Article Inverse Modeling of Nitrogen Oxides Emissions from the 2010 Russian Wildfires by Using Satellite Measurements of Nitrogen Dioxide

Transcription

1 Article Inverse Modeling o Nitrogen Oides Emissions rom the 2010 Russian Wildires by Using Satellite Measurements o Nitrogen Dioide Evgeny V. Berezin 1, Igor B. Konovalov 1, * and Yulia Y. Romanova 2, * 1 Institute o Applied Physics, Russian Academy o Sciences, 46 Ulyanov Str., Nizhniy Novgorod , Russia; e.berezin@appl.sci-nnov.ru 2 Institute or Physics o Microstructures, Russian Academy o Sciences, Aonino, 7 Academicheskaya Street, Nizhniy Novgorod , Russia * Correspondence: konov@appl.sci-nnov.ru (I.B.K.); jul@ipm.sci-nnov.ru (Y.Y.R.); Tel.: (I.B.K.) Academic Editor: Anthony R. Lupo Received: 6 September 2016; Accepted: 12 October 2016; Published: 16 October 2016 Abstract: Observational constraints to biomass burning (BB) NO emissions as provided by satellite measurements o nitrogen dioide (NO2) critically depend on quantitative assumptions regarding the atmospheric NO lietime. In this study, we investigated NO emissions rom the etreme wildires that occurred in the European part o Russia in summer 2010 by using an original inverse modeling method that allowed us to avoid any a priori assumptions regarding the NO lietime. The method was applied to the tropospheric NO2 columns retrieved rom the measurements perormed by the OMI satellite instrument, while the relationship between BB NO emissions and tropospheric NO2 columns was simulated with the CHIMERE mesoscale chemistry transport model. Our analysis indicated that this relationship depends strongly on BB emissions o volatile organic compounds and that a dependence o the eective NO lietime on the NO lues can be essentially nonlinear. Our estimates o the total NO emissions in the study region are ound to be at least 40% larger compared to the respective data rom the GFASv1.0 and GFED4.1s global ire emission inventories. Keywords: nitrogen oides; wildires; satellite measurements; chemistry transport model; inverse modeling; biomass burning emissions 1. Introduction Nitrogen oides (NO = NO + NO2) are important atmospheric constituents aecting climate processes and air quality speciically by serving as precursors o tropospheric ozone [1,2] and inluencing secondary aerosol ormation [3,4]. One o the signiicant sources o NO on a global scale is associated with open biomass burning (BB): it is estimated that BB provides around 15% o total NO emissions globally [5]. On a regional scale, BB NO emissions can play a predominant role in the oidation processes leading to ozone ormation inside smoke plumes rom wildires [6] over other NO sources and provide a major contribution to NO2 column amounts over ire spots [7]. Despite the important role played by BB NO emissions in atmospheric processes, the current knowledge o the amounts o NO emitted into the atmosphere rom wildires is insuicient. Available BB NO emission estimates provided by ire emission inventories (e.g., [8,9]) based on indirect measurements o the amounts o biomass burned have mostly not been validated against atmospheric measurements and are likely to be rather uncertain as indicated by a diversity o the NO emission actor values involved in such inventories and reported in literature or similar types o vegetation cover [10]. Validation o BB NO emission estimates is more diicult compared to Atmosphere 2016, 7, 132; doi: /atmos

2 Atmosphere 2016, 7, o 20 validation o BB emissions o some other species (such as, e.g., CO) particularly due to high reactivity and a relatively short atmospheric lietime o NO, leading to a strong variability o NO concentrations both in space and time. Nonetheless, several studies [7,11 14] have recently eamined the relationship between the satellite-retrieved ire radiative power (FRP) and the tropospheric NO2 columns over ire spots and reported reasonable estimates o the ire emission rate (FER) [7,11] or the emission coeicient (EC) [12 14], both o which allow one to calculate NO emissions, given the observed values o FRP. The measurement-based estimates o FER or EC or dierent regions and dierent types o biophysical model applications (BioMA) are useul, as they provide insight into the capability o available parameterizations to capture spatial and temporal variability o NO emissions. However, such estimates are presently rather uncertain, particularly because they involve eplicit quantitative assumptions regarding the NO lietime and NO2/NO ratio. For eample, it has been argued [13] that as the NO lietime is likely to vary across individual ire plumes between 2 and 6 h, the estimates o EC can be uncertain at least by a actor o 3. In this study, we also use FRP and NO2 columns derived rom satellite measurements, but unlike the previous studies, we do not make any eplicit assumptions on the NO lietime and NO2/NO ratio. Instead, we ollow the inverse modeling approach that has earlier been successully used (e.g., [15 18]) to estimate anthropogenic NO emissions. Within this approach, the relationship between NO2 columns and NO emissions can be simulated by a chemistry transport model, and the NO emissions (or related emission parameters) as well as the NO lietime can be estimated by itting simulated NO2 columns to satellite data. The goals o our study were (1) to eamine easibility o using an inverse modeling approach to estimate BB NO emissions in a region aected by numerous and intense wildires; and (2) to assess the impact o chemical nonlinearities on the relationship between NO emissions rom wildires and corresponding NO2 columns over such a region. We ocused on the disastrous mega-ire event that took place in the European part o Russia in Emissions o some pollutants (such as CO and aerosol) rom the Russian 2010 ires were already investigated using inverse methods (e.g., [19 21]), but, to the best o our knowledge, there have yet been no similar studies ocusing on NO emissions rom those ires. 2. Data and Method 2.1. Input Data Satellite Data We used tropospheric NO2 column amounts retrieved rom measurements o visible and ultraviolet back scattered radiation by the Ozone Monitoring Instrument (OMI) [22] onboard the NASA EOS Aura satellite and provided by KNMI as the DOMINO version 2 data product [23,24]. OMI provides global coverage daily at the km 2 nominal spatial resolution; the measurements are taken in the early aternoon, as the Aura satellite operates in a sun-synchronous polar orbit with an equator crossing at 13:45 LST. The DOMINO version 2 data product was evaluated against ground-based (MAX-DOAS) measurements [25], and only a minor negative bias (~10%) was ound in the satellite data. NO2 data retrieved rom the OMI measurements (although with slightly dierent algorithms) were ound to be useul in earlier studies o BB NO emissions [7,11,12,14]. Following recommendations o the data providers, potentially biased data retrieved or the scenes with the cloud raction and surace albedo eceeding 30% and 0.3%, respectively, were disregarded. The details o the retrieval algorithm can be ound elsewhere [23]. Note that the tropospheric NO2 column retrieval procedure does not eplicitly account or optical eects o BB aerosol. However, such eects were taken into account and corrected or in the same way as similar eects associated with clouds. Scenes strongly contaminated by cloud or aerosol were discarded, and so the bias is likely to remain limited (<20%) [26]. The orbital NO2 data available or the study period that spans rom 15 July to 16 August 2010 were projected onto a 0.5 by 0.5 degree rectangular grid

3 Atmosphere 2016, 7, o 20 covering the study region (48 N 66 N; 20 E 56 E). Any data alling to the same grid on the same day were averaged. As a primary source o inormation on NO emissions rom ires, we used the Fire Radiative Power (FRP) data [27,28] retrieved rom satellite measurements o inrared radiation with the MODIS instruments operating onboard the Aqua and Terra satellites. Both the Aqua and Terra satellites are in sun-synchronous polar orbits with corresponding equator overpass times o 13:30 and 10:30 LST (at daytime) and 01:30 and 22:30 LST (at nighttime). We used the standard MODIS Level 2 thermal anomalies & ire (MOD14 and MYD14, Collection 5) data product available ree rom the NASA Land Processes Distributed Active Archive Center [29]. The data were provided or each satellite overpass at the nominal 1 1 km 2 resolution. The same data have been widely used (e.g., [7,9,21]) to study ire emissions, and details about the retrieval algorithm and potential uncertainties can be ound elsewhere [27]. Finally, our analysis involved satellite retrievals [30] o aerosol optical depth (AOD) at a 550 nm wavelength. The measurements were also perormed by the MODIS instruments onboard Aqua and Terra satellites. We used the Level 3 MYD08_D3/MOD08_D3 data product obtained rom the NASA Giovanni-Interactive Visualization and Analysis system [31] in the ramework o an earlier study [21]. The data had the spatial resolution o 1 by 1 degree and the temporal resolution o one day Simulation Data Along with NO2 data derived rom satellite measurements, we used NO2 columns simulated with the CHIMERE chemistry transport model (CTM) [32]. The CHIMERE CTM is a typical Eulerian three-dimensional model enabling simulations o air pollution on various scales, rom urban to continental ones. It takes into account major physical and chemical processes (including over 300 chemical and photochemical reactions) aecting the evolution o ~80 gaseous species in the lower atmosphere. It also enables simulations o aerosol species but simulated aerosol concentrations were not used in this study. In-detail descriptions o the model structure and o parameterizations o physical and chemical processes is available elsewhere [32]. The model documentation and the source code can be downloaded rom the CHIMERE website [33]. CHIMERE was successully used in a number o studies, including those dedicated to investigations o atmospheric eects o wildires (e.g., [19,21,34,35]). The CHIMERE coniguration was the same as in a previous study [21] o the atmospheric impacts o the 2010 Russian wildires, which was ocused on simulation o BB aerosol. Speciically, the simulations were perormed with a horizontal resolution o with the model grid covering most o European Russia and a portion o Eastern Europe (48 N 66 N; 20 E 56 E). In vertical, the model grid had 12 levels, with the top o the upper level ied at 200 hpa. Fire emissions were calculated and introduced into CHIMERE by using the emission model that had been successully used and described in detail earlier [19,21,36]. Briely, it assumes that the BB emission rate, E s, o a species, s, in a given grid cell o the horizontal domain in an hour, t, is proportional to the daily mean FRP values, Фd, that were inerred rom daily maimum FRP values derived rom the MODIS measurements (see Section 2.1.1) or a given grid cell: s d l s l E ( t) h where α is a conversion actor rom FRP to dry BB rate, which is reerred to below also as the FRP-to-BBR conversion actor and or which a reerence value, αr, o kg s 1 MW 1 is adopted, taking into account the eperimental data [37], are the emission actors or a species s and a land cover type s l l, ρl is the raction o the land cover type l, hel is the diurnal variation o ire emissions, and c is an ad hoc correction actor which was calculated (see [19,21]) using the measured AOD values and was intended to compensate or a possible attenuation o FRP measured rom satellites by very heavy smoke and also to account (at least partly) or emissions rom peat ires. The emission actor values, s, or the three aggregated vegetative land cover types (orest, grassland, and agricultural land) l were taken to be the same as in our previous study (see [21] and Table 2 therein). The emission diurnal l el ( t) c (1)

4 Atmosphere 2016, 7, o 20 cycle was also adopted rom the previous study (see [21] and Figure 1 therein), where it had been derived or the study region rom the FRP measurements using a method developed earlier [36]. The emissions were distributed in the vertical direction by using a well-known parameterization [38] o the maimum injection height o BB emission as a unction o FRP and the boundary layer height. The emission proiles were calculated or each grid cell and each hour by assuming that BB emissions corresponding to individual ire piels are uniormly distributed rom the ground up to the altitude corresponding to the maimum injection height. Note that our ire emission model given by Equation (1) disregards possible variations o the FRP-to-BBR conversion actor both in space and time. It also does not eplicitly take into account emissions rom peat ires which are not visible rom space. These simpliications are mainly due to the lack o observational inormation about such variations. Available estimates o α or dierent vegetation land cover types [9] are based on emission inventory data and, or this reason, were not quite suitable or our study based on the inverse modeling approach. A noteworthy eature o our simulations is that hourly values o the photolysis rates were calculated oline or each grid cell and hour using the TUV model [39] and the satellite AOD data described in Section (see also [19] or details o our calculations). CHIMERE runs were driven with meteorological data obtained rom the WRF-ARW (v. 3.6) model [40]. Further details on meteorological data, anthropogenic and biogenic emission data, and boundary and initial conditions used in the CHIMERE runs perormed in this study can be ound elsewhere [21]. The hourly BB NO2 concentration data obtained rom the CHIMERE runs were matched in time and space to the satellite NO2 data and integrated vertically by applying the averaging kernels [41]. The averaging kernels provide inormation on contribution o dierent atmospheric layers to the tropospheric NO2 column amounts retrieved rom the OMI measurements and were available as a part o the DOMINO data product. The hourly BB NO2 concentrations available or the same day and grid cell were averaged. The model was run or the period rom 12 July to 16 August 2010 both with and without ire emissions. Only the dierence o nitrogen oides concentrations rom the two twin runs was considered in this study in order to minimize an impact o probable biases in biogenic and anthropogenic emissions on our estimates: such a dierence was considered as a contribution o BB emissions to the nitrogen oides concentrations. The irst three days o the model runs were withheld rom the analysis or the model's spin-up. The three-day spin-up was ound to be suicient in the case considered or evaluation o the BB raction o tropospheric NO2 columns, since there had been no signiicant ire emissions in the study region beore the initial date o the simulation Method Estimation o NO emissions was based on a general approach established in previous studies [19,21,36]. The idea is to optimize one or more parameters o the ire emission model by itting modeled concentrations o a species considered to respective observations. Following one o the previous studies [21], we optimized the FRP-to-BBR conversion actor (α) by assuming (in accordance to Equation (1)) that the same value o that actor is applicable to any types o vegetation land cover. Although, in reality, it is not necessarily true, large uncertainties in the measurement and simulation data prevent us rom obtaining estimates o α separately or individual biome types. Variability o the true values o α is epected to be taken into account in the conidence intervals or our estimates. In this study, the optimal value o α, αopt, was inerred by itting the NO2 columns simulated with the CHIMERE model to those derived rom the OMI observations: Nd i 1 i i i C C C 2 arg min (2) opt where Co are the observed tropospheric NO2 columns, Cb are the background parts o the tropospheric NO2 columns (that is, the tropospheric NO2 column amounts which would be present in the atmosphere i there were no ires and which were estimated as eplained below), Cm are the o b m

5 Atmosphere 2016, 7, o 20 simulated tropospheric column amounts o NO2 originating rom ires and i is the inde o a day. The NO2 column amounts (Co, Cb, Cm) involved in Equation (2) are the averages o the corresponding spatially-distributed data or a given day over the study region. Note that the spatial averaging was perormed in order to minimize the impact o possible random uncertainties in the spatial variations o the NO2 columns (due to, e.g., errors in the spatial distributions o ire emissions evaluated using Equation (1)) on our emission estimates. To estimate the background NO2 columns in a given grid cell (i) and in a given day (d), we considered the tropospheric NO2 columns derived rom the OMI measurements or the same grid cell but or another day chosen rom the period rom 15 July to 19 July 2010, when, according to our simulations, a contribution o ires to the mean NO2 columns over the study region was negligible. I such background data were available or more than one day, a nearest, to d, day was selected. On the other hand, i there were no observational data or any o the days in the indicated period, then the given grid cell i was ecluded rom urther analysis. Importantly, estimation o the background NO2 columns in such a way does not rely on the accuracy o the simulations o NO2 columns in the background conditions. This allows us to minimize the eect o possible biases (due to, e.g., uncertainties in anthropogenic and biogenic NO emissions, as well as in the boundary conditions or NO and related species) in the background simulations with CHIMERE or the study region [42,43] on our estimates. Note that Cm involved in Equation (2) was obtained as the dierence between the results o two twin model runs with and without ire emissions (as mentioned in Section 2.1.2), and so Cm also could not be directly inluenced by possible inaccuracies o the simulations or the background conditions. Furthermore, the analysis o our results (see Section 3.3 below) does not show any evidence that the dependence o Cm on BB NO emissions is signiicantly aected by the background conditions in an indirect way (e.g., through chemical interactions) either. Equation (2) was resolved iteratively. In each iteration, Cm, was assumed to depend on BB NO emissions (and, thus, on α) linearly, that is: C ( n) ( n) m ( n1 ) C ( n1) m where α (n) and α (n 1) are values o the conversion actor in two consecutive iterations, n is the iteration inde and the symbol denotes conditioning. Ater substituting α (n) (instead o αopt) and Cm given by Equation (3) into Equation (2), the latter can be resolved analytically: 1 ( n) ( n1) i Nd i2 i Nd i i i i 1 C n i Co Cb C n m 1 ( ) (4) ( 1) ( 1) m In this study, the irst iteration (based on the initial model run) was perormed with the reerence value o the conversion actor, αr (see Section 2.1.2). Three iterations were ound to be suicient in order to ensure that α (n) is converged to the optimal estimate o α with a relative numerical uncertainty o less than 5%. Conidence intervals or αopt were evaluated by means o a Monte Carlo eperiment based on the bootstrapping method [44]. Speciically, the simulations perormed with the optimal value o α were used to represent the true values o the BB raction o the tropospheric NO2 columns. The corresponding observed values were generated by sampling random errors rom the residuals o the modeled (optimal) data or the NO2 columns (see the right-hand part o Equation (2)) and adding them to such true values. The re-sampling procedure and estimation o αopt (assuming a linear dependence o Cm on α) were repeated 1000 times. The conidence intervals were evaluated (and are reported below) in terms o the 95th percentile. In this way, we took into account any model and measurement errors contributing to the cost unction involved in the right-hand part o Equation (1). Such errors may be due to various non-systematic uncertainties in the retrieval procedure used to derive tropospheric NO2 columns rom the OMI measurements [23,26,45], as well due to transport errors in our model. A part o the uncertainties taken into account in the conidence intervals or αopt may be associated with our ire emission model (see Section 2.1.2), speciically with the simpliying (3)

6 Atmosphere 2016, 7, o 20 assumptions concerning the FRP-to-BBR conversion actor as well as with possible inaccuracies o the injection height, with variability o the emission actors and with inaccuracies in the input FRP data. Indeed, it seems reasonable to epect that the errors associated with the emission model varied rom day to day during the study period, relecting temporal and spatial variations in the burning conditions and in the types and density o uel. It should be noted, however, that the conidence intervals reported below cannot ully account or systematic errors (biases) both in the OMI retrieval and in our simulations. Such biases may, in particular, be due to a probable underestimation o the NO2 columns retrieved or the strongly polluted scenes. The biases may also be due to possible inconsistencies between the vertical distribution o NO2 in the real atmosphere and the vertical proiles used in the retrieval procedure [45] and predicted by CHIMERE in this study. Evaluation o such systematic errors and their eects on the corresponding NO emission estimates is a challenging task that goes beyond the scope o this study. Since the tropospheric NO2 columns derived rom satellite measurements provide inormation only on NO emissions, there is an issue concerning speciication o BB emissions o other species (such as CO and volatile organic compounds, VOCs). As the main option, we used the same values o the conversion actor to calculate the BB emissions o all model species. This estimation case is labeled below as VOCbyNO. However, as it was pointed out earlier [21,36], estimates o α derived rom measurements o dierent species may be dierent, partly due to uncertainties in the assumed emission actors. Thereore, as another option, we calculated the BB emissions o CO and VOCs by using the value o α optimized in a previous study [21] against ground-based measurements o CO concentrations in the Moscow region: this CO-measurement-based value o α had been ound to be a actor o 1.88 larger than its reerence value αr. The second estimation case is labeled below as VOCbyCO. Note that the estimation method described above is similar but not identical to those employed in the preceding studies [21,36] where satellite measurements and the CHIMERE model were used to constrain emissions o CO and aerosol rom wildires in Russia. The dierences (that will not be discussed here in detail) are mainly due to the act that CO and aerosol are much less reactive species compared to NO. Some speciic assumptions (e.g., those concerning the origin o possible systematic uncertainties in simulated data) involved in the preceding studies were no longer valid in the case considered in this study and should have been avoided. For eample, it was assumed that the bias in the simulated CO columns in the background conditions was mainly due to errors in the boundary conditions: accordingly, it was estimated by averaging the dierences between the CO columns rom the simulations and measurements over a number o grid cells and days representing similar environment conditions. In contrast, errors in the corresponding NO2 columns may predominantly come rom uncertainties in the biogenic and anthropogenic emissions speciied in the model; urthermore, such errors are likely to strongly vary both in time and space because o a short lietime o NO. Consequently, estimation o BB NO emissions using eactly the same methods as those employed in the previous studies turned out to be ineasible, and such methods had to be modiied accordingly. The major dierence is that here we estimate the background NO2 columns directly rom the measurement data and thus avoid the problem o estimating the bias in the corresponding simulated data. To study a potential impact o chemical nonlinearities on BB NO emission estimates based on satellite measurements, we perormed several additional model runs using BB emissions calculated with dierent values o the conversion actor. Speciically, the reerence value o the conversion actor was scaled with a actor ranging rom 0. 5 to 2, and the resulting values were used in Equations (3) and (4) as the zero-order estimate, α (0), to obtain the irst-order estimate (constrained by satellite measurements according to Equation (2)) o the conversion actor, α (1). Furthermore, we evaluated the eective lietime o NO, τno, corresponding to each value o α (0) as ollows:

7 Atmosphere 2016, 7, o 20 1 Nd Nc ij ij Nd Nc e 1 1 ij ij no, (5) i j i 1 j 1 where e ij are the BB NO emission rates (g hour 1 ) in a grid cell j (in the horizontal domain) and in a day i, ν ij are values (simulated by the CHIMERE model) o the tropospheric column amounts o NO originating rom wildires (g), Nc and Nd are the total numbers o grid cells and days in the study region and period, and θ ij is the selection operator which is equal to unity i the measurement data or a given grid cell and day is available and to zero otherwise. Equation (5) assumes that the evolution o the total (integral) amount o NO rom BB burning over the study region can be roughly approimated by a simple balance equation describing the loss o NO as a irst order decay process, with the rate epressed as τno 1. The deinition given by Equation (5) is based on the known concept [46] o the mean lietime o an atmospheric substance and ollows rom a consideration o a simple balance equation or NO originating rom ires ( NO NO ), describing the loss o three dimensional domain: NO as a irst order decay process in each model grid cell in the [ NO in out d ] ( F F ) ( E L[ NO ]), (6) dt where the square brackets denote the mass concentration o NO, Fin and Fout are the lues o NO in and out o the given grid cell, respectively, E is the local BB NO emission (or injection) rate, and L is the rate o losses o NO due to chemical and deposition processes. Note that the inverse o L gives an estimate o the instant lietime o NO with respect to chemical and deposition processes in a given grid cell. Ater (i) averaging both the let and right parts o Equation (6) over the study region (both in the horizontal and vertical directions) and the study period; (ii) taking into account that concentration o NO approaches zero both at the beginning and at the end o the study period; (iii) disregarding the integral lu o NO rom the study region (over which the persistent blocking anticyclone was developed in summer 2010 [47]); and (iv) multiplying both parts o the equation with the average amount o NO, we arrive to the ollowing approimate relationship: L[ NO ] 1 [ NO ] 1 E [ NO ], (7) where the angle brackets denote the average quantities. We urther note that the term on the lethand side o Equation (7) has the meaning o the mean NO lietime evaluated as the inverse o the weighted average o L (with local NO concentrations being the weights), while the term on the right-hand part o Equation (7) is analogous to the term on the right-hand part o Equation (5). The main dierence between Equation (7) and Equation (5) is that the latter includes only those values o the NO concentrations and BB NO emissions that correspond in space and time to the available OMI observations. Thereore, the eective lietime, τno, deined according to Equation (5) can be regarded as an approimation or the NO lietime given by Equation (7). Note that we considered only those values o NO that have counterparts in the satellite data, because the primary purpose o evaluation o τno in this study was to investigate the eects o chemical nonlinearities on the BB NO emission estimates that can be derived rom satellite measurement using either our method or simpler techniques involving a priori assumptions regarding the NO lietime [12 14]. Note also that the relationship between α (1) and α (0) depends not only on the NO lietime but also on the NO2/NO ratio: this ratio, in turn, depends on ozone concentration and can be aected by changes in the NO emissions. In addition, unlike τno calculated using Equation (5), estimation o α (1) partly takes into account BB NO emissions in grid cells (days) or which the measurement data are not available. Thereore, the dependence o τno on α (0) cannot

8 Atmosphere 2016, 7, o 20 ully characterize that o the NO2 columns retrieved rom OMI measurements on the BB NO emission rate in the study region and period. Nonetheless, we believe that the estimates o τno provided below can be helpul in characterizing the processes determining the relationship between the NO2 columns and NO emissions. Furthermore, such estimates can eventually contribute to reducing uncertainties associated with FER or EC estimates [7,11 14] based on the analysis o the empirical relationships between FRP and tropospheric NO2 columns. However, we do not recommend using our estimates o τno outside o this speciic contet. 3. Results 3.1. Estimation o the FRP-to-BBR Conversion Factor Estimation o the conversion actor (as eplained in the previous section) is based on the analysis o daily time series o spatially averaged NO2 columns derived rom the OMI measurements (Co) and simulated with the CHIMERE model (Cm) and involves evaluation o the background raction (Cb) o the tropospheric NO2 columns. The corresponding time series are presented in Figure 1, along with the time series o the dierence o Co and Cb. Eamples o underlying daily values o tropospheric NO2 columns (speciically, on 29 July and 8 August) spatially distributed inside o the study region are shown in Figure 2. Note that values o Cm shown in Figures 1 and 2 were calculated with the conversion actor equal to its reerence value αr speciied above (see Section 2.1.2). Note also that the spatial distributions o CO and aerosol on the same days were considered in a preceding study [21]. As is evident rom Figure 1, the simulations tend to overestimate the BB part o NO2 columns (that is the dierence o Co and Cb), although the model captures, to a signiicant etent, the temporal variability o NO2 column amounts rom wildires (speciically, the correlation coeicient eceeds a value o 0.75). The act that, or some days, Cb is larger than Co indicates that there may be considerable uncertainties in our estimates o background NO2 values as well as in the original satellite NO2 data. We assume that such uncertainties are mostly random in nature (varying rom day to day) and thus are taken into account in the conidence intervals o our emission estimates reported below. Figure 1. Time series o the daily NO2 column amounts averaged over the study region. Co: the tropospheric NO2 columns retrieved rom OMI measurements; Cb: a background part o tropospheric NO2 columns; Cm: the simulated tropospheric column amounts o NO2 rom wildires. The vertical

9 Atmosphere 2016, 7, o 20 dash-dot lines mark the data or 29 July and 8 August. The RMSE and the correlation coeicient indicated in the igure legend or Cb and Cm were calculated relative to Co and Co Cb, respectively. (a) (b) (c) (d) Figure 2. (a,b) Spatial distributions o the tropospheric NO2 columns (10 15 molec. cm 2 ) derived rom the OMI measurements on 29 July and 8 August 2010; (c,d) the tropospheric NO2 column amounts originated rom wildires, according to the simulations using the reerence value o the FRP-to-BBR conversion actor. Note that the simulation data were processed with the averaging kernels rom the OMI retrievals and are, accordingly, shown only where the measurement data were available. Eamples o spatial distributions o NO2 columns shown in Figure 2 indicate that the NO2 retrievals or many scenes in the study region and period were missing (apparently, due to intererence o BB aerosol that is shielding the radiation coming to a satellite rom the boundary layer). Nonetheless, many moderately polluted scenes showing considerable NO2 amounts which, according to our simulations, originated rom wildires can still be seen in the OMI retrievals. Figure 3 presents the conversion actor estimates, α (n), that were obtained according to Equations (2) (4) by using the simulations in which BB emissions were calculated with the assumed conversion actor values, α (n 1). For convenience, values o both α (n) and α (n 1) are normalized to the reerence value o the conversion actor, αr. Most o the estimates shown in Figure 3 illustrate results o the irst iteration (n = 1), ecept or the optimal estimates (circumvented by black circles in Figure 3) o α (αopt) that were obtained as a result o the third iteration (n = 3). The estimates were obtained separately or the two cases ( VOCbyNO and VOCbyCO ) that involve the dierent assumptions regarding the BB emissions o VOCs and CO. I the relationship between the simulated NO2 columns and the BB NO emissions were linear, then any value o α (n 1) would yield the same value o α (n), or, in other words, α (n) would be independent o α (n 1). Our results show that α (n) is indeed very weakly (positively) dependent on α (n 1), but only or the case VOCbyNO. A more pronounced (decreasing) dependence o α (n) on α (n 1) is ound or the case VOCbyCO. Overall, our results shown in Figure 3 indicate that the eects o chemical nonlinearities on the relationship between the BB emissions and tropospheric NO2 columns in the situation considered cannot be disregarded, at least without certain constraints on VOCs emissions, although these eects are not very strong. In particular, the increase o α (0) rom 0.5 to 2 (that is, by a actor 4) in the VOCbyCO case resulted in the decrease o α (1) rom

10 Atmosphere 2016, 7, o to 0.67 (that is, only by about 30 percent), which is an eect o the corresponding increase (by a actor o ~1.3) o the ratio o the mean BB NO2 column amounts and NO emission rates in the respective simulations. It is especially noteworthy that the sensitivity o the optimal estimates o α to changes in BB VOC emissions turned out to be relatively small: speciically, while BB VOCs emissions or the VOCbyCO case were almost a actor o 3 larger than or the VOCbyNO case (when α = αopt), the optimal estimate o the conversion actor is ound to be only 17 percent larger or the ormer case (0.298 ± kg s 1 MW 1 ) than that or the latter one (0.254 ± kg s 1 MW 1 ). Unortunately, there are presently no data that would allow us to characterize the uncertainty o the regionally averaged emission actors or VOCs, although such uncertainty is likely much smaller than the standard deviation (close to 100% [48]) o individual emission actor measurements. Note that systematic uncertainties in our estimates o α due to biases in the emission actors or VOCs are probably only partly accounted or in the conidence intervals or these estimates, to an etent as such uncertainties are associated with errors in the simulated daily variations o the tropospheric NO2 columns. Figure 3. Dependence o the FRP-to-BBR conversion actor values optimized under the assumption o linearity o the relationship between the NO2 columns and the BB NO emission rate on an assumed value o the same conversion actor. The conversion actor values are normalized to their reerence value. The dependence is shown or the two estimation cases, VOCbyNO and VOCbyCO, see Section 2.2. The estimates circumvented by black circles are the optimal estimates (αopt) obtained as a result o the third iteration (that is, with n = 3) o the iterative procedure described in Section 2.2, whereas the other estimates are a result o the initial iteration (n = 1). Another potential source or a bias in our emission estimates is due to uncertainties in estimates o the background estimates o the NO2 columns. To test the sensitivity o our results to this actor, we repeated our estimation procedure or the VOCbyNO case, but with the background NO2 columns estimated using available observational data or other days (12 14 July 2010 instead o July 2010 as in the base case). The optimal estimate o α or the test case was ound to be only 10 percent larger than or the base case. This result indicates that our BB NO emission estimates are suiciently robust with regard to random uncertainties in the background NO2 column amounts. Furthermore, we investigated whether or not the background NO2 columns changed signiicantly over the study period (e.g., due to changes in meteorological conditions or in biogenic emission rates). To this end, we selected data points (grid cells and days) where the simulated NO2 column amounts originated rom ires were ininitesimal (less than 10 2 % relative to the background NO2 columns estimated or the same grid cells/days using the OMI data or the period July 2010). The selected data were averaged separately over the periods July 2010 and 20 July 16 August As a

11 Atmosphere 2016, 7, o 20 result, the mean value over the second period was ound to be very insigniicantly (~6%) lower than that or the irst period. This result conirms that uncertainties in our estimates o α due to uncertainties associated with evaluation o the background NO2 columns are unlikely to eceed the statistical uncertainties represented by the conidence intervals or the values o αopt. Figure 4 illustrates the perormance o our NO2 simulation using optimized NO emissions or the VOCbyNO case. Note that the simulation data or the VOCbyCO case are almost indistinguishable rom those in the VOCbyNO case and are not presented here. The daily time series shown in Figure 4 are the same as those in Figure 1, ecept that the BB NO emissions used in the simulation presented in Figure 4 were calculated with the optimal estimate o α (αopt) instead o its reerence value (αr) as in Figure 1. Compared to the model results obtained with the reerence value o α, the perormance o the optimal simulation is noticeably better. Speciically, the root mean square error (RMSE) has reduced by more than 20% and a considerable dierence (~40%) between the mean values o the NO2 columns rom the model and measurements has practically been eliminated. Furthermore, the absolute values o the dierences between the daily values o the NO2 columns rom the measurements and simulations have been reduced in the 19 o 28 days considered (ater ecluding the initial 5 days used or estimation o the background NO2 columns). A simple statistical test based on the binomial distribution and involving the statistical hypothesis that our inversion results simply in adding random errors to the simulated NO2 columns conirms that the improvement o the perormance o our NO2 simulations is statistically signiicant with the probability o type 1 error being less than It is noteworthy that the optimization o the NO BB emissions did not aect a value o the correlation coeicient (r). This observation implies that the correlation coeicient in the given case can be considered as an independent (o α) measure o accuracy o the spatial-temporal ield o the BB NO emissions used in our simulations. Figure 4. The same as in Figure 1, ecept that the simulated time series, Cm, was obtained rom the model run using the optimal value o the FRP-to-BBR conversion actor or the VOCbyNO case. The spatial distributions o the average (over the study period) NO2 column amounts obtained using the OMI measurements and given by the CHIMERE model are compared in Figure 5a,b. For the comparison purpose, the simulated BB NO2 column amounts (Cm) calculated with the optimal estimate o the FRP-to-BBR conversion actor or the VOCbyNO case were added to the background values o the NO2 columns (Cb). Visual comparison o Figure 5a,b reveals that such composed data demonstrate a rather reasonable, although evidently imperect, agreement with the measurement data. In particular, the both two-dimensional ields o NO2 columns ehibit the large-scale negative

12 Atmosphere 2016, 7, o 20 gradients in the north-east direction, as well as the hot spots around Moscow. There are also the enhancements in the NO2 column amounts south o Moscow. Quantitatively, a degree o agreement between the NO2 columns in the corresponding grid cells can be characterized by the correlation coeicient that is ound to be o This value indicates that the composed data (Cm + Cb) account or about a hal o the total variance o the data. The rest o the variance includes, in particular, small scale variability in the both ields, which may partly be due to random uncertainties in Co and Cb. A part o the discrepancies between the composed data and the measurements may also be due to errors in the BB emission data as well as due to the limited spatial coverage o the satellite measurements in the situation considered. For eample, one o the prominent discrepancies is that spots with high NO2 columns east o Moscow at approimately 45 E in Figure 5b,c (which, according to our additional analysis, are mostly due to strong emissions rom ires on 26 July) are missing in Figure 5a. This discrepancy may be due to, e.g., mismatches between the MODIS and OMI measurement in space and time: that is, the smoke plumes rom the ire spots which were seen by the MODIS instruments could be outside the vision ield o the OMI instrument (at least partly). Nonetheless, the reasonable overall agreement between the data shown in Figure 5a,b provides one more argument in avor o reliability o the results o our analysis. Note that, in accordance to our results shown in Figure 4, a contribution o ires to the tropospheric NO2 columns is ound to be relatively small (see Figure 5c). (a) (b) (c) Figure 5. The spatial distributions o the average (over the study period) NO2 column amounts obtained using the OMI measurements (a); The modeled data that were composed by summing up the simulated tropospheric column amounts o NO2 rom ires and the background values o the tropospheric NO2 columns (b); and The simulated tropospheric column amounts o NO2 rom ires (c) are also shown separately.

13 Atmosphere 2016, 7, o NO Lietime Estimates The NO lietime estimates (τno) obtained in accordance to Equation (5) are presented in Figure 6 as a unction o α or the two estimation cases ( VOCbyNO and VOCbyCO ). Similar to the dependencies o α (n) on α (n 1) that are shown in Figure 3, the dependencies o τno on α or the two cases are qualitatively dierent. Speciically, while the case VOCbyNO eatures almost constant NO lietimes irrespectively o the assumed values o the conversion actor, the values o τno obtained or the VOCbyCO case are increasing by a actor o 2.1 with the increase o α by a actor o 4. The NO lietime values corresponding to the optimal estimates o the conversion actor are 4.2 and 3.1 h or the VOCbyNO and VOCbyCO cases. These values are within a rather broad range o lietimes (2 6 h) assumed in earlier studies o BB NO emissions [11 14]. Importantly, the relative dierence between the lietimes calculated or the VOCbyNO and VOCbyCO cases is much larger compared to the relative dierence between values o αopt (see Figure 3) or the same cases. As mentioned above (see Section 3.1), our estimates o α depend on the ratio o the mean BB NO2 column amounts and NO emissions. Since this ratio depends not only on the NO lietime but also on the NO2/NO ratio, a comparison o the results shown in Figures 3 and 6 indicates that the dierence o the lietimes is counterbalanced by the dierence o the NO2/NO ratio. Figure 6. The eective NO lietime (τno) estimated a unction o the FRP-to-BBR conversion actor (α) normalized to its reerence value (αr). Black circles depict the NO lietime values corresponding to the optimal estimates o α or the two estimation cases ( VOCbyNO and VOCbyCO ) considered in this study. Ehaustive theoretical interpretation o the dependencies shown in Figure 6 goes beyond the scope o this study. Here, we discuss the chemical mechanisms underlying these dependencies only very briely. In particular, we note that NO is mainly removed rom the atmosphere in a reaction o NO2 with OH [1]. Both OH and NO2 concentrations are typically growing unctions o ozone (O3) concentration [1]. Our simulation data or O3 and OH concentrations (not presented here) indicate that hot spots eaturing strong BB emissions typically coincide in space and time with depleted concentrations o these species, as well as with reduced values o NO2/NO ratio. It can be speculated that the behavior o the O3 ormation rate is the main actor determining a nearly linear dependence o NO2 concentration on the BB NO emission rate over the hot spots. Speciically, in the high NO limit (in which OH concentration is relatively small), the O3 ormation rate is known to be proportional to the ratio o VOC to NO concentrations [1] and thus can be epected to be independent o α in the VOCbyNO case. In contrast, the VOC emission rate is independent o α in

14 Atmosphere 2016, 7, o 20 the VOCbyCO case. Thus the O3 ormation rate is likely to decrease with an increase o α, leading to the observed increase in the NO lietime. Such an interpretation o the results presented in Figure 6 is consistent with the dependencies presented in Figure 3. Indeed, given that the NO lietime is almost independent o α in the VOCbyNO, the results shown in Figure 3 or the same case imply that the NO2/NO ratio and thus O3 concentration are also kept nearly constant. On the other hand, the act that the relative range o the changes o α (n) in the VOCbyCO case in Figure 3 is much smaller than that o the changes o τno in Figure 6 indicates that the NO2/NO ratio decreases with an increase o α in that case: this observation, in turn, implies that O3 concentration accordingly decreases NO Emission Estimates We used our optimal estimate o the FRP-to-BBR conversion actor (0.254 kg s 1 MW 1 ) or the VOCbyNO case to calculate NO emission rates rom wildires in the study region and period. Figure 7 shows the time series o the daily BB NO emission rates that were obtained by integrating either over all grid cells and hours o a given day or only over selected grid cells/hours that were provided with the OMI NO2 data. According to our calculations, the ull emissions tended to increase until 6 August and then started to decrease, consistently with results o other studies [9,19,21,49] that reported on the strong air pollution in the study region in the beginning o August. Two major peaks o the BB NO emission rate occurred on 4 August and 6 August. A dierence between the two time series indicates to what etent the inormation provided by the OMI retrievals on BB NO emissions is representative o the whole study period. Evidently, the BB NO emission rates corresponding to the scenes provided with the OMI NO2 retrievals represent only a part (around 30%, on the average) o the integral emission rates in the study region and period. Thereore, there is a question whether or not our optimal estimate o the FRP-to-BBR conversion actor is representative (at least, on average) o all ires in the study region. Figure 7. Time series o the daily mean BB NO emission rates that were obtained by integrating either over all model grid cells (see the green curve) or only over selected grid cells that are provided with the OMI NO2 data or a given day (see the brown curve). Figure 8 presents the spatial distributions o the BB NO emission rates averaged over the study period. The two dierent distributions (shown in Figure 8a,b) correspond to the two respective time series presented in Figure 7. Epectedly, the distribution obtained only with the selected data typically eatures smaller BB NO emission rates compared to the one obtained using all the available emission data. Nonetheless, it is important to note that both the coverage and spatial structure o the emission rates shown in the dierent distributions are rather similar. In particular, both distributions

15 Atmosphere 2016, 7, o 20 show that there were strong NO emissions rom ires in the regions south-east and east o Moscow as well as around Russia's border with Kazakhstan (near the bottom right-hand corner o the model domain). Furthermore, only 9% o the ire spots (grid cells with non-zero ire emissions) shown in Figure 8a are missing in Figure 8b. Thereore, a comparison o the spatial distributions shown in Figure 8a,b provide evidence that our optimal conversion actor estimate that was used to scale BB NO emissions rates across the study region is, to a considerable etent, representative o the whole region and period considered. Note that, according to a special test involving splitting the study period into two approimately equal parts, the dierence between the optimal values o the conversion actor or those periods was less than the statistical uncertainty o our estimate or the whole period. Although the result o this test provides urther evidence in avor o applicability o our optimal conversion actor estimates across the study region and period, the question regarding the representativeness o the conversion actor estimates derived rom observations o the limited number o ire scenes still remains and requires urther investigations that go beyond o the scope o this study. (a) (b) Figure 8. Spatial distributions o the BB NO emission rates averaged over the study period. The two dierent distributions are obtained by averaging either over all model grid cells and hours in the study period (a); or only those grid cells/hours which are provided with the OMI NO2 retrievals (b) and correspond to the green and brown curves in Figure 7, respectively. Figure 9 compares our estimates o the total BB NO emissions in the study region and period with corresponding estimates obtained using the data o ire emission inventories, such as GFED4.1s [8] and GFASv1.0 [9]. The GFED inventory is based on satellite measurements o burned area and other parameters rom a biosphere model, while the GFAS inventory iners BB emissions rom the FRP satellite measurements by using a method similar to one employed in this study. However, unlike our method, where the conversion actor is optimized against atmospheric composition measurements, the GFAS inventory uses conversion actor values calibrated with the GFED data. Thus the GFED4.1s and GFASv1.0 inventories are not ully independent, although the spatial structure and temporal variability o regional emission ields provided by these inventories may be signiicantly dierent. Our results suggest that both GFED4.1s and GFASv1.0 considerably underestimate BB NO emissions (by 33% and 30%, respectively, relative to our estimate or the VOCbyNO case) and that the dierence between our emission estimate and the respective inventory data cannot be eplained by statistical uncertainties in our estimates. This inding is qualitatively consistent with the results o other studies (see, e.g., [19 21,36]) where those inventories were ound to underestimate emissions o other species (such as CO and aerosol) rom wildires in Russia. It should be kept in mind, however, that the conidence intervals shown in Figure 9 may not ully account or possible biases in our estimates due to several actors mentioned above, including uncertainties in the VOC BB emissions and in background NO2 column amounts, probable negative biases in the satellite NO2 retrievals or the strongly polluted scenes, possible inconsistencies between the NO2 vertical proiles used in the retrieval and predicted by our model, as well as possibly

16 Atmosphere 2016, 7, o 20 insuicient spatial and temporal representativeness o the observational constraints provided by the available NO2 satellite measurements to BB NO emissions. Quantitative understanding o these actors and their impacts on BB NO emission estimates inerred rom satellite NO2 measurements calls or urther dedicated studies. Figure 9. Optimal estimates o the total NO emissions in the study region and period in comparison with corresponding estimates based on the data o the GFED4.1s and GFASv1.0 emission inventories. The optimal estimates are shown separately or the two estimation cases, VOCbyNO (i) and VOCbyCO (ii). The conidence intervals represent statistical uncertainty evaluated in terms o the 95th percentile. 4. Conclusions In this paper, we eamined the easibility o estimating NO emissions rom wildires within an inverse modeling approach by using satellite measurements o NO2 columns. For this purpose, we considered the situation o the mega-ire event that occurred in the European part o Russia in July August We used the tropospheric NO2 columns available rom satellite measurements perormed with the OMI instrument along with the NO2 column amounts simulated with the CHIMERE chemistry transport model (CTM) in which emissions o NO and other species rom biomass burning (BB) were taken into account. The BB emissions were calculated using the Fire Radiative Power (FRP) data that were available rom the MODIS satellite instrument. The measured and simulated NO2 columns were combined in the ramework o a simple inverse modeling method aimed at estimating the conversion actor (α) rom FRP to BB rate. BB emissions o individual model species were then calculated using estimates o α and typical emission actors reported in literature. The ocus in our study was given to analysis o the eect o chemical nonlinearities on the relationship between BB NO emissions and respective NO2 column amounts. While in our study this relationship was simulated with CTM, several previous studies (mentioned in the Introduction) in which BB NO emissions had been investigated using NO2 OMI measurements, involved eplicit quantitative assumptions regarding the dependence o NO2 columns on BB NO emissions. To test the robustness o such assumptions, we perormed several model runs where a typical reerence value o the conversion actor (and thus the mean BB NO emission rate) was scaled with a actor ranging rom 0.5 to 2.0. The simulations were done or the two cases involving dierent assumptions regarding emissions o volatile organic compounds (VOCs) and CO. Speciically, in one o the cases ( VOCbyNO ), VOCs and CO emission rates were scaled in the same way as the NO emissions, while in another case ( VOCbyCO ), the VOCs and CO emission rates were calculated with the α value inerred in a previous study rom ground-based measurements o CO and were kept constant irrespectively o changes in the NO emission rates. The corresponding linearized relationships