1 Top-down constraints on 2007 and 2010 South America fire Carbon loss: 2 supplementary information. 3 4 1. MOPITT CO emissions adjoint inversion 5 6 Monthly inverse estimates of CO fluxes are performed at 4°×5° using the NASA Carbon 7 Monitoring System flux tools; these are based on MOPITT CO data [Worden et al., 8 2010], the GEOS-Chem model and a 4D-Var assimilation approach [Jiang et al., 2014]. 9 In this approach, we minimize the cost function of the form, 10 11 π½(π±) = (π (π±) β π²)π ππΊβπ (π (π±) β π²) + (π± β π± π )π ππβπ (π± β π± π ) 12 13 where x is the state vector of CO emissions, y is a vector of observed concentrations, and 14 F(x) is the forward model which represents the transport of the CO emissions in the 15 GEOS-Chem model and accounts for the vertical smoothing of the MOPITT retrieval, 16 π± π is the a priori estimate, and ππΊ and ππ are the observational and a priori error 17 covariance matrices, respectively. The cost function is minimized using the adjoint of 18 GEOS-Chem model in a 4D-Var approach [Henze et al., 2007]. 19 20 Total CO emission estimates are partitioned into anthropogenic, biomass burning and 21 biogenic emissions based on the relative contribution of prior emission estimates from 22 each category. We anticipate that CO emission partitioning errors are small, given that 23 biomass burning CO emissions are (a) spatially distinct from anthropogenic (non-biomass 24 burning) CO emission sources, and (b) approximately an order of magnitude greater than 25 biogenic CO emissions during the 2007 and 2010 fire years. To determine CO emissions 26 within the study area we (i) re-grid the CO emission estimates at a 1°×1° resolution and 27 (ii) integrate across grid-points within the study area (25°S β 5°S, 90°W β 30°W). 28 29 2. TES CH4/CO 30 31 To quantify large-scale fire emissions CH4-to-CO ratios [Worden et al., 2013b], we use 32 the TES Lite Products version 7 (tes.jpl.nasa.gov/data/) to determine CH4-to-CO slope 33 (CH4/CO) within the study area (25°S β 5°S, 90°W β 30°W) on a monthly basis. We 34 determine mean column CH4 and CO concentrations and their associated uncertainties 35 based on the Worden et al. [2013b] CH4 & CO measurement selection criteria. We use a 36 weighted total least squares algorithm to quantify the CH4/CO slope and associated 37 uncertainties [Krystek & Anton, 2007]. 38 39 To distinguish between the regional fire plume and background air masses within our 40 study area, we derive CH4/CO based on co-located CH4 and CO values where CO > 41 110ppb (i.e., CO cutoff = 110ppb). We examine the sensitivity of selecting a CO cutoff = 42 110ppb by imposing cutoff values ranging from 50ppb to 150ppb. We show the resulting 43 CH4/CO values in Figure S1. The results presented henceforth and in the main text are 44 converted from volumetric CH4/CO [ppm (CH4) ppm-1 (CO)] to mass CH4/CO [g (CH4) 45 g-1 (CO)]. 46 47 CH4/CO values where the Pearson correlation is insignificant at a 99% level (pvalue >0.01) 48 are excluded from our analysis. For the 2007 and 2010 fires, at a cut-off of CO = 110ppb 49 CH4/CO values are significant in September 2007 (CH4/CO = 0.0669±0.0083), October 50 2007 (CH4/CO = 0.0936±0.0117) and September 2010 (CH4/CO = 0.1093±0.0116). For a 51 ±10 ppb change in the prescribed CO-cutoff value (i.e., 100ppb β€ CO cutoff β€ 120ppb, 52 see Figure S1) the CH4/CO value standard deviations are 0.0054, 0.0042 and 0.0092, 53 respectively. 54 55 3. Monte Carlo Bayesian inversion approach 56 57 For each species s and land-cover type b, the fire-season emissions πΉπ ,π are estimated as 58 follows: 59 60 πΉπ ,π = π΄π πΆπ΅π·π πΈπ ,π (1) 61 62 where π΄π , πΆπ΅π·π and πΈπ ,π are the burned area, combusted biomass density and emission 63 factor (see main text). We refer to the 18 unknown parameters (i.e., πΆπ΅π·π and πΈπ ,π , for s 64 = [CH4, CO], b = [savannas, forest and agriculture], and y = [2007, 2010]) as parameter 65 vector x: π¦ π¦ 66 67 2007 2007 2007 2007 2007 2007 2007 2007 2007 x = [πΆπ΅π·π ππ£. , πΆπ΅π·πππ. , πΆπ΅π·ππππ. , πΈπ ππ£.,πΆπ , πΈπ ππ£.,πΆπ»4 , πΈπππ.,πΆπ , πΈπππ.,πΆπ»4 , πΈπππ.,πΆπ , πΈπππ.,πΆπ»4 , 68 2010 2010 2010 2010 2010 2010 2010 2010 2010 πΆπ΅π·π ππ£. , πΆπ΅π·πππ. , πΆπ΅π·ππππ. , πΈπ ππ£.,πΆπ , πΈπ ππ£.,πΆπ»4 , πΈπππ.,πΆπ , πΈπππ.,πΆπ»4 , πΈπππ.,πΆπ , πΈπππ.,πΆπ»4 ] (2) 69 70 According to Bayesβ theorem, the probability of x given the observations O, p(π±|π) can 71 be described as: 72 73 p(π±|π) β p(π|π±) × p(π±) (3) 74 75 where p(π|π±) is the likelihood of x given O (or the probability of O given x) and p(π±) is 76 the prior probability of π± (Smith et al., 2013). The observations consist of MOPITT 77 derived CO emissions and TES CH4/CO: specifically we use (a) the mean May-Dec 78 2007, May-Dec 2010 CO emissions, (b) the normalized difference between May-Dec 79 2007 and May-Dec 2010 CO emissions, and (c) TES CH4/CO ratios in 2007 and 2010 80 (see table 1). We assume no covariance between these observational constraints, 81 therefore we define p(π|π±) (the likelihood of the parameters given the observations O) as 82 follows: 83 84 p(π|π±) = exp (β0.5 βπ π=1 (Mπ βOπ )2 Ο2π ) (4) 85 86 where for N observations, ππ and ππ are the nth observation and corresponding model 87 estimate (based on equation 1 in main text) and ππ2 is the variance of ππ [e.g., Ziehn et 88 al., 2012]. 89 90 Quantifying uncertainties (ππ2 ) associated with MOPITT CO fluxes remains a 91 computationally challenging task when using adjoint inversion approach (see section S1) 92 to estimate surface fluxes. We assign a conservative estimate of 20% for the accuracy 93 and 5% for the precision in the CO emissions, consistent with the CO emissions estimates 94 shown in Worden et al. [2013a]. The 5% precision estimate is one of the key 95 uncertainties driving the conclusions in this paper, as the conclusions are insensitive to 96 the CO emissions accuracy. We expect that a 5% precision in the normalized difference 97 between 2010 and 2007 CO emissions is a conservative estimate, as the change in 98 emissions are driven by the observed change in CO: within the study area (25°S β 5°S, 99 90°W β 30°W), the observed change in CO has a <2% precision on the standard error of 100 the mean (2 ppb) for 2007 and 2010 peak fire season values of 147 ppb and 131 ppb 101 respectively (figure 1 in main text). We determine the sensitivity of our results to the 102 prescribed uncertainty estimates (i.e. 20% accuracy, 5% precision) in section 4. TES 103 CH4/CO uncertainty estimates are based on the weighted least-squares CH4/CO slope 104 uncertainty (see section 2). The observations and the corresponding model estimates are 105 summarized in Table 1. We quantify the sensitivity of our approach to each uncertainty 106 estimates in section 4. 107 108 We define the prior probability of x, p(π±), as follows: 109 110 p(π±) = pEF (π±) × pCF (π±) × pΞECF (π±) (5) 111 112 where pEF (π±) is the prior probability of the emission factors, pCF (π±) is the prior 113 probability of the combustion factor (discussed below), and pΞECF (π±) is the prior 114 probability of the 2007-to-2010 change in emission and combustion factors. The 115 uncertainty terms in the prior probability of x consist of (a) emission factor uncertainty; 116 (b) combustion factor uncertainty; and (c) the uncertainty in the 2007-to-2010 change in 117 emission and combustion factors. 118 119 We characterize pEF (π±) based on the Andreae and Merlet [2001] emission factors for 120 forests, savannas and agriculture: to ensure a positive-definite distribution for all CH4 and 121 CO emission factor priors, we approximate the mean and standard deviation of Andreae 122 and Merlet [2001] CH4 and CO emission factors (in agriculture, forests and savannas) as 123 a log-normal probability distribution: 124 2 125 log(πΈπΉ(p) ) β log(πΈπΉ0(π) ) pEF (π±) = exp (β0.5 × β12 π=1 ( log(1+ ππΈπΉ 0(π) ) πΈπΉ0 (π) ) ) (6) 126 127 where πΈπΉ(π) represents the pth emission factor in x (equation 2), and EF0 (π) , ππΈπΉ0 (π) 128 represent the corresponding emission factor prior and uncertainty. We use the EF0 (π) and 129 ππΈπΉ0 (π) values reported under the βSavanna and Grasslandβ, βTropical Forestsβ and 130 βAgricultural Residueβ categories by Andreae and Merlet [2001]. We note that a log- 131 normal distribution is preferable to truncating the pEF (π±) at zero: fire conditions with low 132 CH4 and CO emission factors become increasingly unlikely as combustion efficiency 133 approaches 1 (i.e. 100% combustion efficiency). 134 135 We impose an upper limit on πΆπ΅π·π based on the above-ground carbon density π΅π within 136 burned area π΄π . The combustion factor [CFb units: kg (C combusted) kg-1 (C)] can be 137 expressed as 138 139 CFb =πΆπ΅π·π /π΅π . (7) 140 141 We use the Saatchi et al. [2011] above-ground biomass density map to determine π΅π 142 within each monthly MODIS burned area Ab. Although CFb is naturally limited between 143 0<CFb<1, we note that (a) the Saatchi et al. [2011] biomass map is temporally fixed 144 (circa 2003); (b) Saatchi et al. [2011] biomass density pixel-scale errors are typically 30- 145 50% within the study area; and (c) small but significant carbon losses can also occur from 146 the below-ground biomass pools. Moreover litter pools (not included Saatchi et al. 147 [2011]) typically amount to a smaller but significant carbon stock compared to total 148 biomass in savannas and forest ecosystems [e.g., Ward et al., 1996; Prasad et al., 2001; 149 Malhi et al., 2009]. We therefore allow CFb (i.e. the combustion factor relative to Saatchi 150 et al. [2011] Bb) to range between 0 and 2, and pCF(x) is defined as: 151 152 pCF (π±) = 1 if 0 < π΅πΆπ·π π΅π < 2, otherwise pCF (π±) = 0. (8) 153 154 For the sake of simplicity, we optimize CFb instead of CBDb/π΅π (and later multiply CFb 155 by Bb to obtain CBDb). Parameter vector x (equation 1) consists of all unknown 156 parameters (combusted biomass density and emission factors for each land-cover type, in 157 2007 and 2010: based on equations 2 and 7, we convert combusted biomass density to 158 combustion factor, therefore x is now defined as: 159 160 2007 2007 2007 2007 2007 2007 2007 2007 2007 2010 x = [πΆπΉπ ππ£. , πΆπΉπππ. , πΆπΉππππ. , πΈπ ππ£.,πΆπ , πΈπ ππ£.,πΆπ»4 , πΈπππ.,πΆπ , πΈπππ.,πΆπ»4 , πΈπππ.,πΆπ , πΈπππ.,πΆπ»4 , πΆπΉπ ππ£. , 161 2010 2010 2010 2010 2010 2010 2010 2010 πΆπΉπππ. , πΆπΉππππ. , πΈπ ππ£.,πΆπ , πΈπ ππ£.,πΆπ»4 , πΈπππ.,πΆπ , πΈπππ.,πΆπ»4 , πΈπππ.,πΆπ , πΈπππ.,πΆπ»4 ] (9) 162 163 To avoid un-realistic changes in πΆπΉπ and πΈπ ,π between 2007 and 2010 within each land- 164 cover type b, we define pΞECF (π±) as: 165 π±π,2007 166 pΞECF (π±) = exp (β0.5 × β9π=1 ( log( π±π,2010 log(π) ) 2 ) ) (10) 167 168 where π₯π,2007 are all 2007 parameters in x, π₯π,2010 are their 2010 counterparts (see 169 equation 9), and f denotes the uncertainty in 2007-to-2010 change in xp. Specifically, the 170 pΞECF (π±) term dictates a 68% probability that each 2007 parameter is within a factor of f 171 of its 2010 counterpart, and a >99% probability that each 2007 parameter is within a 172 factor of 2f of its 2010 counterpart (for example, if f = 1.5, then there is a 68% 173 probability that 174 of these parameters. However, given the seasonal range of reported combustion factor 175 and efficiency measurements [e.g., Korontzi et al., 2003; Hély et al., 2003], we anticipate 176 that f = 2 (i.e. a 1-sigma probability for a factor of 2 change in CFb and Es,b between 2007 177 and 2010) is a conservative constraint on inter-annual variations in fire characteristics. 1 π±π,2007 <π± 1.5 π,2010 < 1.5). We are not aware of any inter-annual measurements 178 179 We use a Metropolis Hastings Markov Chain Monte Carlo (MHMCMC) algorithm to 180 draw 2 ×105 samples of x from p(π|π±). Markov Chain Monte Carlo algorithms have 181 been widely used to solve ecological parameter optimization problems [Hurtt and 182 Armstrong, 1996, Braswell et al., 2005, Ziehn et al., 2012, amongst others]. The 183 MHMCMC algorithm used here is fully described by Bloom and Williams [2014]. The 184 MHMCMC code is available (in matlab) upon request. 185 186 4. Sensitivity to Uncertainty Estimates 187 188 The posterior probability density function of x, p(π±|π), is dependent on the prescribed 189 observation and prior uncertainty estimates reported in section 3. To determine the 190 sensitivity of our MHMCMC results to imposed uncertainty values, we perturb individual 191 observation and prior value uncertainties. We categorize observation and prior 192 uncertainties into five groups (the uncertainty groups are summarized in Table 2). We 193 perform 5 sensitivity tests (S1-S5) by (i) perturbing (increasing) the prescribed 194 uncertainty estimates, and (ii) repeating the MHMCMC optimization of all parameters 195 (x). For test S1 we double the uncertainty of mean CO emissions (from 20% to 40%). For 196 test S2 we double the uncertainty of the normalized difference in CO (from 0.05 to 0.10). 197 For test S3 we double the uncertainty of TES CH4/CO ratios (double the values shown in 198 Table 1). For test S4 we double all emission factor uncertainties reported by Andreae and 199 Merlet [2001]. For test S5 we double the combustion and emission factor 2007-to-2010 200 change uncertainty (from f=2 to f=4 in equation 10). 201 202 Based on 105 MHMCMC samples for each sensitivity test (S1-S5), we re-calculate the 203 probability of each hypothesis (see main text), and compare it against the unperturbed 204 hypothesis probabilities (henceforth S0). The results are shown in Table 3. We find that 205 the hypothesis probability results are most sensitive to the 2007-to-2010 emission and 206 combustion factor change uncertainty (test S5), and least sensitive to the mean CO flux 207 uncertainty (test S1). 208 209 5. Repeat fires 210 211 We examine the locations of (a) 2007-2009 fires and (b) 2010 fires to determine the 212 maximum effect of 2007-2009 fire carbon losses on 2010 fires. We derive the location of 213 2007-2010 Jul-Oct fires at a 1km × 1km resolution (based on the MCD14ML product: 214 https://earthdata.nasa.gov/active-fire-data). We select MCD14ML fire events with a 215 confidence value >95%. We find that there is a 9.19% overlap at a 1km × 1km between 216 2007-2009 fire locations and 2010 fire locations (see Figure 3). We multiply the Saatchi 217 et al. [2011] above-ground biomass map by the 1km × 1km gridded fire locations (i.e. 218 equivalent to assuming a 100% burned area within each grid-cell) to determine potential 219 C losses within each pixel. The upper limit of C loss is the unlikely instance of (a) 100% 220 overlap within each 1km × 1km gricell (b) 0% biomass recovery, and (c) a 100% 221 combustion completeness. Given these upper-limit C losses for 2007-2009 fires, C losses 222 from 2010 fires would be 8.25% lower than 2007 fire C losses. Given that litter and foliar 223 C stocks are likely to partially recover, fires do not necessarily overlap at sub-pixel scale, 224 and combustion completeness is always < 100%, the effect of repeat fires on 2010 C loss 225 is most likely substantially below the 8.25% limit. 226 227 6. OMI NO2 228 229 Fire season (July-October) monthly mean NO2 values for 2006-2011 are shown in Figure 230 S3. Monthly mean NO2 estimates are based on the OMI standard NO2 product [Bucsela 231 et al., 2013; OMNO2, 2013], averaged to a 0.2°× 0.2° grid, excluding any measurements 232 that are affected by the row anomaly [KNMI, 2012] and that contain cloud radiance 233 fractions (CRF) of 80% or higher [Stammes et al., 2008]. The choice of such a high CRF 234 is motivated by the fact that the smoke plumes from biomass burning events are flagged 235 erroneously as cloud contamination by the OMI O2-O2 cloud product, and a smaller CRF 236 threshold would result in discarding otherwise cloud-free biomass burning events. For the 237 data points in Figure S3, the 0.2°×0.2° monthly averages are aggregated further over the 238 study area (5°S-25°S, 30°W-90°W). No correction has been made to account for 239 anthropogenic emissions (e.g., over Manaus, Brazil), because their overall contribution to 240 the total NO2 signal over the Amazon is relatively constant from month to month, and 241 negligible during times of biomass burning. 242 243 7. 2007 and 2010 GOME-2 chlorophyll fluorescence 244 245 We compare satellite-based observations of mean solar-induced fluorescence (SIF) 246 during the 2007 and 2010 months (February-June) preceding the major fire months (July- 247 Oct): we use the 0.5° × 0.5° gridded monthly GOME-2 version 25 SIF product [Joiner et 248 al., 2013] (data available at avdc.gsfc.nasa.gov). On a 0.5° × 0.5° grid within the study 249 area, we determine pixels where SIF observations were available each month: we also 250 evaluate SIF within pixels where the July-October fires were of similar size during both 251 years (0.5 < BA2007/BA2010 < 2). This area accounts for 78% of the total burned area 252 during May-Dec 2007 and 74% of the total burned area during May-December 2010. By 253 selecting these pixels we aim to minimize the effects of fire location and SIF sampling 254 differences between the two years. The selected pixels within the study are shown in 255 figure S4. Based on mean (area weighted) SIF, we find 2010 SIF is 4% lower relative to 256 2007 SIF during the pre-fire months (mean SIF2007 = 1.14 mW/m2/nm/sr and mean 257 SIF2010 = 1.09 mW/m2/nm/sr). In comparison, mean SIF within all consistently sampled 258 SIF pixels was 6% lower (mean SIF2007 = 1.04 mW/m2/nm/sr and mean SIF2010 = 0.98 259 mW/m2/nm/sr) in 2010 relative to 2007. To account for anomalously low sampling of SIF 260 retrievals in April 2007, we repeat the SIF calculation without April SIF values: we find 261 an overall 5%-7% SIF reduction for the February-June (minus April) time-period. In 262 addition to sampling differences between 2007 and 2010, we note that residual cloud and 263 aerosol differences, as well as intra-month vegetation trends, are additional potential 264 sources of bias in our estimate of 2007-to-2010 change in SIF [Joiner et al., 2013]. 265 266 References 267 268 Andreae, M. O., and Merlet, P. (2001). Emission of trace gases and aerosols from biomass burning. Global 269 Biogeochemical Cycles, 15(4), 955-966. doi: 10.1029/2000GB001382 270 271 Bucsela, E. J., et al. (2013). A new stratospheric and tropospheric NO 2 retrieval algorithm for nadir- 272 viewing satellite instruments: applications to OMI. Atmospheric Measurement Techniques, 6(10), 2607- 273 2626. doi:10.5194/amt-6-2607-2013 274 275 Bloom, A. A., and Williams, M. (2014). Constraining ecosystem carbon dynamics in a data-limited world: 276 integrating ecological "common sense" in a model-data-fusion framework. Biogeosciences Discussions, 277 11(8), 12733-12772. doi:10.5194/bgd-11-12733-2014 278 279 Braswell, B. H., Sacks, W. J., Linder, E., and Schimel, D. S. (2005). Estimating diurnal to annual 280 ecosystem parameters by synthesis of a carbon flux model with eddy covariance net ecosystem exchange 281 observations. Global Change Biology, 11(2), 335-355. doi: 10.1111/j.1365-2486.2005.00897.x 282 283 Hély, C., Alleaume, S., Swap, R. J., and Justice, C. O. (2003). SAFARI-2000 characterization of fuels, fire 284 behavior, combustion completeness, and emissions from experimental burns in infertile grass savannas in 285 western Zambia. Journal of Arid Environments, 54(2), 381-394. doi: 10.1006/jare.2002.1097 286 287 Henze, D. K., Hakami, A., and Seinfeld, J. H. (2007). Development of the adjoint of GEOS-Chem. 288 Atmospheric Chemistry and Physics, 7(9), 2413-2433. 289 290 Hurtt, G. C., and Armstrong, R. A. (1996). A pelagic ecosystem model calibrated with BATS data. Deep 291 Sea Research Part II: Topical Studies in Oceanography, 43(2), 653-683. doi:10.1016/0967- 292 0645(96)00007-0 293 294 Joiner, J., et al. (2013). Global monitoring of terrestrial chlorophyll fluorescence from moderate-spectral- 295 resolution near-infrared satellite measurements: methodology, simulations, and application to GOME-2. 296 Atmospheric Measurement Techniques, 6(10), 2803-2823. doi:10.5194/amt-6-2803-2013 297 298 KNMI, OMI Row Anomaly web page, http://www.knmi.nl/omi/research/product/rowanomaly- 299 background.php 300 301 Korontzi, S., et al. (2003). Seasonal variation and ecosystem dependence of emission factors for selected 302 trace gases and PM2. 5 for southern African savanna fires. Journal of Geophysical Research: Atmospheres 303 (1984β2012), 108(D24). doi: 10.1029/2003JD003730 304 305 Krystek, M., and Anton, M. (2007). A weighted total least-squares algorithm for fitting a straight line. 306 Measurement Science and Technology, 18(11), 3438. doi:10.1088/0957-0233/18/11/025 307 308 Malhi, Y., Saatchi, S., Girardin, C., and Aragão, L. E. (2009). The production, storage, and flow of carbon 309 in Amazonian forests. Amazonia and Global Change, 355-372. DOI: 10.1029/2008GM000733 310 311 OMNO2 (2013). OMI NO2 Standard Data Product, available at http://disc.sci.gsfc.nasa.gov/Aura/data- 312 holdings/OMI/omno2_v003.shtml, 2013. 313 314 Prasad, V. K., Kant, Y., Gupta, P. K., Sharma, C., Mitra, A. A., and Badarinath, K. V. S. (2001). Biomass 315 and combustion characteristics of secondary mixed deciduous forests in Eastern Ghats of India. 316 Atmospheric Environment, 35(18), 3085-3095. doi:10.1016/S1352-2310(01)00125-X 317 318 Saatchi, S. S., et al. (2011). Benchmark map of forest carbon stocks in tropical regions across three 319 continents. Proceedings of the National Academy of Sciences, 108(24), 9899-9904. 320 doi:10.1073/pnas.1019576108 321 322 Smith, M. J., Purves, D. W., Vanderwel, M. C., Lyutsarev, V. and Emmott., S. (2013). The climate 323 dependence of the terrestrial carbon cycle, including parameter and structural uncertainties. Biogeosciences 324 10(1) 583-606. doi:10.5194/bg-10-583-2013 325 326 Stammes, P., Sneep, M., De Haan, J. F., Veefkind, J. P., Wang, P., & Levelt, P. F. (2008). Effective cloud 327 fractions from the Ozone Monitoring Instrument: Theoretical framework and validation. Journal of 328 Geophysical Research: Atmospheres (1984β2012), 113(D16). doi:10.1029/2008GL033521. 329 330 Ward, D. E., Hao, W. M., Susott, R. A., Babbitt, R. E., Shea, R. W., Kauffman, J. B., and Justice, C. O. 331 (1996). Effect of fuel composition on combustion efficiency and emission factors for African savanna 332 ecosystems. Journal of Geophysical Research: Atmospheres (1984β2012), 101(D19), 23569-23576. doi: 333 10.1029/95JD02595 334 335 Worden, H. M., Deeter, M. N., Edwards, D. P., Gille, J. C., Drummond, J. R., and Nédélec, P. (2010). 336 Observations of nearβsurface carbon monoxide from space using MOPITT multispectral retrievals. Journal 337 of Geophysical Research: Atmospheres (1984β2012), 115(D18). doi: 10.1029/2010JD014242 338 339 Worden, J., et al. (2013a). CH4 and CO distributions over tropical fires during October 2006 as observed 340 by the Aura TES satellite instrument and modeled by GEOS-Chem. Atmospheric Chemistry and Physics, 341 13(7), 3679-3692. doi:10.5194/acp-13-3679-2013 342 343 Worden, J., et al. (2013b). El Niño, the 2006 Indonesian peat fires, and the distribution of atmospheric 344 methane. Geophysical Research Letters, 40(18), 4938-4943. doi: 10.1002/grl.50937 345 346 Ziehn, T., Scholze, M., and Knorr, W. (2012). On the capability of Monte Carlo and adjoint inversion 347 techniques to derive posterior parameter uncertainties in terrestrial ecosystem models. Global 348 Biogeochemical Cycles, 26(3). doi: 10.1029/2011GB004185 349 350 351 352 1 353 354 Figure S1. Monthly CH4/CO values (denoted as colors) based on TES mean column CH4 355 and CO retrievals within the study area. The CH4/CO color-scale is shown on the right. 356 The βCO cutoffβ corresponds to the minimum CO value used to distinguish CH4/CO in 357 fire plumes from background CH4/CO. The dashed line shows the CO cutoff value 358 selected for this study. Inset: map of study region. 359 360 361 Figure S2. 2007-2009 MCD14ML fire locations (blue) 2010 MCD14ML fire locations 362 (red) and overlap between 2007-2009 and 2010 fire locations (green). The fire locations 363 were gridded at a 1km × 1km resolution. The overlapping locations (green area) account 364 for 9.19% of all 2010 fire locations within the study region. 365 366 367 368 Figure S3. Mean monthly OMI NO2 aggregated within the study region for 2006-2011 369 fire seasons. Inset: map of study region. 370 371 1 2 372 373 Figure S4. Top left: Map Study region; Top-right: 0.5 × 0.5 degree pixels with (i) SIF 374 observations for every month during Jan-Apr 2007 and Jan-Apr 2010, and (ii) 375 comparably sized burned area during both 2007 and 2010 (0.5 < BA2007/BA2010 < 2); 376 Bottom-left: Mean Jan-Apr 2007 SIF where 0.5 < BA2007/BA2010 < 2; Bottom-right: Mean 377 Jan-Apr 2010 SIF where 0.5 < BA2007/BA2010 < 2. 378 379 380 381 382 383 384 385 Table 1: Observational constraints, associated uncertainties, and equivalent model 386 quantities. n Description Observational constraint (Oπ ) Uncertainty Model (Mπ ) (Οπ ) 1 Mean May-Dec MOPITT (MCO2010 + MCO2007 )/2 20% 2007 2010 (πΉπΆπ + πΉπΆπ )/2 CO (MCO) derived emissions 2 Normalized Difference between 2007 and 2010 5%) 394 395 396 397 2007 2010 [πΉπΆπ + πΉπΆπ ] 2007 2010 [πΉπΆπ + πΉπΆπ ] πππ‘07 πππ‘07 πΉπΆπ»4 /πΉπΆπ TES CH4/CO mass ratios CH4/CO Sep 07 = 0.0669 0.0083 * 4 (see section 2) CH4/CO Oct 07 = 0.0936 0.0117 * CH4/CO Sep 10 = 0.1093 0.0116 * agriculture fluxes. 393 2 3 388 392 normalized derived emissions * 391 0.05 (i.e. difference of 387 390 [MCO2010 β MCO2007 ] [MCO2010 + MCO2007 ] (May-Dec) MOPITT CO 5 389 2 πππ10 πππ10 πΉπΆπ»4 /πΉπΆπ πππ‘10 πππ‘10 πΉπΆπ»4 /πΉπΆπ Fluxes πΉπΆπ and πΉπΆπ»4 correspond to the total CO and CH4 fluxes from savanna, forest and 398 Table 2: Sensitivity tests1 Sensitivity test Quantity Uncertainty Perturbation S0 All observations/priors None S1 Mean [2007 & 2010) MOPITT CO based CO Double π1 (see Table 1) flux estimate S2 Normalized MOPITT CO based CO flux Double π2 (see Table 1) difference (between 2010 and 2007) S3 TES CH4/CO ratios Double π3,4,5 (see Table 1) S4 prior emission factor values Double ΟEF0 (p) (see equation 7) S5 prior constraint on 2007-to-2010 change factor Double uncertainty (change f=2 in EFs,b and CFb to f=4 in equation 10) 399 1 400 values. 401 402 403 404 405 406 407 408 409 410 411 412 The sensitivity tests used to quantify the dependence of results to prescribed uncertainty 413 Table 3: Probability of hypotheses H1-H41 Sensitivity H1 H2 H3 H4 H1 & H2 H2 & H3 Test S0 28% 60% 12% 0% 88% 72% S1 27% 61% 12% 0% 88% 73% S2 28% 57% 14% 0% 86% 72% S3 33% 57% 10% 0% 90% 67% S4 23% 53% 24% 0% 76% 77% S5 40% 50% 10% 0% 90% 60% 414 1 415 based on observations & priors with enhanced uncertainty estimates (sensitivity tests S1- 416 S5). Table 2 summarizes the uncertainty perturbations for individual sensitivity tests (S1- 417 S5). 418 419 420 421 The probability of hypotheses H1-H4 based on original uncertainty estimates (S0) and
© Copyright 2026 Paperzz