Supplement to: Sixteen years of MOPITT satellite data strongly constrain Amazon CO fire emissions

constrain Amazon CO fire emissions Stijn Naus1,2, Lucas G. Domingues3,4, Maarten Krol1,5, Ingrid T. Luijkx1, Luciana V. Gatti4,6, John B. Miller7, Emanuel Gloor8, Sourish Basu9,10, Caio Correia4,6, Gerbrand Koren1, Helen M. Worden11, Johannes Flemming12, Gabrielle Pétron7,13, and Wouter Peters1,14 1Meteorology and Air Quality, Wageningen University and Research, The Netherlands 2SRON Netherlands Institute for Space Research, Utrecht, The Netherlands 3National Isotope Centre, GNS Science, New Zealand 4Nuclear and Energy Research Institute, São Paulo, Brazil 5Institute for Marine and Atmospheric Research, Utrecht University, The Netherlands 6National Institute for Space Research (INPE), São José dos Campos, Brazil 7Global Monitoring Laboratory, National Oceanographic and Atmospheric Administration, Boulder, CO, USA 8School of Geography, University of Leeds, Leeds, UK 9Earth System Science Interdisciplinary Center, University of Maryland, MD, USA 10NASA Goddard Space Flight Center, Greenbelt, MD, USA. 11Atmospheric Chemistry Observations and Modeling, National Center for Atmospheric Research, Boulder, CO, USA 12European Centre for Medium-Range Weather Forecasts (ECMWF), Reading, UK 13Cooperative Institute for Research in Environmental Sciences, University of Colorado, Boulder, CO, USA 14Centre for Isotope Research, University of Groningen, The Netherlands Correspondence: Stijn Naus (s.naus@sron.nl)

measurement uncertainty of 2 ppb CO for surface observations in the inverse system.
In Fig. S1, the difference between simulated and observed CO mole fractions is shown at four representative surface sites, per month, per site), satellite observations are weighted more heavily than surface observations. We have tried to reduce this 15 effect by inflating the error on satellite observations by a factor √ 50 (as in Nechita-Banda et al. (2018)), but near the zoom domain satellite observations still seem to dominate. This can be seen from the relative good match at ALT (and most other sites; not shown) compared to the other three sites shown in Fig. S1. Especially Ragged Point, Barbados (RPB) and Ascension Island (ASC) are important for the inversion, since these sites roughly represent the inflow conditions of the Amazon domain (e.g. Gatti et al., 2014). The posterior differences between simulated and observed mole fractions at these sites fall outside 20 the prescribed observational uncertainties for surface data and, especially at RPB, are not always an improvement on the prior simulation.
To investigate the importance of our difficulty in reproducing mole fractions at surface sites near the zoom domain, we performed inversions for 2010, 2015 and 2016, in which we reduced the error on CO surface observations from 2 ppb to 0.2 ppb. As expected, in these new inversions the agreement with observations at all sites improves compared to our standard 25 inversions (black lines in Fig. S1), while MOPITT CO columns in these three new inversions are reproduced equally well as in our standard inversions (not shown). Biomass burning emissions derived in these three inversions are lower than those derived in our standard inversions (Fig. S2), but the difference is small and consistent between years (7−11 Tg/year). We note that this is consistent with the posterior mismatch at RPB (orange line in Fig. S1). Namely, the simulated mole fractions at RPB are consistently too low in our standard inversions, while in the inversions with a 0.2 ppb error simulated mole fractions at RPB of derived emissions to the boundary conditions. Intuitively, this can be understood, since, during the dry season, advection of CO into the Amazon domain is less important than local emissions. In conclusion, we find some inconsistencies in the Amazon boundary conditions in our standard inversions, but we also find limited sensitivity of derived emissions to these inconsistencies.
S2 Other sensitivities in the inverse system 50 In our main text, we have presented the sensitivity of derived fire emissions to fire emission prior, and to assimilating IASI instead of MOPITT satellite data. In this supplement, we discuss sensitivities of the derived CO emissions to the other important components of the inverse system: CO production from NMHC and OH chemistry. In general, we find that the absolute magnitude of the derived CO emissions, and the spatio-temporal patterns therein, are robust.

55
In this section we investigate the influence of natural CO emissions on our derived fire emissions. The aggregated category of natural emissions as defined here includes secondary production of CO (from CH 4 and NMVOCs), and direct biogenic CO emissions. This source category partly correlates with fire emissions, as CO production from NMVOCs in particular peaks during the dry season, because of significant NMVOC emissions from fires, as well as higher biogenic emissions of natural emitted NMVOCs like isoprene during dry months.

60
The natural emissions used in our standard emissions were annually repeating and retrieved from a full-chemistry simulation of TM5 for 2006. While we consider the spatio-temporal patterns of this distribution realistic, the absolute amount of natural CO emissions was very high. Globally, 1750 Tg CO/year is emitted in this aggregated category, compared to, for example, 1400 Tg/year suggested in Huijnen et al. (2010). When we compare our natural CO emissions with CO emissions calculated based on the Model of Emissions of Gases and Aerosols from Nature (MEGANv2.1) (Sindelarova et al., 2014), we find confirmation that 65 our reference estimate of natural emissions is too high (Fig. S3). Therefore, as a sensitivity test, we have performed inversions with natural emissions scaled down by 30% (green dashed line in Fig. S3), which is more in line with other estimates.
In the inversions where natural emissions have been scaled down, we derive higher biomass burning emissions (Fig. S2).

S2.2 OH chemistry fields
The final uncertainty in the CO budget that we consider is loss to oxidation by OH. Since OH has a lifetime of seconds, its atmospheric abundance is determined by local atmospheric conditions, which can be disturbed by high emissions from 85 fires: precisely our regime of interest. Our default OH field (from Spivakovsky et al. (2000) in the troposphere, scaled by a factor 0.92; from Brühl and Crutzen (1993) in the stratosphere) is climatological, and it therefore does not include the interannual disturbances in fire emissions. Since CO has a typical atmospheric lifetime of one month, variations in OH will drive an integrated, slow response in atmospheric CO abundance, which suggests that the impact of varying OH on emission localization and timing will be small. However, if we systematically over-or underestimate OH abundance, then this can cause In inversions with CAMS OH fields, we find lower biomass burning emissions by 35 Tg and 19 Tg in 2010 and 2011 respectively (Fig. S2). This is one of the largest sensitivities in our inverse system with significant interannual variability. It is 95 driven by the large difference between the two different OH fields we have tested: in some regions CAMS OH is lower than our default OH fields by a factor 100 (Fig. S4). This difference is largely consistent between months and between years.
The gap in OH fields over remote forests is a known attribute of OH fields derived in some full-chemistry models and it is likely related to incomplete recycling mechanisms for OH. In-situ measurements of OH over remote forests have revealed Figure S4. The ratio between OH fields from the CAMS reanalysis product for 2010, and climatological OH fields from Spivakovsky et al. (2000), scaled by 0.92. OH fields were first averaged over our April−December inversion window and over the lowest 10 model layers, corresponding to approximately 500 hPa, or 5 km. The color scale is logarithmic.
higher OH concentrations, even under low-NO x conditions (Lelieveld et al., 2008). One explanation is that under these pristine 100 conditions, isoprene oxidation can sustain a high OH recycling efficiency (e.g. Lelieveld et al., 2008;Taraborrelli et al., 2012), although more recent work has suggested that natural NO x emissions have been underestimated over these regions (Wells et al., 2020). Whatever the driving mechanism, the gap in OH over the Amazon in CAMS-OH could explain why, in the CAMS reanalysis for CO (Flemming et al., 2017), GFAS emissions do not result in an underestimate of CO over the Amazon: too-low OH and too-low fire emissions cancel out.

105
Given these considerations, we deem our climatological OH fields more realistic over the Amazon region than the OH fields from the CAMS reanalysis. The large sensitivity of derived emissions to the OH field used is likely a reflection of the extreme difference between the OH fields we have tested, rather than a reflection of a large intrinsic sensitivity in the inverse system. We note that OH variations have a relatively diffuse, slow impact on CO, which means that its contribution to the sharp dry season peak in atmospheric CO abundance is limited. However, loss to OH does contribute substantially to the Amazon CO budget, 110 and efforts to understand and constrain OH over remote regions (e.g. Fu et al., 2019;Nölscher et al., 2016)  in the Amazon domain.
6 Figure S5. Simulated and observed monthly mean MOPITT CO columns over the Santarém aircraft site. Columns are averaged over a 1 • by 1 • area centered on Santarém (56 • E; 2.8 • S). CO columns sampled from the prior (green) and from the posterior (orange) simulation are shown.

S3 CO columns over the Santarém aircraft site
Out of the five sites for which aircraft profiles are compared to model results, we only found substantial differences between observed and TM5-simulated aircraft profiles at the Santarém aircraft site. Specifically, aircraft profiles sampled in a simulation with MOPITT-optimized emissions were systematically too low ( Figure 2 in the main text). In contrast, MOPITT CO columns 120 over Santarém are well reproduced ( Figure S5). This effect is similar to what we observed at the NOAA surface site of Ragged Point, Barbados, where simulated CO mole fractions in the standard inversion were also lower than those observed (Supp. S1). This could point to a systematic bias in the MOPITT CO columns. However, surface observations, aircraft profiles and MOPITT CO columns all have different vertical sensitivities and therefore systematic uncertainties in the vertical transport of TM5 also affect this comparison.