we assimilate both surface and satellite observations combined with a bias correction (see section 2.4), small errors appeared necessary to prevent overfitting of the data and aliasing between emissions and the bias correction.

[18] To decrease the effective number of variables to be optimized in the system, spatial and temporal correlations are proposed in the prior error covariance matrix B. We use a Gaussian spatial correlation length of 100 km. This means that the prior emissions in two adjacent grid boxes (in the 1° × 1° zoom region) have a correlation coefficient of 0.29. Similarly, we use an e-folding temporal correlation length of 3 months which is equivalent to a month-to-month correlation coefficient of 0.72.

2.3. Observations

[19] With the current setup of the 4D-Var system, either surface flask observations from the National Oceanic and Atmospheric Administration Earth System Research Laboratory (NOAA/ESRL) Global Monitoring Division (GMD) can be assimilated or CO columns derived from space-borne satellite instruments, or both. Independent observations used to validate the optimized emission estimates are also described here.

2.3.1. Assimilated Observations

[20] Surface flask observations from 34 NOAA background sites are assimilated in the 4D-Var system. We use exactly the same list of stations as in our previous study [Hooghiemstra et al., 2012], but the data (up to 2010) have been updated (downloaded from in November 2011) and show slight differences with respect to the data we used in our earlier work. The reason for the change of data is a change in the calibration of the gas chromatograph used for measuring CO mixing ratios in the NOAA laboratory in Boulder, Colorado (P. C. Novelli, personal communication, 2011). For most observations, the differences are very small, but for remote sites in the SH, like the South Pole station, the new calibration results in higher CO mixing ratios up to 10% in particular in Austral summer when CO mixing ratios are low. Available NOAA stations in the vicinity of South America are shown in Figure 1.

[21] In contrast with our previous work, modeled CO mixing ratios are no longer temporal averages over 3 hours. Instead, modeled CO mixing ratios are instant model values at the time of the observation. The observation error for the flask measurements consists of a fixed measurement error of 3 ppb (1.5 ppb for remote SH stations to give these measurements more weight in the cost function) and a model representativeness error. The latter is slightly different compared to our previous work and is solely based on modeled 3-D CO mixing ratio gradients.

[22] CO total columns from the Measurement of Pollution in the Troposphere (MOPITT) instrument are assimilated in the system. MOPITT was launched in 1999 onboard of NASA's Terra satellite. In this study we use the MOPITT version 4(V4) total column product [Deeter et al., 2010] which is based on the MOPITT retrieval of CO in the thermal-infrared (TIR) at a wavelength of 4.7 μm [Deeter et al., 2003]. These columns are mainly sensitive to CO in the free troposphere. As described in Deeter et al. [2010], the MOPITT V4 product suffers from a temporal bias drift. The CO total column drifts of by about 1 ppb/yr and may be due to long-term changes in the modulation cells or other instrumental parameters. Although such a temporal bias might influence the results of an inversion, we believe that the effect on posterior emission estimates will be small.

[23] However, there is currently also a MOPITT V5 product based on retrieved CO columns in both the TIR and the shortwave-infrared (SWIR at 2.3 μm), which is also sensitive to CO in the boundary layer. This product is not used in the current study because it typically comes with larger random errors and is not yet fully validated. Here we assimilate MOPITT V4, Level 3, daytime observations which are gridded on a 1° × 1° resolution. In contrast to our previous study, in which only MOPITT ocean pixels were assimilated, here we also use MOPITT over South America to obtain better coverage specifically over this region during the biomass burning season and to pick up specific biomass burning events.

[24] We linearize the MOPITT V4 averaging kernel as before [Hooghiemstra et al., 2012] to keep the model linear and the cost function quadratic such that the CONGRAD method is a feasible minimizer for the cost function. No additional corrections for the linearization are used in this study. The linearized MOPITT averaging kernels are applied to model columns but in the current study we interpolate the (coarse) model columns to the location of the satellite observations, which are provided on a 1° × 1° grid.

2.3.2. Validation Observations

[25] Independent observations of CO are used to validate the optimized emissions. CO flask observations are performed by the Instituto de Pesquisa Energéticas e Nucleares (IPEN), Laboratorio de Química de la Atmosferíca (LAQAT) at station Arembepe, located at the east coast of Brazil (12.77°S, 38.17°W, see Figure 1) from 2006 onwards (no data for 2010 released yet). Since this station is most likely sensitive to biomass burning CO in the period August–October, simulated mixing ratios will be compared to these observations in section 3.3.

[26] The Infrared Atmospheric Sounding Interferometer (IASI) onboard the MetOp satellite (launched in October 2006) also measures CO columns at a wavelength of 4.7 μm [Clerbaux et al., 2009]. IASI has a somewhat smaller footprint and a larger swath width compared to MOPITT. This results in nearly daily global coverage by this instrument. Due to the smaller footprint, IASI observations are less often masked by clouds and hence pick up even more specific biomass burning events over South America. The altitude sensitivity specified by the averaging kernel for IASI is comparable to MOPITT and IASI is thus mainly sensitive to free tropospheric CO. Here we use IASI day time total column observations, retrieved using the FORLI (Fast Optimal Retrievals on Layers for IASI) optimal estimation algorithm [Turquety et al., 2009; George et al., 2009]. Data for the August-October period for the years 2007–2010 over the 1° × 1° zoom region are used for validation.

2.4. Bias Correction

[27] Most recent CO inverse modeling studies used satellite observations as constraints for the emissions. However, due to possible biases in both atmospheric transport models and satellite retrieved CO columns, simulations with optimized emissions are significantly overestimating observed surface CO mixing ratios in the remote Southern Hemisphere [Arellano et al., 2006; Kopacz et al., 2010; Fortems-Cheiney et al., 2011; Hooghiemstra et al., 2012]. Since not all biases in satellite observations are fully understood at the moment, we attempt to use the surface observations as an anchor point in the inversion. Ideally, one combines the assimilation of surface observations (to obtain a correct background CO distribution) with assimilation of satellite observations to constrain the emissions in the Tropics and other regions where only a few surface observations are available.

[28] Hooghiemstra et al. [2012] found that in particular for the Tropics and the SH, surface flask and satellite based inferred emission estimates are inconsistent, at least in the TM5 transport model. In this study, we assimilate both surface and satellite observations jointly in the 4D-Var system and in addition fit a set of bias parameters to obtain realistic emissions that are in agreement with both observational data sets. For inverse modeling of methane using surface network observations and SCIAMACHY columns, Meirink et al. [2008b] and Bergamaschi et al. [2009] fitted a polynomial bias correction along the latitude direction, based on 4 bias parameters. Here, we propose a bias correction scheme with one parameter for each degree in latitude because the bias is in particular apparent on the remote SH. Since we assimilate MOPITT CO columns between 65°N and 65°S only, we use a set of 130 bias parameters. In order to obtain a smooth bias correction and to reduce the number of effective bias correction parameters, the parameters are correlated in space using a correlation length of 5° (correlation coefficient of 0.85 for two parameters 2° apart). A rather tight prior error of 0.5 ppb per parameter is chosen in order to prevent aliasing (in which the bias correction is used to compensate for unphysical emission changes). The bias parameters are set to zero a priori. For each MOPITT observation (gridded on 1° × 1°) we subtract the value of the bias parameter corresponding to the latitude of the observation, from the MOPITT column-averaged CO column. Hence, a positive (negative) bias parameter decreases (increases) the corresponding MOPITT column.

3. Results

[29] In this section we present the results of the base inversion for the period 2006–2010. The following paragraphs discuss the fit of the prior and posterior modeled CO mixing ratios with the assimilated observations and the resulting bias correction. Next, we analyze the interannual variability of CO mixing ratios and emissions over South America. This section concludes with the validation of the derived emissions with independent observations.

3.1. Fit With Assimilated Observations

3.1.1. NOAA Surface Stations and the Bias Correction

[30] The prior (yellow) and posterior (blue) simulation at 3 NOAA stations in or close to South America are plotted in Figure 2 together with the South Pole results. The NOAA flask observations are shown as open symbols and the error bars denote the 1-σ observation error. A scatterplot of modeled versus observed CO mixing ratios is also shown for each station. For station Ascension Island, the highest CO mixing ratios are observed from August to October mainly due to biomass burning in Africa and South America. Interannual variability in observed CO mixing ratios is rather small, indicating that enhanced CO over Ascension Island is mainly due to emissions on the African continent, where the IAV in biomass burning CO is less pronounced compared to South America [Torres et al., 2010; van der Werf et al., 2010].

Figure 2.

Prior (yellow) and posterior (blue) simulated mixing ratio for 4 NOAA stations that have been assimilated. NOAA flask observations (including a 1-σ error) are shown in black. The inset shows the fit with the data as scatterplots for the prior and posterior simulations. The posterior simulation corresponding to a sensitivity inversion for 2010 without bias correction is shown in red. For South Pole stations the fit with the observations typically deteriorates near the end of the year since the inversions all finish at December 31.

Figure 2.


[31] For station Easter Island, the prior simulation typically underestimates observed CO mixing ratios by 5–10 ppb (up to 20 ppb in the dry season). After assimilation the posterior fit improved, but some high mixing ratios in summer are not fitted well. This is likely caused by the coarse resolution of the model outside the zoom region over the ocean [Hooghiemstra et al., 2011].

[32] Station Tierra del Fuego represents the remote SH. The prior simulation underestimates NOAA flask observations in the last part of the year. The posterior simulation increases the modeled mixing ratios by 5–10 ppb, from July onwards. For this station (as for Easter Island station) we do observe IAV, likely due to transport from biomass burning CO emissions. The fit with the measurements remains remarkably good for the posterior simulation even though we assimilated MOPITT columns also in our inversions. Previous inverse modeling studies using MOPITT to constrain the emission estimates reported large differences when comparing their posterior simulations with NOAA observations in the remote SH [Arellano et al., 2006; Kopacz et al., 2010; Fortems-Cheiney et al., 2011; Hooghiemstra et al., 2012]. In the inversions presented here, we fitted a bias correction to the MOPITT columns and used the station observations as anchor points for the inversion. Even observed background CO mixing ratios at South Pole station are nicely fitted by the inversion, in sharp contrast with previous inversions that overestimated observed mixing ratios at the South Pole by 20 ppb. A sensitivity inversion assimilating both datasets without fitting a bias correction is shown in red for the year 2010 only. Clearly, the assimilation of the NOAA surface stations already brings the modeled CO mixing ratios close to the observations compared to earlier satellite only inversions, but the added effect of the bias correction (see below) improves the fit even more.

[33] The bias correction obtained for the years 2006–2010 is shown in Figure 3. The positive bias on almost the entire SH may be associated with a retrieval bias in very clean environments. Note that the value of a bias parameter is subtracted from the MOPITT columns at the corresponding latitude. Hence, with a positive bias on almost the entire SH, MOPITT columns are reduced by the inversion by a maximum of 5 ppb for the latitude range 30°–50°S. This implies that optimized CO emissions for the South American continent are somewhat lower in our approach compared to a MOPITT-only inversion. Given that the 5 year period is treated as 5 separate inversions, the derived bias is remarkably robust for the period 2006–2010. Assuming a column-averaged CO mixing ratio between 50 and 100 ppb in the SH, we infer a positive bias in MOPITT V4 of 5 to 10% in the SH. This is significantly lower compared to the bias in MOPITT V3 in the SH (up to 20%), which was deduced from comparison with aircraft profiles [Emmons et al., 2007, 2009]. In contrast, the NH Tropics show a small negative bias (meaning that MOPITT is too low according to the inversion) of 3 to 5 ppb. At NH mid and high-latitudes, the inferred bias is close to zero.

Figure 3.

Bias correction for the joint inversions for 2006–2010. Note that the bias (in ppb) is subtracted from the MOPITT column-averaged CO mixing ratio. Therefore, a positive bias (leading to smaller observed columns) typically leads to a reduction in the emissions compared to no bias.

3.1.2. Comparison of Prior and Posterior Simulation with MOPITT

[34] Figure 4 shows monthly composites of the prior and posterior modeled CO columns and the assimilated MOPITT observations (with the bias correction applied to these observations) for August 2008 (top) and 2010 (bottom). This figure clearly illustrates the power of the 4D-Var approach: although we optimize monthly emissions only, and fire emissions typically change on shorter timescales, the fine scale pollution events (due to fires) are fitted very well. For 2008 the agreement mainly improves over the northeast part of the zoom region where background CO levels are increased from the prior to the posterior simulation. For 2010, however, the prior simulation largely overestimates observed columns in Western Brazil, Bolivia and parts of Paraguay. The posterior simulation (with reduced emissions) is indeed much closer to the observed columns.

Figure 4.

Prior and posterior modeled column-averaged CO mixing ratios and MOPITT observations for August (top) 2008 and (bottom) 2010. Monthly average wind fields at 600 hPa are shown in the middle panels. The observations have been corrected with the posterior bias (Figure 3). Note the different scales between the top and bottom panels.

[35] Although the main biomass burning emission hot spots in the GFED3 inventory are located in the states of Mato Grosso and Pará (in particular for 2010), satellite-observed CO hot spots are mainly visible more westward in Rondônia and southwards to Bolivia. This feature is due to transport of biomass burning pollution plumes to the west and southwest, where CO accumulates at the foot of the Andes mountain range. The westerlies then transport CO enhanced air southeastward to the Atlantic ocean. These transport patterns are also shown in Figure 4 (middle). A different pathway may also transport CO enriched air further to the west toward the Pacific ocean.

[36] Previous inverse modeling studies showed difficulties in fitting both surface and satellite observations over the mid- and high latitudes of the SH [Arellano et al., 2006; Kopacz et al., 2010; Fortems-Cheiney et al., 2011; Hooghiemstra et al., 2012]. Since we specifically fit both surface and satellite observations using a bias correction scheme it is important to analyze the fit with MOPITT observations in the remote SH. Figure 5 (top) shows global prior and posterior model-data differences defined as MOPITT-TM5 for August 2010. The large prior model underestimate over the tropical Atlantic Ocean is significantly reduced in the posterior model simulation as well as the large overestimate over much of the South American continent. Note that the fit with the data is also significantly improved on the NH. Figure 5 (bottom) shows frequency plots of the model-data mismatch for the rectangular regions shown in the corresponding upper panels. By design of the system the fit with the data improves drastically for the 1° × 1° zoom region. The prior underestimate in the rectangular box corresponding to the remote SH peaks at 5 ppb and is nearly reduced to 0 in the posterior simulation, likely due to the inferred bias correction.

Figure 5.

Model-observation differences defined as TM5-MOPITT for (top left) the prior simulation and (top right) the posterior simulation for August 2010. (bottom) A frequency diagram of prior (yellow) and posterior (blue)differences corresponding to the rectangles in the top panels. Note that MOPITT observations have been corrected for with the optimized bias.

3.2. Interannual Variability of CO Emissions Over South America

3.2.1. Time Series of Column-Averaged CO Mixing Ratios

[37] Interannual variability of CO mixing ratios is clearly visible in Figure 6, which shows modeled and observed column-averaged CO mixing ratios for 4 regions in and close to South America. For example, the MOPITT observed columns (dashed red lines) show peak values over the Amazon of 200 ppb in 2007, but the maximum in 2009 is below 100 ppb. For the South American outflow regions IAV is also observed, albeit less pronounced compared to the continental region. Similar IAV is modeled using the prior emissions (yellow line) and is caused by the IAV in the biomass burning emissions of GFED3. The prior simulation largely underestimates observed CO columns (except for 2010), in particular for the Atlantic outflow regions. The prior underestimate outside the biomass burning season (April to June and November and December) indicates that the background CO levels due to anthropogenic and natural CO emissions are too low. An underestimation of biomass burning emissions in the GFED3 prior may also play a role. Note that the dashed green lines represent the bias corrected MOPITT columns, which deviate from the uncorrected MOPITT columns only over the Southern Atlantic. The bias correction reduces the MOPITT columns by approximately 5 ppb at these latitudes (see Figure 3).

Figure 6.

Five-year time series for four regions in and close to South America. Prior (yellow), posterior (blue) and observed (red dashed and green dashed) column-averaged CO mixing ratios are shown for the period April–December for each year. A bias correction (per latitude, in ppb) is fitted to the observed MOPITT columns. The prior bias is zero for all latitudes, the posterior bias is non-zero (further illustrated in Figure 3) and subtracted from the original MOPITT columns leading to the green dashed line. Modeled and observed columns have been averaged over 10 day periods and the individual data points are indicated in the prior simulation by rectangular markers. Note the data gap in August and September of 2009 due to a malfunctioning cooler of the MOPITT instrument.

[38] If we zoom in on smaller regions (illustrated in Figure 7), similar IAV is observed as compared to the large regions. However, for region 1 (in the arc-of-deforestation in the states of Mato Grosso and Pará in Brazil [Morton et al., 2006]), the prior simulation overestimates the observations in 2007 and in particular in 2010, in which the Amazon suffered an extreme drought [Lewis et al., 2011; Xu et al., 2011]. Extremely dry years usually trigger larger biomass burning events due to extensive drying of the biomass [Aragão et al., 2008] and hence, peak CO columns >300 ppb are simulated. The overestimation in these two years may indicate that GFED3 emissions are locally somewhat too high in these years.

Figure 7.

As in Figure 6, but for small regions in South America. Region 1 covers parts of the states of Mato Grosso and Pará, both located within the arc-of-deforestation [Morton et al., 2006], region 2 is the Brazilian Cerrado (mainly savanna and grasslands) and region 3 is the state of São Paulo.

[39] The posterior simulation (shown in blue), fits the bias corrected observed columns (dashed green lines) very well, especially when averaged over large regions as shown in Figure 6. For the smaller regions in Figure 7, the fit with the observations is also remarkably good given the constraint that emissions may vary only on monthly timescales. However, for the Brazilian Cerrado (region 2 in Figure 7), the posterior simulation still underestimates the observations by up to 20–40 ppb. This illustrates the compromise in the 4D-Var approach: the posterior simulation tries to fit the available observations as good as possible but deviations from the prior emissions also result in a penalty in the cost function (equation (1)). In addition, the number of observations available for each region (which are given per month in Table 1) plays a role here. The low number of observations in 2009, in particular for the small regions results in emissions that remain close to the prior estimate.

Table 1. Number of Gridded (1° × 1°) MOPITT Observations per Month per Regiona
  • a

    The eight regions that have been used in this paper are the 1° × 1° zoom region (see Figure 1), the four large regions (Figure 6) and the three small regions (Figure 7). Note that due to a cooler malfunction, the MOPITT instrument was turned off from July 28 to September 28, 2009. For comparison, the number of IASI observations (also binned to 1° × 1°) for August 2010 for region 1 equals 32523 and for region 8 IASI reported 942 observations.


3.2.2. Optimized Emissions

[40] The total prior and posterior emissions (including errors), aggregated per month for the 1° × 1° zoom region are shown in Figure 8. As expected, the total emissions are increased for all months (except for August in 2007 and 2010).

Figure 8.

Prior and posterior total monthly CO emission estimates for the South American 1° × 1° zoom region from 2006 to 2010. Note that emissions for the period January to March have been omitted since our inversions ran from April to December. Error bars denote the prior and posterior 1-σ uncertainty in the emission estimates.

[41] For the dry season, total emissions are increased except for the year 2010. We assume here that the IAV in dry season total CO emissions can be attributed entirely to the IAV in biomass burning CO emissions. Biomass burning CO emissions (for both August and September) are therefore calculated by subtracting a constant background from the total CO emissions summed over August and September for each year in the period 2006–2010. The background term is the sum of anthropogenic and natural CO emissions corrected for additional biomass burning emissions outside the dry season (estimated as 1 Tg CO/month). By inspection of the posterior CO emissions outside the dry season (Figure 8), we estimate the background emission term to be 7 Tg CO/month. The resulting IAV of biomass burning CO emissions is detailed in Table 2. The dry season of 2007 and 2010 show very high biomass burning emissions of 92 and 93 Tg CO, respectively. In contrast, in 2006 and 2008, 60 and 42 Tg CO is emitted and in 2009 (a very wet year), emission estimates are as low as 16 Tg CO. In comparison with the prior emission inventory from GFED3, the IAV has reduced. This is indicated by the smaller spread in the inferred emission estimates compared to the prior (see bottom row of Table 2 ). Using prior emissions that are constant from year to year (i.e., start with zero IAV, see Appendix A and column 4 of Table 2), the IAV in inferred emissions is further reduced, confirming that the prior information in these inversions remains important. The aggregated posterior errors are significantly smaller (uncertainty reduction up to 60%) than the prior errors in particular in the dry season, indicating that the assimilated observations constrain the 1° × 1° region well. The posterior errors and corresponding sensitivity studies are further discussed in Appendix A.

Table 2. Calculated Interannual Variability in Biomass Burning CO, Total CO, and Surface Emissionsa
YearBiomass Burning CO (Corrected)Total COSurface Emissions
August + SeptemberApril - DecemberApril - DecemberYear
PriorPosteriorPriorPosteriorPriorPosteriorPosteriorFortems-Cheiney et al. [2011]
  • a

    Calculated interannual variability in biomass burning CO emissions for August and September together in Tg CO (columns 2–4). Interannual variability in biomass burning emissions for the complete inversion period of 9 months (columns 5–7). Total CO (from anthropogenic, biomass burning and the natural source, including NMVOC-CO) is shown in columns 8 and 9. Finally, an estimate of yearly emissions of anthropogenic (ANT) and biomass burning (BB) emissions are given for our study (column 10) and the numbers reported by Fortems-Cheiney et al. [2011] (column 11). We have estimated our annual surface emissions by subtracting the NMVOC-CO for the 9-month period (total of 48 Tg CO), adding the prior anthropogenic and biomass burning emissions for 3 months (totaling to 7 Tg CO) and the annual contribution of direct emissions from vegetation (12 Tg CO/yr).

  • b

    The term GFED-mean refers to the prior biomass burning emissions calculated from GFED3 by taking the 5-year mean emission estimate. Prior biomass burning emissions for these inversions are set to 51 Tg CO (for August and September) and 63 Tg CO for the April - December period.

Mean ± STD52 ± 3961 ± 3068 ± 1565 ± 4298 ± 34103 ± 19128 ± 42161 ± 34132 ± 34149 ± 29

[42] For 2006 to 2010, the derived spatial patterns of the total CO emissions for August and September are shown in Figure 9. Figures 9 (top) and 9 (middle) present the prior and posterior total CO emissions, respectively. Figure 9 (bottom) shows the percentage emission increment, relative to the prior emissions. Two-month total CO emissions (in Tg CO) aggregated over the 1° × 1° zoom region are given in brackets. Focusing on the arc-of-deforestation, the region in northern Mato Grosso and southern Pará (region 1 in the top left panel), emissions in this region increase compared to the prior in 2006 and 2008 and to a lesser extent in 2009. In contrast, large reductions in emissions for this region are inferred for 2007 and 2010. Since the biomass burning CO emissions dominate the prior spatial emission pattern (not shown), we attribute these increments to under- (2006, 2008 and 2009) and overestimates (2007 and 2010) in the GFED3 inventory. A second region of interest is the Brazilian Cerrado (from the north- to the southeast, region 2). Here, the inversions increase emissions for all years and the increments are likely attributed to too low emissions in GFED3 possibly combined with too low background emissions from both the natural (NMVOC) source and the anthropogenic source. A third region of interest is the state of São Paulo (region 3). CO emissions in this region are associated with pre-harvest burning of sugar cane plantations. Despite the fact that by accident the sugarcane emissions from both the GFED and the EDGAR inventories were included in the prior emissions and the prior estimate was therefore too high, the inversion further increased the emissions by 50–100% for all years. Given the growing demand for sugar cane ethanol as a fuel and as an export product [Cardoso et al., 2012], it is likely that both prior inventories underestimate these emissions for the period 2006–2010.

Figure 9.

Prior emissions, posterior emissions and increments of the total CO emissions for August and September for the years 2006–2010 in g CO (2 month)−1 m−2. Numbers in the titles represent the two monthly total emission (Tg CO) for the 1° × 1° South American zoom region.

3.3. Validation With Independent Observations

[43] Independent observations, i.e. observations that have not been assimilated in the 4D-Var system, are used to validate the inferred emission estimates. Figure 10 shows the comparison with independent CO flasks from station Arembepe, Brazil: The prior simulation for 2007 to 2009 (in yellow) and the posterior simulation (in blue) are shown as well as the flask observations. A scatterplot for all observations and for the dry season observations is presented in Figure 10 (bottom). The comparison with the flask observations improves in particular in the dry season, for which the prior simulation was too low in 2007–2009. The posterior simulation shows excellent agreement with peak CO observations due to pollution plumes which are well captured by the high resolution model.

Figure 10.

Validation with independent data from station Arembepe, Bahia, Brazil for the period 2007–2009. This station has not been assimilated in the inversion. (top) The prior and posterior simulations are shown in yellow and blue, respectively. Flask observations are denoted by open symbols. (bottom) Scatterplots of modeled versus observed mixing ratios for (left) all observations and (right) observations in the dry season, defined as the months August to October. Numbers in the panels present the statistics for the bias, root mean square (RMS) and the correlation coefficient for the prior and posterior simulation, respectively.

[44] Prior and posterior model simulations are also compared to IASI columns. Figure 11 shows the prior (Figure 11, left) and posterior (Figure 11, middle) modeled CO total column in 1018 molecules cm−2 (based on our joint MOPITT/NOAA inversion) and the IASI total columns (Figure 11, right) for August, September and October 2010. The model is sampled with collocated IASI observations and the IASI averaging kernels are used for a proper comparison. The comparison with IASI columns confirms that the prior emissions were too high in August and September 2010. Hence, the system reduces the emissions and the posterior simulation shows very good agreement with IASI columns. For October 2010 however, there is poor agreement with IASI columns. This can be explained by the low amount of data available from MOPITT for this month compared to the previous months (see Table 1, compare the number of MOPITT observations for September and October 2010 for regions 2 and 6). This was also found for August and September 2009 (not shown). In 2009, no MOPITT measurements were available from July 28 to September 28 and emissions were only constrained by observations from October onwards.

Figure 11.

Prior and posterior modeled total column CO (in 1018 molecules/cm2) and IASI observations for August to October 2010.

4. Discussion

[45] The resulting emission estimates from application of the high resolution 4D-Var system presented so far show significant differences compared to the GFED3 prior inventory and strong year-to-year variation. Below we discuss potential explanations for differences found between our emission estimates and the GFED3 prior and we discuss possible drivers of the inferred interannual variability in CO emissions.

4.1. GFED3

[46] For the period 2006 to 2010 significant differences between our emission estimates and the GFED3 prior inventory were shown. We acknowledge that since we can not separate the individual CO emission categories in the 4D-Var framework, our derived biomass burning CO emissions (as given in Table 2) carry some uncertainties. However, deficiencies in the GFED3 product also play a role. The GFED3 CO emission inventory [van der Werf et al., 2010] for biomass burning is based on four quantities: burned area, fuel loads, combustion completeness and emission factors, all bearing their own uncertainties. We discuss the impact of these uncertainties below:

[47] 1. In this study we have used the GFED3 emissions on a monthly resolution. In reality, these fires vary probably on a day-to-day basis. In our current framework, we can only optimize grid-scale emissions on a monthly time-scale and hence, CO emissions due to multiple fire events in a single grid box are averaged over the whole month.

[48] 2. GFED3 uses constant emission factors throughout the year per biome. Our inversion indicated that high fire years were overestimated, and low fire years underestimated in GFED3. One potential reason could be that these emission factors vary over time and space, something that is not taken into account in GFED3. During drought years, burning efficiency is higher and the amount of CO released per unit biomass consumed is lower. In the future, building dynamic emission factors into GFED [van Leeuwen and van der Werf, 2011] that would dampen interannual variability may thus explain part of the discrepancies found. This reasoning, however, does not explain the shift in peak emissions from August to September.

[49] 3. Another source of uncertainty in this region relates to the combustion completeness (CC). In GFED3, CC is modeled as a function of climatological conditions and a fire persistence metric based on active fire observations as a proxy for the amount of repeated burning [Morton et al., 2008].

[50] 4. The same fire persistence metric is used to boost burned area estimates [Giglio et al., 2010] in deforestation regions to better mimic reported deforestation rates [van der Werf et al., 2010]. Both the CC and burned area formulation in GFED are thus crude, and better comparisons against ground measurements is required to understand the underlying reasons behind the differences between bottom-up and top-down estimates.

4.2. IAV in Biomass Burning Emissions

[51] We have inferred significant IAV in biomass burning emission estimates for the period 2006–2010. Year-to-year variability in precipitation combined with IAV in deforested area is likely the main driver of IAV in biomass burning CO emissions. For example, during an anomalous wet year such as 2009 [Torres et al., 2010], biomass burning due to deforestation may be postponed because of high fuel moisture. This would typically lead to a reduction in CO emissions in that year but the emissions associated with that specific biomass debris will be added to the biomass burning CO emissions of the next year. Such a mechanism could for example explain our inferred CO emissions estimates for 2009 and 2010. In contrast, in very dry years such as 2007 and 2010 [Lewis et al., 2011], biomass burning fires may easily run out of control due to the extensively dried vegetation leading to high CO emissions. In addition, smoldering ground fires release large amounts of CO to the atmosphere. Torres et al. [2010] found high correlations between OMI derived aerosol loads, precipitation and fire counts, concluding that the IAV in aerosols from 2001–2009 was due to biomass burning. The year 2008 showed very low aerosol loads and fire counts but was certainly not an above average wet year. Therefore, Torres et al. [2010] concluded that perhaps economic or regulatory factors were the main drivers of the low fire activity and aerosol load in 2008. The inferred CO emissions from our 5 year inversion show a similar correlation with the rainfall anomalies as the aerosol observations presented by Torres et al. [2010] in which the year 2008 stands out for its low emissions. The IAV in inferred CO emission estimates was further compared to CO emission estimates reported by Fortems-Cheiney et al. [2011]. Overall, the IAV in their CO emissions agrees quite well with our findings, although their estimates are somewhat larger. This comparison is further detailed in Appendix B.

[52] IAV in deforested areas may be linked to the global demand for goods such as beef and soy and their associated prices. For example, Laurance [2007] described how a global increase in the price of soy has an impact on deforestation in Amazonia. The higher global demand for soy directly triggers forest conversion to soy plantations in the Amazon region. Either intact forest is cleared or pasturelands from ranchers are purchased and converted for crop production. Cattle ranchers in turn are pushed into the forest frontier clearing new forest for pasturelands. Governmental legislation also influences deforestation. For example, the ‘Forest Law’ (active since 1965), requires protection of 80% of the forest of private lands in the Amazon. This law is currently about to change, which may lead to more deforestation in the coming years [Sparovek et al., 2010]. Moreover, the ‘Forest Law’ only protects the Amazon region whereas the Brazilian Cerrado is unprotected and converted rapidly to agricultural lands. Deforested area in the Amazon region is closely monitored from space by the Brazilian institute for Space Research (INPE) and shows a steadily decreasing trend in deforested area since 2004 (with only one small increase in 2008). In contrast, although our inferred CO emission estimates decrease from 2007 onwards (Table 2), CO emissions returned to high levels in 2010.

[53] Top-down estimates of biomass burning emissions as presented in this paper would be a valuable addition to existing methods to monitor deforestation (INPE, GFED). However, it is difficult to link deforestation one-on-one to CO emissions for the following reasons. First, since not all deforested biomass may be burned in the year of deforestation, piles of biomass debris may be ignited several times in the years after the deforestation took place [van der Werf et al., 2009] and hence, emit much more CO than expected based on the deforested area. Second, besides deforestation fires, also savanna and woodland fires in the Brazilian Cerrado contribute to the total CO emissions from biomass burning. In the future, a closer collaboration between bottom-up approaches to understand and partition fire sources and top-down work as presented here to constrain total emissions may yield better insights into fire processes and the quantitative effect of biomass burning on the composition of the atmosphere.

4.3. Expansion of Sugar Cane Plantations

[54] Our inversions strongly increase CO emission estimates over São Paulo state, likely associated to pre-harvest burning of sugar cane plantations. With a growing world wide demand for sugar cane ethanol as a fuel and associated expansion of sugar cane plantations, the question is whether this leads to additional deforestation. Some studies claim that most new sugar cane plantations are started on lands previously used for cattle ranching or crop growing and therefore do not lead to more deforestation [Walter et al., 2011; Rovere et al., 2011]. However, other studies expect that the expansion of sugar cane plantations have an indirect effect on deforestation by pushing ranchers into the forest frontier [Lapola et al., 2010]. The benefit of avoiding greenhouse gas emissions by using biofuels instead of fossil fuels may be offset by enhanced greenhouse gas emissions associated with deforestation. However, with proper regulation, it is suggested that the growing demand for sugar cane ethanol can be achieved without additional deforestation if strategies for cooperation between ranchers and plantations are formed [Lapola et al., 2010]. Although pre-harvest burning of sugar cane will be phased out by 2021 (in São Paulo state), high resolution satellite CO observations during the coming decade in combination with inverse modeling, may be useful to identify land-use changes, by monitoring pre-harvest fires on sugar cane plantations.

5. Summary and Conclusions

[55] In this study we have performed a 5-year inversion for CO, specifically focusing on the biomass burning season in South America for 2006 to 2010. The inversions were performed in 9-month periods from April to December using the TM5-4D-Var system. Although the inversions optimized CO emissions on a global scale, we used the nested zooming capability over South America to resolve the emissions in this region on a high spatial model resolution of 1° × 1°. To our knowledge, this is the first study that assimilated column-averaged CO mixing ratios from MOPITT over the oceans and the South American continent in combination with NOAA surface flask observations. A bias correction scheme was used to obtain posterior emission estimates that were consistent with both observational data sets. The main conclusions of this study are listed below:

[56] 1. Assimilation of MOPITT and NOAA observations in combination with the bias correction leads to a significantly improved fit with the surface observations compared to a MOPITT-only inversion. The resulting bias correction indicates a robust positive bias in MOPITT V4 in the SH up to 5 ppb (in column-averaged CO mixing ratio) between latitudes 30°S and 50°S and a slightly negative bias in the NH Tropics in all years of the assimilation.

[57] 2. South American total CO emissions show large interannual variability over the period 2006–2010 and this IAV is most likely driven by biomass burning CO emissions. A priori IAV in the GFED3 inventory is somewhat reduced by the inversion. The dry season (August and September) biomass burning CO emissions over the South American zoom region for the 5 year inversion are 60 (34), 92 (86), 42 (22), 16 (7) and 93 (109) Tg CO/yr for 2006 to 2010, respectively (GFED3 prior emission estimates are shown between brackets). These total emissions are found to be robust as a sensitivity inversion with 5 year mean GFED3 emissions resulted in emission estimates within 20% except for 2009.

[58] 3. The biomass burning inventory GFED3 seems to estimate too high emissions for 2007 and 2010 in the arc-of-deforestation in Brazil. However, emissions in the savanna regions in eastern Brazil tend to be underestimated. For the other years of this study, the inferred emission estimates are higher than GFED3.

[59] 4. For the entire 2006–2010 period, CO emissions are increased in São Paulo state by 50 to 100%.These emissions, probably from pre-harvest burning of sugar cane plantations seem to be underestimated in current inventories.

[60] 5. Validation with independent observations from IASI gives a very good comparison for months in which enough MOPITT data is available for assimilation. However, large discrepancies are observed for months which were not constrained by MOPITT data (e.g., October 2010), for which the posterior emission estimate stays very close to the prior.

[61] 6. The interannual variability in derived biomass burning CO emissions does not agree with the steady decrease in deforestation in Brazil since 2006 as reported by INPE. Although a decrease in deforestation as monitored from space may lead to less biomass burning CO emissions, climatic influences (in the form of droughts) may have very large effects as can be seen in the large increase in biomass burning CO emissions in 2007 and 2010.

[62] In combination with a state-of-the-art prior emission set, combined assimilation of accurate and high-precision flask measurements and satellite observations with global coverage has been shown to be an effective way to obtain high resolution CO emission estimates for South America.

[63] Our study period of 5 years is likely too short to draw conclusions on long-term trends in CO emissions over South America. In the coming decade it will become clear if the Brazilian government is able to reduce deforestation by 2020 to 20% of the average 1996–2005 levels [Nepstad et al., 2009]. With the expected expansion of sugar cane and soy plantations for the large-scale production of biofuels, the question remains how this will influence deforestation and carbon emissions in this region. In this light, top-down estimates as presented in this paper may provide additional and independent information about deforestation and associated biomass burning emissions.

[64] For future inverse modeling studies, we recommend to optimize biomass burning emissions on a sub-monthly temporal resolution, but only if combined with high resolution satellite observations. For example the IASI instrument results in a denser set of observations compared to MOPITT. Assimilation of the MOPITT V5 product, which exploits both the thermal and short-wave infrared channels of the MOPITT instrument may be useful since it provides sensitivity to surface CO, where the emissions occur. The upcoming TROPOMI mission (planned for launch in 2015) will also measure CO in the short-wave infrared with high resolution and is expected to be an additional valuable source of information on CO.

Appendix A:: Sensitivity Studies and Error Analysis

[65] In this appendix we discuss the effect of different prior emission settings and the biomass burning injection height. In addition, the achieved uncertainty reduction on the scale of the 1° × 1° zoom region and the grid-scale is discussed.

A1. Large and Small-Scale Uncertainty Reduction

[66] The base inversion for 2006–2010 shows large interannual variability in biomass burning emissions, with only 16 Tg CO in 2009 and up to 93 Tg CO in 2010 (Table 2). The robustness of the derived interannual variability of the total emissions is analyzed by repeating the same 5-year inversions, but using a 5-year GFED3 mean as prior biomass burning emission. Hence, the prior biomass burning emission in a certain box is the average of the GFED3 emissions in this box for the 2006 to 2010 period. The GFED3 prior emission for both August and September in these inversions amounts to 51 Tg CO. The derived biomass burning CO emissions from these sensitivity inversions for August and September are shown in Table 2. As in the base inversion, the years 2007 and 2010 show the highest emissions (83 and 88 Tg CO/yr respectively), but approximately 10% lower compared to the base inversion. The other years show higher biomass burning CO emissions, in particular 2009, in which we derive 51 Tg CO due to biomass burning. This is explained by the fact that there are no MOPITT measurements from July 28, 2009 up to September 29, 2009 and hence, the 2009 biomass burning season is poorly constrained by the observations. However, comparison of the posterior base simulation (16 Tg CO biomass burning CO) with IASI (not shown) showed much better agreement compared to this sensitivity inversion, indicating that the derived emission estimate of 51 Tg CO in 2009 is based on the prior emission estimate due to the poor coverage of MOPITT for this period. Overall, the total annual emissions in the zoom region are rather close to the base inversion, indicating that the observations constrain the emissions well on this spatial scale.

[67] Although Figure 4 gives convincing evidence that the 4D-Var system is indeed capable to optimize monthly emissions on a high resolution that fit MOPITT columns well, the question remains how much this fit depends on the prior assumptions we made. For example, how important is the spatial pattern of the biomass burning emissions we use in the prior? To investigate this issue, we performed an additional inversion for 2010 using uniform biomass burning emission instead of GFED3. The uniform prior has the same annual total biomass burning emissions as GFED3, but the emissions per grid box are constant throughout the year. Therefore, the total biomass burning emissions in the dry season are significantly lower than GFED3. For all sensitivity studies described in this appendix, the prior and posterior emissions for August and September 2010 and the increments are illustrated in Figure A1. Figure A1 (bottom) shows scatterplots of the prior and posterior fit with MOPITT columns over the South American zoom region.

Figure A1.

Prior, posterior and incremental total CO emissions for August and September for the year 2010 in g CO (2 month)−1 m−2 for a series of sensitivity inversions. Numbers in the titles represent the two monthly total emission (Tg CO) for the 1° × 1° South American zoom region. The bottom row plots show scatterplots of modeled versus observed CO columns for the prior (yellow) and posterior (blue) simulation.

[68] For the uniform prior, the 4D-Var system increases the total emissions in the dry season over South America from 57 Tg CO to 91 Tg CO, which is still 20% lower than the base inversion for 2010 (resulting in 107 Tg CO). Moreover, the spatial patterns for this inversion are completely different as there are no emission hot spots in the posterior emission estimate. The reason for the derived uniform emissions is 1) the lack of emission hot spots in the prior and 2) our methodology for setting the prior errors as a percentage of the corresponding emissions. Hence, for a spatially heterogeneous prior such as GFED3, the system adjusts regions with high emissions rather than regions with low emissions to fit the observations. For a spatially homogeneous prior this is not the case and the whole emission field is scaled to fit the observations. The fit with MOPITT columns is compared in the scatterplots (Figure A1, bottom). In particular the high CO columns observed by MOPITT (due to pollution events) are significantly underestimated. This sensitivity study thus clearly illustrates the need of a realistic prior emission inventory like GFED. The result indicates that on the grid-scale the observations put a poorer constrain on the emissions compared to the scale of the 1° × 1° zoom region. This is in agreement with a much smaller grid-scale uncertainty as shown in Figure A2. Largest uncertainty reduction is achieved in regions where 1) prior emissions and hence, prior errors are large and 2) a large amount of MOPITT observations constrain the emissions. As discussed in [Meirink et al., 2008a], the mean grid-scale uncertainty reduction is indeed much smaller compared to the aggregated uncertainty reduction.

Figure A2.

Grid-scale uncertainty reduction for the base inversion aggregated over August and September 2010 and the total number of MOPITT observations per grid box.

A2. Uncertainty in Natural Emissions

[69] The publicly available POET database consists of high resolution (1° × 1°) inventories of several precursors of CO (e.g., isoprene, terpenes, methanol and acetone) as well as inventories of CO emissions from plants and the ocean. We constructed a prior natural emission inventory based on POET using the following yields for CO production [Duncan et al., 2007]: 0.2 for isoprene and terpenes, 0.67 for acetone and 1 for methanol.

[70] The sensitivity inversion starting from the POET based natural CO emissions results in somewhat larger total CO emissions in both the prior (139 Tg CO versus 123 Tg CO in the base) and the posterior emission estimate (118 versus 107 Tg CO). From the increment plot, it is observed that apart from a reduction of biomass burning CO emissions in the arc of deforestation, the emissions in the Northwestern part of the Amazon are slightly reduced, indicating that the POET based natural CO emissions might be too high over this region. This can be explained by our assumption of releasing all CO precursors directly as CO, neglecting their atmospheric lifetimes. For the longer lived precursors (like acetone), this results in a CO overestimate.

A3. Biomass Burning Injection Height

[71] Apart from assumptions for the prior emissions, model uncertainties also play a role. Recently, Jiang et al. [2011] and Hooghiemstra et al. [2012], showed that in particular the choice of OH field may lead to large differences on both regional and global scales for the inferred emissions. Here we focus on the vertical distribution of the biomass burning emissions. As in our previous study, the biomass burning emissions in the base inversions are released below 2 km. However, what would be the effect on the inferred emissions if we release a fraction of the emissions above the boundary layer? One would expect that those emissions are observed by the satellite earlier which may lead to somewhat lower emission estimates. For this sensitivity study we use an injection height that is defined as follows: 65% is released uniformly below 2 km. Additionally, 25% is released between 2 and 3 km and the remaining 10% is released uniformly between 3 and 6 km altitude. This inversion was performed for the year 2010 and the spatial pattern of the inferred emissions and the total CO emission were very close to the base inversion (not shown). Hence the different injection height has almost no impact on the inversion. The explanation is visualized in Figure A3. In Figure A3 (left), the prior modeled CO mixing ratios at different vertical model layers are shown for the base and the sensitivity simulation for a small area dominated by biomass burning emissions. In the sensitivity simulation relatively more CO is injected at higher altitudes and hence, the surface CO mixing ratios decrease in comparison with the base simulation (yellow and green markers). Similarly, due to enhanced injection of CO around 700 hPa in the sensitivity simulation, CO mixing ratios are increased compared to the base simulation (blue markers). However, for typical altitudes at which MOPITT is most sensitive, there are no significant differences in the modeled CO mixing ratios for both injection heights (red and cyan markers)). Hence, the modeled column-average CO mixing ratios in Figure A3 (right) are very close to 1:1 line when either the MOPITT averaging kernel is taken into account or not. We acknowledge that for specific situations in which long term accumulation of CO in the boundary layer due to favorable meteorological conditions is suddenly vented out to the free troposphere, the injection height may indeed play a role. In addition, for a Short Wave Infrared (SWIR) satellite instrument (such as SCIAMACHY, the MOPITT V5 product or the upcoming Tropospheric Ozone Monitoring Instrument (TROPOMI)), with sensitivity to the surface, a different vertical distribution could have effect on optimized CO emissions because these satellite instruments typically view the whole column.

Figure A3.

A priori modeled CO mixing ratios for September 2010 over a small region in Mato Grosso. (left) CO mixing ratios for specific vertical model layers for the base injection height (horizontal axis) versus the altered injection height (vertical axis). The legend indicates the pressure corresponding to the model layers in hPa. (right) Column-averaged CO mixing ratios before (red) and after (green) application of the MOPITT averaging kernels. The black dashed line represents the 1:1 relation.

Appendix B:: Comparison With Recent Work

[72] Our inferred IAV of biomass burning emissions for South America is compared to the posterior emissions derived from a 10 year MOPITT inversion by Fortems-Cheiney et al. [2011]. They reported total CO emissions (from fossil fuel and biofuel combustion and biomass burning) for the South American Temperate region, which covers South America south of the Equator. However, we only inverted for the period April to December and hence, we have to make an estimate of the emissions in the period January to March. Also, we have included the NMVOC-CO source in our total emission estimate. In an effort to make a fair comparison we subtract the prior NMVOC-CO emissions for April to December (48 Tg CO) and add the prior anthropogenic and biomass burning emission for January to March (7 Tg CO) and the annual contribution of vegetation for this region (12 Tg CO) to our posterior emission estimates.

[73] The comparison of our results with reported CO emission estimates is shown in the last two columns of Table 2. In general our estimates are lower by 10 to 20 Tg CO/yr and even 52 Tg CO lower in 2009. The large 2009 difference is likely caused by the different prior emissions for biomass burning that were used. Fortems-Cheiney et al. [2011] used GFED2 which does not supply biomass burning emission estimates for 2009 and Fortems-Cheiney et al. [2011] used the 2008 prior estimate. However, the GFED3 prior estimate is significantly lower leading to our posterior estimate of 80 Tg CO in 2009. The remaining discrepancies are likely due to other differences in the inversion setup such as the OH field, the exact treatment of the NMVOC-CO emissions, our inclusion of a bias correction scheme and different spatial model resolution. The interannual variability of the total emissions show similar patterns in which the year 2007 really stands out.


[74] This research was supported by the Dutch User Support Programme 2006–2010 under project GO-AO/05. The Dutch National Computer Facility (NCF) is acknowledged for computer resources. We would like to acknowledge the IASI team and the French Ether database for providing the ULB/LATMOS IASI CO product, and we would like to thank in particular M. George for helpful discussion on the IASI averaging kernels. We are most thankful to S. Basu for help during the model development phase. We are thankful to S. Myriokefalitakis for supplying the a priori fields of NMVOC-CO production. We also thank S. Houweling for optimized methane mixing ratio fields.