Nine years of global hydrocarbon emissions based on source inversion of OMI formaldehyde observations

Abstract

As formaldehyde (HCHO) is a high-yield product in the oxidation of most volatile organic compounds (VOCs) emitted by fires, vegetation, and anthropogenic activities, satellite observations of HCHO are well-suited to inform us on the spatial and temporal variability of the underlying VOC sources. The long record of space-based HCHO column observations from the Ozone Monitoring Instrument (OMI) is used to infer emission flux estimates from pyrogenic and biogenic volatile organic compounds (VOCs) on the global scale over 2005–2013. This is realized through the method of source inverse modeling, which consists in the optimization of emissions in a chemistry-transport model (CTM) in order to minimize the discrepancy between the observed and modeled HCHO columns. The top–down fluxes are derived in the global CTM IMAGESv2 by an iterative minimization algorithm based on the full adjoint of IMAGESv2, starting from a priori emission estimates provided by the newly released GFED4s (Global Fire Emission Database, version 4s) inventory for fires, and by the MEGAN-MOHYCAN inventory for isoprene emissions. The top–down fluxes are compared to two independent inventories for fire (GFAS and FINNv1.5) and isoprene emissions (MEGAN-MACC and GUESS-ES).

The inversion indicates a moderate decrease (ca. 20 %) in the average annual global fire and isoprene emissions, from 2028 Tg C in the a priori to 1653 Tg C for burned biomass, and from 343 to 272 Tg for isoprene fluxes. Those estimates are acknowledged to depend on the accuracy of formaldehyde data, as well as on the assumed fire emission factors and the oxidation mechanisms leading to HCHO production. Strongly decreased top–down fire fluxes (30–50 %) are inferred in the peak fire season in Africa and during years with strong a priori fluxes associated with forest fires in Amazonia (in 2005, 2007, and 2010), bushfires in Australia (in 2006 and 2011), and peat burning in Indonesia (in 2006 and 2009), whereas generally increased fluxes are suggested in Indochina and during the 2007 fires in southern Europe. Moreover, changes in fire seasonal patterns are suggested; e.g., the seasonal amplitude is reduced over southeast Asia. In Africa, the inversion indicates increased fluxes due to agricultural fires and decreased maxima when natural fires are dominant. The top–down fire emissions are much better correlated with MODIS fire counts than the a priori inventory in regions with small and agricultural fires, indicating that the OMI-based inversion is well-suited to assess the associated emissions.

Regarding biogenic sources, significant reductions in isoprene fluxes are inferred in tropical ecosystems (30–40 %), suggesting overestimated basal emission rates in those areas in the bottom–up inventory, whereas strongly positive isoprene emission updates are derived over semiarid and desert areas, especially in southern Africa and Australia. This finding suggests that the parameterization of the soil moisture stress used in MEGAN greatly exaggerates the flux reduction due to drought in those regions. The isoprene emission trends over 2005–2013 are often enhanced after optimization, with positive top–down trends in Siberia (4.2 % year${}^{-\mathrm{1}}$) and eastern Europe (3.9 % year${}^{-\mathrm{1}}$), likely reflecting forest expansion and warming temperatures, and negative trends in Amazonia ($-$2.1 % year${}^{-\mathrm{1}}$), south China ($-$1 % year${}^{-\mathrm{1}}$), the United States ($-$3.7 % year${}^{-\mathrm{1}}$), and western Europe ($-$3.3 % year${}^{-\mathrm{1}}$), which are generally corroborated by independent studies, yet their interpretation warrants further investigation.

Introduction

Complementary to bottom–up methodologies for deriving emissions estimates, inverse modeling has the potential to improve those estimates through the use of atmospheric observations of trace gas compounds, in particular over regions undergoing fast economic development and facing intense air pollution problems, like eastern China but also on the global scale . Pollutants like CO and NO${}_{\mathrm{2}}$ are directly detected from satellite and their emissions have been inferred using inversion techniques on different scales (e.g., ). The detection of formaldehyde columns from satellite sensors measuring in the UV-visible spectral window opened the way for the derivation of fluxes of non-methane volatile organic compounds (NMVOCs), a broad class of formaldehyde precursors emitted by vegetation, fires, and anthropogenic activities . These compounds have a profound impact on air quality and climate, owing to their influence on OH levels and the methane lifetime and to their role as precursors of ozone and secondary organic aerosols . The accurate estimation of their fluxes is therefore of utmost importance.

Natural emission from vegetation is the dominant volatile organic compound (VOC) source. The global annual flux is estimated at ca. 1000 Tg VOC, with isoprene accounting for half of this emission . Despite a general consensus on the isoprene emission patterns, including their dependence on temperature and light density responsible for their marked diurnal and seasonal variations, these emission estimates come, however, with large uncertainties, associated with the strong variability of emission factors and the extrapolation of sparse measurements to larger scales. An uncertainty of a factor of 2 in global and regional isoprene fluxes was reported based on a compilation of numerous literature studies , whereas emission models were found to be strongly sensitive to choices of input variables, leading to even wider uncertainty, of ca. 200–1000 Tg C year${}^{-\mathrm{1}}$ globally .

The global biomass burning fluxes are estimated by bottom–up inventories to be ca. 1300–2200 Tg C on a yearly basis, which corresponds to 40–100 Tg VOC year${}^{-\mathrm{1}}$ using emission factors from the compilation of or . These estimates, however, depend on assumptions made in fire emission models regarding fuel loading and consumption efficiency and on the quality of land cover maps and fire proxies from satellite .

Formaldehyde is a high-yield product in the oxidation of a large majority of NMVOCs. Isoprene alone is responsible for approximately 30 % of the global HCHO burden according to model estimates , whereas the contribution of vegetation fires is globally small (3 %) but can be locally very important. Spaceborne vertical columns of HCHO retrieved from GOME, SCIAMACHY, the Ozone Monitoring Instrument (OMI), and GOME-2 sensors have been used to constrain the VOC budget on different scales (e.g., ). Top–down flux estimates deduced from two satellite sensors with different overpass times showed a good degree of consistency over the Amazon and globally . The latter study using GOME-2 (09:30 LT) and OMI (13:30 LT) HCHO observations in 2010 reported a good agreement between the inversion results over most areas and identified large regions where the derived emissions were highly consistent (e.g., Amazonia, southeastern US). Encouraged by those results, and relying on a multiyear record of HCHO columns observed by the OMI sensor, we use inverse modeling to derive top–down pyrogenic and biogenic VOC estimates over 2005–2013. The satellite data offer an unparalleled opportunity to bring new insights in our understanding of emissions and their quantification, to infer long-term seasonal and interannual flux variability, and to detect potential emission trends that may not be well represented in bottom–up inventories. To this purpose, we use a global chemistry-transport model (CTM), coupled with an inversion module and a minimization algorithm adjusting the emissions used in the model, in order to achieve an optimal match between the modeled and the observed HCHO columns while accounting for errors in the a priori emissions and the HCHO observations. The optimized fluxes are compared with independent bottom–up pyrogenic and biogenic emission inventories as well as with previous literature studies. The methodology is briefly presented in Sect. , and an overview of the results is discussed in Sect. . The top–down fluxes and comparisons to bottom–up inventories over big world regions are discussed thoroughly in Sects. – and emission trends in Sect. . Conclusions and final remarks are presented in Sect. .

Methods

We used formaldehyde observations retrieved from the OMI spectrometer aboard the Aura mission and fully documented in a recent study . The retrievals are based on an improved DOAS algorithm that reduces the effect of interferences between species and ensures maximum consistency between the OMI and GOME-2 columns. The current data version (v14) uses an iterative algorithm to remove spikes in the residuals of the slant columns and a procedure based on the background normalization to remove striping artifacts due to calibration problems . In addition to the destriping procedure, in order to reduce the effect of the OMI row anomaly issue affecting the spectra after 2007 (http://www.knmi.nl/omi/research/product/rowanomaly-background.php), the OMI rows presenting higher levels of noise and fitting residuals than the average were systematically removed from the dataset . Although this filtering leads to a loss of coverage, the resulting dataset is more appropriate for addressing trend studies, as explained in .

The IMAGESv2 global model calculates the concentrations of 131 transported and 41 short-lived trace gases with a time step of 6 h at 2${}^{\circ }$ $\times$ 2.5${}^{\circ }$ resolution between the surface and the lower stratosphere. The effect of diurnal variations is accounted for through correction factors on the photolysis and kinetic rates obtained from model simulations with a time step of 20 min, which are also used to calculate the diurnal shapes of formaldehyde columns required for the comparison with satellite data. A detailed model description is provided in . Meteorological fields are obtained from ERA-Interim analyses of the European Centre for Medium-range Weather Forecasts (ECMWF). The model uses anthropogenic NO${}_x$, CO, SO${}_{\mathrm{2}}$, NH${}_{\mathrm{3}}$, and total NMVOC emissions from the Emission Database for Global Atmospheric Research (EDGAR4.2, http://edgar.jrc.ec.europa.eu), which is overwritten by the EMEP inventory (http://www.ceip.at/ms) over Europe and by the REASv2 inventory over Asia. The NMVOC speciation is obtained from REASv2 over Asia and from the RETRO inventory elsewhere. The emissions over the US are scaled according to the NEI national totals for all years between 2005 and 2013 (http://www.epa.gov/air-emissions-inventories/air-pollutant-emissions-trends-data).

Biomass burning emissions are taken from the latest version of the Global Fire Emissions Database, GFED4s (July 2015), which includes the contribution of small fires based on active fire detections (; ). The GFED data are available on a daily basis at 0.25${}^{\circ }$ $\times$ 0.25${}^{\circ }$ resolution from 1997 through the present at http://www.globalfiredata.org. Those emissions are distributed vertically according to .

A priori isoprene emissions are obtained from the MEGAN-MOHYCAN model for all years of the study period at a resolution of 0.5${}^{\circ }$ $\times$ 0.5${}^{\circ }$ (http://tropo.aeronomie.be/models/isoprene.htm). Besides the emission dependence on leaf temperature, photosynthetically active radiation (PAR), leaf area, and leaf age, the model accounts for the inhibition of isoprene emissions in very dry soil conditions through a dimensionless soil moisture activity factor (${\mathit{\gamma }}_{\text{SM}}$) expressed as a function of volumetric soil moisture content obtained from the ERA-Interim reanalysis. The parameterization of ${\mathit{\gamma }}_{\text{SM}}$ bears large uncertainties, as it is based on scarce (and sometimes contradictory) field data, and its implementation can lead to very different results depending on the choice of database for soil moisture data . It reduces the emissions by ca. 20 % globally according to MEGAN-MOHYCAN, with strongest effects (up to factor of 3 or more) over Australia and southern Africa and to a lesser extent over northern Africa (Sahel), the western US, and the Middle East.

Global distributions of mean 2005–2013 HCHO columns for January and June observed by OMI (upper panels), modeled using emissions (middle panels) and inferred after optimization (lower panels). The columns are expressed in 10${}^{15}$ molec. cm${}^{-\mathrm{2}}$. The observed monthly averages exclude scenes with cloud fractions higher than 40 % and land fractions lower than 20 %, as well as data with a retrieval error higher than 100 %. The four lower panels illustrate the model–data difference before and after optimization for January and July.

[Figure omitted. See PDF]

The chemical degradation mechanism of pyrogenic NMVOCs is largely described in , with only minor modifications. This mechanism includes an explicit treatment for 16 pyrogenic formaldehyde precursors. The emissions of other pyrogenic compounds is represented through a lumped compound (OTHC) with a simplified oxidation mechanism designed in order to reproduce the overall formaldehyde yield of the explicit NMVOC mix it represents. The oxidation mechanism for isoprene is based on , modified to account for the revised kinetics of isoprene peroxy radicals according to the Leuven Isoprene Mechanism version 1 (LIM1) , as well as for the chemistry of the isoprene epoxides (IEPOX) following the Master Chemical Mechanism (MCMv3.2; http://mcm.leeds.ac.uk/MCMv3.2/). Using a box model, the formaldehyde yield in isoprene oxidation by OH is calculated to be 2.4 mol mol${}^{-\mathrm{1}}$ in high NO${}_x$ (1 ppbv NO${}_{\mathrm{2}}$, after 2 months of simulation) and 1.9 mol mol${}^{-\mathrm{1}}$ for 0.1 ppbv NO${}_{\mathrm{2}}$. It should be stressed that the isoprene mechanism still bears important uncertainties at low-NO${}_x$ conditions, as both the oxidation products of the isoprene epoxides and the isomerization products of isoprene peroxy radicals have complex degradation mechanisms that are still far from being well elucidated despite recent progress . Note that suppressing the isomerization channel in the isoprene degradation resulted in only slightly higher model HCHO columns over isoprene-rich regions .

The mismatch between the CTM and the observations, quantified by the cost function $J$, $$J\left(\mathit{f}\right)={\displaystyle \frac{\mathrm{1}}{\mathrm{2}}}\left(\right(H\left(\mathit{f}\right)-\mathit{y}{)}^{\text{T}}{\mathbf{E}}^{-\mathrm{1}}\left(H\right(\mathit{f})-\mathit{y})+{\mathit{f}}^{\text{T}}{\mathbf{B}}^{-\mathrm{1}}\mathit{f}),$$

is minimized through an iterative quasi-Newton optimization algorithm, which is based on the calculation of the partial derivatives of $J$ with respect to the input variables. In our case the input variables are scalars $\mathit{f}=\left(f_j\right)$, such that the optimized flux can be expressed as

$${\mathrm{\Phi }}_i^{\text{opt}}(x,t)=\sum _{j=\mathrm{1}}^m{e}^{f_j}{\mathrm{\Phi }}_i(x,t),$$ with ${\mathrm{\Phi }}_i(x,t)$ being the initial flux depending on space (latitude, longitude) and time (month) and $m$ being the emission categories or processes. In Eq. (), $H\left(\mathit{f}\right)$ denotes the model acting on the variables, $\mathit{y}$ the observation vector, $\mathbf{E}$ and $\mathbf{B}$ the covariance matrices of the errors in the observations and on the a priori parameters $\mathit{f}$, respectively, and ${}^{\text{T}}$ means the transpose of the matrix. The partial derivatives of $J$ with respect to $\mathit{f}$ are calculated by the discrete adjoint of the IMAGESv2 chemistry-transport model (CTM) . The derivation of monthly pyrogenic and biogenic fluxes is carried out on a global scale at the resolution of the model (2${}^{\circ }$ $\times$ 2.5${}^{\circ }$), as described in detail in . The inversions are performed separately for all years of the study period (2005–2013), and about 60 000 flux parameters are optimized per year globally.

Upper panel: mean (2005–2013) annual biomass burning emission estimates in Tg C/grid per year according to the a priori inventory GFED4s and to the OMI-based biomass burning emissions. Lower panel: mean (2005–2013) annual isoprene emission estimates in Tg isoprene per grid cell per year according to the a priori MEGAN-MOHYCAN inventory and from the OMI-based inversion.

[Figure omitted. See PDF]

The covariance matrix of the observational errors is assumed to be diagonal. The errors are calculated as the squared sum of the retrieval error and a representativity error set to 2 $\times$ 10${}^{15}$ molec. cm${}^{-\mathrm{2}}$. The assumed error in the a priori biogenic and pyrogenic fluxes is factor of 3. This choice reflects the high variability of the biomass burning emission source and the strong uncertainties associated with the biogenic emissions, as demonstrated by the large range of literature emission estimates . The spatiotemporal correlations among the a priori errors in the flux parameters are defined as in . About 20–40 iterations are needed to reach convergence, which is attained when the gradient of the cost function is reduced by a factor of 1000 with respect to its initial value. The cost function generally decreases by ca. 45–55 % in comparison to its initial value.

Figure  illustrates a comparison between observed monthly mean HCHO column densities over 2005–2013 and monthly columns simulated by the IMAGESv2 model sampled at the time and location of the satellite measurement. The observed monthly averages exclude scenes with cloud fractions higher than 40 % and land fractions lower than 20 %, as well as data with a retrieval error higher than 100 %. The number of effective observational constraints is highest in the first years of the OMI mission (ca. 17 000 per year) and declines by about 15 % after 2009 due to instrumental degradation effects , whereas the data availability is higher during the summer than in the winter in the Northern Hemisphere (ca. 1600 vs. 1200 measurements per month). The satellite columns are freely available at the BIRA-IASB website (http://h2co.aeronomie.be). The OMI-based emission fluxes presented in this study are available at the GlobEmission web portal (http://www.globemission.eu).

Mean a priori and OMI-based emission estimates compared to independent emission inventories for open biomass burning and isoprene emissions calculated for different world regions and globally. Regions are defined in Fig. . The means are taken over the period of data availability, i.e., over 2005–2013 for all inventories, except for MEGAN-MACC (2005–2010) and GUESS-ES (2005–2009). NH: Northern Hemisphere; SH: Southern Hemisphere.

	North	South	Europe	NH	SH	Russia	Southeast	Australia	Global
	America	America		Africa	Africa		Asia
Biomass burning emissions (burned biomass in Tg C year${}^{-\mathrm{1}}$)
GFED4s	105	319	31	418	684	130	237	104	2028
OMI-based	86	273	35	320	530	112	203	95	1653
GFAS	187	328	22	333	431	264	246	126	1938
FINNv1.5	112	452	34	278	415	114	579	22	2006
GFED4	84	231	17	279	479	97	156	95	1438
Isoprene emissions (Tg isoprene year${}^{-\mathrm{1}}$)
MEGAN MOHYCAN	32	141	6.8	50	29	9.4	36	38	343
OMI-based	26	97	8.4	35	28	11	31	36	272
MEGAN-MACC	34	173	7.8	103	67	12	80	94	570
GUESS-ES	44	143	18.1	77	60	20	63	26	452

Overview of the results

The source optimization leads to a good overall agreement with the OMI observations (Fig. ), in particular in the tropics, as a result of the high signal-to-noise ratio in the observations at these latitudes. The a posteriori columns remain close to the a priori at high latitudes, mainly due to lower data availability and higher observational errors at these latitudes . The inferred mean HCHO columns over the study period are generally decreased by 20–25 % over the Amazon and equatorial Africa, whereas a mean decrease of about 13 % is found in the southeastern US during summertime (cf. Supplement, Fig. S1). The HCHO columns are increased in a few regions after inversion, especially during biomass burning events. The annually averaged global distribution of pyrogenic and isoprene emissions over 2005–2013 before and after optimization is illustrated in Fig. . Figure  displays the extent of the regions over which comparisons will be discussed. Bottom–up and top–down emission estimates are summarized in Tables  and .

Definition of big and small regions used in this study. Big regions are N America (13–75${}^{\circ }$ N, 40–170${}^{\circ }$ W), S America (60${}^{\circ }$ S–13${}^{\circ }$ N, 90${}^{\circ }$ W–30${}^{\circ }$ E), Europe (37–75${}^{\circ }$ N, 15${}^{\circ }$ W–50${}^{\circ }$ E), NH Africa (0–37${}^{\circ }$ N, 20${}^{\circ }$ W–65${}^{\circ }$ E), SH Africa (0–40${}^{\circ }$ S, 20${}^{\circ }$ W–65${}^{\circ }$ E), Russia (37–75${}^{\circ }$ N, 50–179${}^{\circ }$ E), SE Asia (10${}^{\circ }$ S–37${}^{\circ }$ N, 65–170${}^{\circ }$ E), and Australia (10–50${}^{\circ }$ S, 110–179${}^{\circ }$ E). Small regions are the SE US (26–36${}^{\circ }$ N, 75–100${}^{\circ }$ W), Amazonia (5–20${}^{\circ }$ S, 40–75${}^{\circ }$ W), W Europe (37–71${}^{\circ }$ N, 10${}^{\circ }$W–20${}^{\circ }$ E), E Europe (37–71${}^{\circ }$ N, 20–50${}^{\circ }$ E), northern Africa (0–16${}^{\circ }$ N, 15${}^{\circ }$ W–35${}^{\circ }$ E), southern Africa (15–35${}^{\circ }$ S, 10–55${}^{\circ }$ E), Siberia (57–75${}^{\circ }$ N, 60–140${}^{\circ }$ E), south China (18–32${}^{\circ }$ N, 109–122${}^{\circ }$ E), Indochina (9–30${}^{\circ }$ N, 94–109${}^{\circ }$ E), Indonesia (10${}^{\circ }$ S–7${}^{\circ }$ N, 90–140${}^{\circ }$ E), N Australia (10–24${}^{\circ }$ S, 110–150${}^{\circ }$ E), and S Australia (24–38${}^{\circ }$ S, 110–155${}^{\circ }$ E).

[Figure omitted. See PDF]

Global a priori and OMI-based emission estimates per year. Fire estimates are expressed in Tg C year${}^{-\mathrm{1}}$, isoprene in Tg of isoprene per year.

Year	A priori	Optimized	A priori	Optimized
	fires	fires	isoprene	isoprene
2005	2252	1936	349	282
2006	2207	1721	339	280
2007	2202	1966	340	285
2008	1873	1605	324	263
2009	1862	1504	339	269
2010	2150	1679	363	272
2011	1872	1404	341	258
2012	2058	1676	350	276
2013	1773	1383	338	265
2005–2013	2028	1653	343	272

The OMI-based fire flux estimates are compared with two independent inventories: GFAS and FINNv1.5. The Global Fire Assimilation System (GFAS) is based on the assimilation of fire radiative power observed from the MODIS instruments aboard the Terra and Aqua satellites and provides daily global fire emission estimates at 0.5${}^{\circ }$ $\times$ 0.5${}^{\circ }$ and 0.1${}^{\circ }$ $\times$ 0.1${}^{\circ }$ resolution for 2003 onwards (http://eccad.sedoo.fr). The Fire Inventory from NCAR (FINN) version 1.5 is an updated version of the FINN daily global high-resolution inventory . In addition to GFED4s, we also used GFED4. Both version have adopted lower fuel consumption rates than the previous version GFED3 to better match field observations , but in GFED4s this decrease is compensated for by the addition of small (s) fire-burned area. It is available at http://www.globalfiredata.org. The average 2005–2013 global burned biomass is estimated at 1938, 2006, and 1438 Tg C year${}^{-\mathrm{1}}$ in GFAS, FINNv1.5 and GFED4, respectively (Table ; Fig. S2).

The isoprene emission estimates are compared to two bottom–up inventories: MEGAN-MACC and GUESS-ES (Fig. S3). MEGAN-MACC relies on the MEGANv2.1 model for biogenic volatile organic compounds (BVOC) and is based on the MERRA reanalysis fields . The emissions are provided at 0.5${}^{\circ }$ $\times$ 0.5${}^{\circ }$ resolution and on a monthly basis from 1980 through 2010. The GUESS-ES isoprene inventory is based on the physiological isoprene emission algorithm described by and updated by . It is coupled to the dynamic global vegetation model LPJ-GUESS and is driven by the CRU (Climatic Research Unit) monthly meteorological fields at 1${}^{\circ }$ $\times$ 1${}^{\circ }$ resolution between 1969 and 2009. Both inventories are available from the ECCAD data portal (http://eccad.sedoo.fr). The mean isoprene emission amounts to 452 and to 570 Tg year${}^{-\mathrm{1}}$, according to the GUESS-ES and to the MEGAN-MACC inventory, respectively (Table ), and both lie much higher than the a priori MEGAN-MOHYCAN inventory (343 Tg year${}^{-\mathrm{1}}$ average over 2005–2013). The large discrepancy between MEGAN-MACC and MEGAN-MOHYCAN datasets, both relying on the MEGAN emission model and the same version of basal emission factors (version 2011), can be explained to a large extent by (i) the neglect of soil moisture stress effects in MEGAN-MACC, (ii) a reduction by a factor of 4.1 of the basal emission factors for forests in Asia in MEGAN-MOHYCAN as suggested by field observations in Borneo , and (iii) the use of the crop distribution database of in MEGAN-MOHYCAN, along with the necessary adjustment of the other plant functional type distributions, leading overall to larger crop extent and lower total emissions.

The average global fire flux, expressed as burned biomass, is reduced from 2028 Tg C year${}^{-\mathrm{1}}$ (GFED4s) to 1653 Tg C year${}^{-\mathrm{1}}$ after optimization (Table ). Note that the inversion provides updated VOC emissions of HCHO precursors. However, to ease the comparison with other inventories, VOC emissions are converted to carbon emissions through the use of emission factors obtained from the compilation of (with 2011 updates). It should be acknowledged that the top–down estimates given here for fuel consumption may be affected by errors in the emission factors as well as by errors in the formaldehyde yields per VOC. The strongest emission decreases are induced over Africa (23 %), South America, and southeast Asia (15 %), whereas in Europe the fire fluxes are 12 % higher than in GFED4s. The reduced top–down emission agrees within 15 % with the GFED4 inventory (1438 Tg C year${}^{-\mathrm{1}}$) and is ca. 18 % lower than the GFAS and FINN global estimates. The lower a posteriori emissions in Africa are supported by the independent inventories, and the flux updates in Europe and Russia are in good agreement with the FINN fluxes. At tropical latitudes, the estimates from the independent inventories often exhibit large discrepancies, underscoring the large uncertainty of this source, while the top–down emissions lie generally within their range (Table ; Fig. ).

Interannual variability expressed as coefficient of variation, defined as the standard deviation of the emissions divided by the mean of the emissions, given for the a priori, for the OMI-based emission estimates, and for the independent emission inventories of biomass burning (upper panel), and isoprene emissions (lower panel) over the big regions defined in Fig. .

[Figure omitted. See PDF]

Temporal correlation between monthly MODIS fire counts, GFED4s, and OMI-based fluxes over the regions selected based on literature evidence for the occurrence of small fires. The regions are shown on the MODIS land cover map in Fig. S4.

Region	Coordinates	Fire type	MODIS vs. GFED4s	MODIS vs. OMI-based
N Africa	4–16${}^{\circ }$ N, 15${}^{\circ }$ W–15${}^{\circ }$ E	agricultural${}^{\text{a}}$	0.89	0.96
Maranhão	6${}^{\circ }$ S–2${}^{\circ }$ N, 44–52${}^{\circ }$ W	agricultural${}^{\text{a}}$	0.56	0.91
Mato Grosso	7–15${}^{\circ }$ S, 50–60${}^{\circ }$ W	small-scale deforestation${}^{\text{b}}$	0.95	0.97
SE US	30–36${}^{\circ }$ N, 75–100${}^{\circ }$ W	agricultural${}^{\text{a}}$	0.36	0.65
N China	30–40${}^{\circ }$ N, 111–122${}^{\circ }$ E	agricultural${}^{\text{a}}$	0.66	0.85
Indochina	6–27${}^{\circ }$ N, 87–110${}^{\circ }$ E	agricultural${}^{\text{a}}$ and small-	0.84	0.95
		scale deforestation${}^{\text{d}}$
Indonesia	10${}^{\circ }$ S–5${}^{\circ }$ N, 93–130${}^{\circ }$ E	agricultural${}^{\text{a}}$ and peat${}^{\text{c}}$	0.85	0.89
NW India	29–33${}^{\circ }$ N, 70–79${}^{\circ }$ E	agricultural fires${}^{\text{a}}$	0.75	0.87
Russia	52–60${}^{\circ }$ N, 55–90${}^{\circ }$ W	agricultural${}^{\text{a}}$ and peat${}^{\text{c}}$	0.81	0.94
Eq. Africa	14${}^{\circ }$ S–2${}^{\circ }$ N, 10–25${}^{\circ }$ E	agricultural${}^{\text{e}}$	0.96	0.99
E Australia	20–40${}^{\circ }$ S, 145–155${}^{\circ }$ E	agricultural${}^{\text{e}}$	0.55	0.86
Madagascar	12–26${}^{\circ }$ S, 43–50${}^{\circ }$ E	agricultural${}^{\text{e}}$	0.90	0.96

${}^{\text{a}}$ Region dominated by cropland according to the MODIS land cover change . ${}^{\text{b}}$ Region with a high number of small deforestation fires . ${}^{\text{c}}$ Region with peat fires selected based on . ${}^{\text{d}}$ Region with a high number of small deforestation fires . ${}^{\text{e}}$ Region where GFED4s emissions are predominantly associated with small fires .

The OMI-based fire emissions present a marked interannual variability, between a minimum of 1383 Tg C in 2013 and a maximum of 1966 Tg C in 2007 (Table ). Figure  illustrates the coefficient of variability, defined as the standard deviation of the emissions divided by the mean, which is a measure of the interannual variability of the emissions . The GFED4s coefficient is lowest over Africa (less than 0.15) and highest in South America, southeast Asia, Russia, and Australia (0.35–0.57). The low variability over Africa can be explained by the dominance of intense savanna fires that are highly regular throughout the years. According to the source attribution of GFED4s, deforestation fires are by far the prevailing source, responsible for 80 % of the total emission in South America, while the rest is due to savanna burning occurring in northeastern South America. The coefficient of variation of South American deforestation fires amounts to 0.74, pointing to the strong effect of climate variability, caused by, e.g., the strong El Niño–Southern Oscillation (ENSO), on the fire occurrence in the Amazon , and the rapid decline in deforestation rates over 2005–2013 . In addition, the estimated coefficient for savanna fires (0.41) is substantially higher than for the African savannas due to the strong variability of fire burning in northern South America . In Australia, savanna, grassland, and shrubland fires are responsible for the high interannual variability of the GFED4s inventory (0.42). In southeast Asia the contribution of peat burning to the total fire flux varies strongly from year to year (0–38 %) and drives the high coefficient of variability (0.45) . After inversion, the coefficient of variability is reinforced over Europe and Southern Hemisphere (SH) Africa but is reduced in the tropics, especially over southeast Asia and South America, where the decreased top–down variability is supported by comparisons with GFAS and FINN (Fig. ). This interannual variability of the optimized fluxes will be thoroughly discussed in the following sections (Sects. –).

Updates (percentage change from the a priori) in annually averaged biomass burning emissions suggested by the flux inversion for all years of the study period.

[Figure omitted. See PDF]

Updates (percentage change from the a priori) in annually averaged isoprene emissions inferred by the optimization for all years of the study period.

[Figure omitted. See PDF]

The global mean 2005–2013 isoprene emission is reduced from 343 to 272 Tg year${}^{-\mathrm{1}}$ after inversion (Table ), with the largest reductions inferred in Northern Hemisphere (NH) Africa and South America (ca. 30 %) and in the southeastern US (35 %). In contrast to the emission decrease suggested by satellite, the isoprene fluxes estimated by MEGAN-MACC and GUESS-ES are substantially higher, by 100 and 66 %, respectively. The interannual variation of the isoprene fluxes is low in all regions, with the coefficient of variability close to 0.04 in the tropics and up to 0.07 in extratropical regions. The satellite columns suggest stronger interannual variability over all regions, except in South America where it is slightly reduced. The interannual variation of isoprene fluxes is low for all inventories, generally stronger in MEGAN-MACC (up to 0.1) and weaker in GUESS-ES (Fig. ).

The monthly variation of the a priori and the OMI-based emissions is compared directly to MODIS Aqua (MYD14CM, 13:30 LT) fire counts (http://reverb.echo.nasa.gov) over 12 smaller regions selected based on literature evidence for the occurrence of small fires (Table ; Fig. S4). Higher spatial and temporal correlations are calculated after the inversion in all selected areas, especially over agricultural regions, like the southeastern US, eastern Australia, and Maranhão, where the correlation improves significantly from 0.36 to 0.65, from 0.55 to 0.86, and from 0.56 to 0.91, respectively. This shows that satellite HCHO observations do detect the contribution of small fires. It also explains the improved correlation of OMI-based emissions with GFAS and FINN in South America, northern Africa, and southeast Asia, regions where this contribution is important .

Interannual variation of burned biomass (in Tg C year${}^{-\mathrm{1}}$) over 2005–2013 from the a priori inventory (black), the satellite-based estimates (OMI in red), and from other bottom–up inventories (GFED4 in green, GFAS in orange, FINN in blue) over small regions defined in Fig. . Units are Tg C year${}^{-\mathrm{1}}$. The GOME-2-inferred estimate is shown as magenta circle.

[Figure omitted. See PDF]

Interannual variation of isoprene fluxes over 2005–2013 from the a priori inventory (black), the satellite-based estimates (OMI in red), MEGAN-MACC (in green), and GUESS-ES (in orange) over regions (red boxes) defined in Fig. . Units are Tg of isoprene per month. The GOME-2-inferred estimate is shown as magenta circle.

[Figure omitted. See PDF]

The ratio of the optimized to the a priori annual fluxes for biomass burning and isoprene emissions is presented in Figs.  and , respectively. The interannual flux variation is displayed in Figs.  and , and the seasonal variation of the fluxes over different regions (Sects. –) is shown in Figs.  and –. We present detailed results for regions where the satellite observations suggest important changes relative to the a priori fluxes.

Amazonian emissions

The OMI columns suggest important fire flux decreases during years with strong a priori fluxes, by 16 % in 2005, 22 % in 2007, and 32 % in 2010. The inferred flux reduction in 2010 is corroborated by earlier inversion studies constrained by GOME-2 HCHO columns , MOPITT CO observations , and a multisensor-based emission estimate above Mato Grosso . The top–down interannual fire variability is marked but less pronounced compared to the a priori, with the lowest emission inferred in 2009 (80.2 Tg C) and the highest in 2007 (387.4 Tg C; Fig. ), and it is corroborated by the GFAS and FINN inventories (Fig. ). The time and duration of the fire season is not modified by the optimization (Fig. ). The OMI-derived fluxes display the same pronounced seasonality as GFED4s, with fire emissions peaking between August and September and a rapid decline in October and November, as found in previous studies . The independent inventories, however, indicate generally higher fluxes than the top–down fluxes from October to January (Fig. ).

Seasonal and interannual variation of biomass burning emissions and isoprene emissions from bottom–up and top–down estimates over Amazonia (Fig. ). Units are Tg C per month for biomass burning fluxes and Tg of isoprene per month for biogenic emissions. The annual emission flux per inventory is given as inset numbers.

[Figure omitted. See PDF]

Regarding isoprene, the inversion infers generally lower fluxes than the a priori inventory for all years of the study period, with a 38 % mean annual reduction over 2005–2013, as illustrated in Figs.  and . The top–down annual isoprene flux ranges between 59 Tg in 2013 and 70 Tg in 2007, and the a priori interannual and seasonal variability is generally preserved after inversion (Figs. , ) and is similar in all inventories, with minimal emissions during the wet-to-dry season transition (April–June) and higher fluxes during the dry season (July–October). The peak-to-trough ratio is about a factor of 2 for the a priori and optimized fluxes, whereas it is weaker in the GUESS-ES inventory (1.6) and stronger in MEGAN-MACC (2.4). During the wet-to-dry transition season (April–June), top–down estimates from GOME-2 and OMI show better consistency than in the dry season (Fig. , ). An all-year-round emission decrease in most bottom–up inventories was also required in order to reconcile the GEOS-Chem model with SCIAMACHY and OMI HCHO columns . The strong seasonal variation and low emissions during the wet-to-dry transition are most likely due new leaf growth and lower flux rates from young leaves .

Comparison of a priori (black) and satellite-based (red) isoprene fluxes with ground-based flux measurements (colored numbered squares). The a priori and a posteriori isoprene fluxes are averaged over the full period from 2005 to 2013 for the grid. To ensure meaningful comparison, the ground-based flux measurements are corrected for the diurnal variation in isoprene fluxes; cf. Table S1 for more details.

[Figure omitted. See PDF]

Figure  shows a comparison of modeled isoprene fluxes with flux measurements from 12 field campaigns performed in the Amazon. The comparison accounts for the diurnal variations in the fluxes through correction factors used to scale the measured fluxes to daily averages (cf. Table S1). Direct comparisons between modeled fluxes and field measurements should, however, be considered with caution mainly due to the coarse resolution of the modeled emissions but also to the fact that flux measurements were often performed outside the study period (2005–2013). The observed isoprene fluxes exhibit strong local differences within the forest (up to 5 mg m${}^{-\mathrm{2}}$ h${}^{-\mathrm{1}}$, ), as well as significant differences from one day to another (up to 0.5 mg m${}^{-\mathrm{2}}$ h${}^{-\mathrm{1}}$; ), whereas they may exhibit differences of up to 1 mg m${}^{-\mathrm{2}}$ h${}^{-\mathrm{1}}$ associated with the use of different measurement techniques . Overall, the emission reduction inferred by the satellite observations lies within the variability of the field measurements, while the discrepancies between the observed fluxes are often larger than the differences between the a priori and a posteriori fluxes. The field studies generally agree on higher fluxes during the dry and the dry-to-wet transition season between July and December , while a recent field campaign suggests much lower fluxes (by ca. a factor of 3) compared to the top–down estimates, most likely related to a local effect of leaf flushing at the measurement location .

African emissions

In northern Africa, the biomass burning source is reduced by the inversion by 15–38 % for the different years and lies closer to GFED4, GFAS, and FINN estimates (Table ; Fig. ). In this region, both natural and agricultural fires peak in December, but the agricultural fire season, from September to May, lasts longer than the season of natural fires, which generally occurs between November and March . The OMI observations suggest a ca. 50 % emission decrease in the fire peak season, which is supported by comparisons with GFAS and FINN inventories (Fig. ), and a moderate increase from February to April when the agricultural fires are dominant and when the fraction of small fires is largest according to GFED4s. Higher emissions from February to April are also supported by GFAS and FINN, suggesting an even stronger shift in the fire season, with higher fire emissions lasting until May. The reduced emission amplitude and the longer burning season in northern Africa are corroborated by an inversion study using CO columns from the MOPITT instrument .

In Africa south of the equator, the OMI-based fire source is 23 % lower than the bottom–up estimate and lies closer to the estimates of GFED4, GFAS, and FINN (Table ; Fig. ). In terms of seasonal variation, the natural fires open the fire season between April and October, followed by agricultural fires lasting from June to November . The inversion infers 21 % lower emissions in the beginning of the fire season, when fires are predominantly natural, a reduction by 43 % during the fire peak between July and September, and 20 % higher emissions than GFED4s in October, when agricultural fires are the prevalent source (Fig. ). The GFED4s inventory allocates the maximum of the small-fire fraction to the peak of the fire season , resulting in an enhanced emission peak in July–August, rather than in September, as suggested by the OMI observations. This seasonality shift of the burning season was also reported in past inversion studies constrained by SCIAMACHY and GOME-2 HCHO and MOPITT CO observations .

As Fig.  but for Africa.

[Figure omitted. See PDF]

Southern Hemisphere Africa can be divided into two regions based on the fire source updates suggested by OMI (Fig. ). In its northern part, reduced emissions are systematically derived for all years, by up to 65 %, with regard to the a priori flux, whereas in its southern part (southern Africa in Fig. ), the emissions exhibit a stronger variability, increasing significantly until 2010 but remaining closer to the a priori in the subsequent years, as illustrated in Fig. . The a posteriori emissions during the peak fire season in September are found to be up to a factor of 3 higher than FINN and 50 % higher than GFAS and GFED4. The largest top–down flux in this region is inferred in September 2008, estimated to be 50 % higher than the a priori, due to record-high wildfires in Mozambique, South Africa, and Swaziland in that year .

The OMI observations suggest a decrease in isoprene fluxes over the African continent by ca. 20 % for all years of the target period, from 79 Tg year${}^{-\mathrm{1}}$ in the a priori to 63 Tg year${}^{-\mathrm{1}}$, as shown in Table . This decrease is very similar to the result obtained from an inversion study constrained by the NASA OMI HCHO retrieval product reporting an emission reduction in African isoprene fluxes, from 87 Tg year${}^{-\mathrm{1}}$ in the a priori to 68 Tg year${}^{-\mathrm{1}}$ through 2005–2009 . In the latter study, the flux decrease was strongest over equatorial and northern Africa, in very good agreement with the updates shown in Fig. . In a follow-up inversion study also based on OMI observations, invoked a reduction in MEGAN emission factors for broadleaf trees and shrub (ca. factor of 2) and woody savannas (20 %) in Africa in order to reconcile the model with the observations, whereas the reported comparisons with ground-based measurements suggested that even lower isoprene flux rates may be necessary.

The isoprene fluxes in northern Africa exhibit a weak interannual variability (Figs. , ). The OMI observations point to a mean (2005–2013) decrease of 26 % in this region with respect to the bottom–up estimate. The geographical extent of the emission updates (Fig. ) is in agreement with previous satellite-based results using SCIAMACHY and GOME-2 HCHO columns . As seen in Fig. , the seasonality of isoprene emissions in northern Africa is characterized by two emission maxima, driven by the two equatorial rainy seasons occurring from March to May and from August to November. The satellite columns indicate a change in the seasonal profile, from two equally strong emission maxima in April–May and in October to a peak in March and a weaker second peak in October–November (Fig. ). This agrees with the seasonality derived from GOME-2 observations and is similar to the seasonality change reported by . The stronger emissions in the first half of the year are also consistent with the independent inventories, whereas the secondary peak is better represented in the GUESS-ES inventory.

The isoprene emissions in southern Africa peak during the Southern Hemisphere summer, when both temperature and precipitation rates are higher (Fig. ). Both MEGAN-MACC and GUESS-ES emission estimates are about a factor of 2 higher than the top–down estimates. The discrepancy with MEGAN-MACC is partly explained by the neglect of the soil moisture stress effect (${\mathit{\gamma }}_{\text{SM}}$) in the standard version of the MEGAN-MACC model. Its inclusion in MEGAN-MACC was found to have a strong impact, leading to a flux decrease by 50 % on a global scale and even stronger decreases in Africa and South America . Interestingly, the inversion suggests a large increase in isoprene emissions (up to a factor of 2) southward of 15 ${}^{\circ }$S and particularly in the very dry southwestern part of the continent (west of ca. 30${}^{\circ }$ E), where the soil moisture stress effect is strongest in the MEGAN-MOHYCAN emissions (Fig.  and Fig. 2 in ). The spatial coincidence of the largest emission updates inferred by the inversion with the areas where the soil moisture stress effect is strongest is a first indication that its parameterization in MEGAN overestimates the impact of very low soil moisture on the emissions in dry subtropical environments like southern Africa (also Australia; see Sect. ). A second, even stronger indication is provided by the interannual variability of the emission updates in southwestern Africa (15–35${}^{\circ }$ S, 10–30${}^{\circ }$ E) shown on Fig. . These updates are indeed found to be well correlated ($r=0.81$) temporally with the factor by which the emissions are reduced due to the soil moisture activity factor ${\mathit{\gamma }}_{\text{SM}}$. In other words, the emission increments are largest when and where ${\mathit{\gamma }}_{\text{SM}}$ is lowest.

MEGAN simulates the isoprene response to soil moisture stress with a simple parameterization that shuts off isoprene emission when soil moisture drops to the level where plants can no longer draw moisture from the soil, known as the wilting point. While the MEGAN soil moisture stress effect uses a simple concept, the implementation is difficult due to the need to accurately model soil moisture, soil wilting point, and plant rooting depth. evaluated the MEGAN response to soil moisture stress by comparison to measured whole canopy isoprene fluxes and found that the algorithm performed poorly with the default soil wilting point but worked well when a more accurate value was used.

Interannual evolution of the factor by which the annual isoprene flux is reduced due to soil moisture stress vs. the isoprene flux increment inferred from OMI data (in %) in southwest Africa (top) and southern Australia (bottom).

[Figure omitted. See PDF]

As Fig.  but for southeast Asia.

[Figure omitted. See PDF]

Emissions in southeast Asia

The fire season in southeast Asia is characterized by a first peak in March, associated with aboveground vegetation burning in Indochina, and a second peak in August to October caused by peat combustion occurring in Indonesia (Fig. ). The GFED4s fluxes vary considerably across the years, ranging between a minimum of 123 Tg C (in 2011) and 277 Tg C (in 2006). The top–down estimates remain generally close to the a priori, except in 2006 and 2009, when the satellite observations suggest a significant decrease in the fluxes associated with peat burning in Indonesia by almost a factor of 3 (Fig. ). The optimized fluxes generally increase in March and decrease from August to October, while the amplitude of the seasonal pattern is reduced, with the emissions in March being generally larger than the peat burning emissions in August. In addition, the higher a posteriori correlation with monthly MODIS fire counts in Indochina (Table ) indicates an improved representation of the seasonal natural fires in March–April and agricultural waste burning in April–May .

In Indonesia, the fire season extends from June to November and comprises intense peat burning, in particular during extreme drought conditions caused by El Niño . The GFED4s estimates are generally lower than 100 Tg C year${}^{-\mathrm{1}}$ but significantly higher for El Niño years, e.g., 2006 (350.3 Tg C) and 2009 (191.6 Tg year${}^{-\mathrm{1}}$). The inferred flux drop in 2006 and 2009 is supported by GFAS and FINN, but in all other years both FINN and GFAS are relatively close to GFED4s. The lower 2006 flux suggested by the observed columns is corroborated by an independent carbon emission estimate based on burned area in a small region of Borneo in 2006 (Central Kalimantan, approximately 13 % of the Indonesian peatland) reporting peat fire emissions of 49 Tg C during the 2006 El Niño episode . This estimate is about half of the GFED4s value (109 Tg C) and closer to the OMI-based estimate of 33 Tg C for the same area and year. Note, however, that this independent estimate does not account for aboveground biomass burning.

As mentioned in the previous sections, the updated isoprene emissions are systematically decreased in tropical regions, by about 40 % on average in Amazonia and equatorial Africa (Fig. ), pointing to potentially overestimated emission factors used in the MEGAN model for tropical forests. In contrast to these regions, the emission reduction for the tropical rainforests of southeast Asia is much weaker ($<$ 20 %; Figs. , ) due to the lower basal emission rates incorporated in MEGAN-MOHYCAN based on OP3 campaign measurements in the rainforest of Borneo . The relatively small discrepancy between the model and the satellite HCHO columns in southeast Asia supports the use of lower isoprene flux rates for the Asian rainforests.

In China, most of the fires are agricultural and their emissions are generally low, except for the North China Plain (Fig. , ). The isoprene fluxes in China are also reduced after optimization, from 7.3 Tg year${}^{-\mathrm{1}}$ in MEGAN-MOHYCAN to 5.8 Tg year${}^{-\mathrm{1}}$ on average over the study period, but the decrease is stronger in south China, ranging between 27 and 45 % depending on the year. The emissions peak in summertime and present weak interannual variability (Fig. ), with a maximum in 2007 (2.6 Tg year${}^{-\mathrm{1}}$) and a minimum in 2010 (1.7 Tg year${}^{-\mathrm{1}}$; Fig. ). The OMI-based flux in 2010 is in good agreement with an earlier estimate inferred from GOME-2 HCHO observations (2.4 Tg year${}^{-\mathrm{1}}$; Fig. ) .

As Fig.  but for northern and southern Australia.

[Figure omitted. See PDF]

Australian emissions

Northern Australia is a major fire-prone area, where bushfires occur during many months every year . The peak of the fire season is observed between September and November, but its magnitude depends strongly on the year. The fire season begins between April and June, with the beginning of the dry season, is reinforced by the hot temperatures and winds of the subsequent months, and lasts until December. The OMI data suggest top–down fluxes close to the a priori in all years, except for 2011, when the emission maximum is decreased by about 25 % with respect to GFED4s (Fig. ), whereas the estimates from GFAS and FINN in this region differ by more than a factor of 10. In southern Australia (Fig. ), the fire fluxes are generally half those in northern Australia, and bushfires are again the main fire type in this region. This region, and in particular the state of Victoria, sometimes experiences extreme fire events, like the 2006–2007 bushfires, which were some of the worst on record, and the “Black Saturday” bushfires in February 2009. The satellite columns of HCHO lead to a significant reduction (Fig. ) in the fire emission during the aforementioned major fire events in comparison to the GFED4s inventory, in good agreement with the FINN estimates.

As Fig.  but for Europe and the southeast US.

[Figure omitted. See PDF]

The optimization indicates negative isoprene updates in the tropical and subtropical ecosystems of northern Australia, which are dominated by woodland and grasslands, and generally positive flux increments in the southern part of the continent, where temperate forests and grasslands are prevalent (Figs. , ). The mean reduction over 2005–2013 in northern Australia amounts to ca. 20 % with respect to the a priori (24.4 Tg year${}^{-\mathrm{1}}$) and is supported by the inversion study based on GOME-2 HCHO columns as shown in Fig. , pointing to possibly overestimated emission factors assumed in MEGAN for tropical ecosystems. In southern Australia, the a posteriori isoprene fluxes are increased by about 20 % on average over the study period, from 12.5 to 15 Tg year${}^{-\mathrm{1}}$, and show small interannual and seasonal variability (Fig. ). Although the MEGAN-MACC emissions are much higher than the other inventories over Australia, a sensitivity calculation accounting for the soil moisture stress activity factor in the MEGAN-MACC model resulted in a substantial flux decrease of about 70 % with respect to the reference MEGAN-MACC simulation , stressing the important role of soil moisture stress in these very dry environments. As for southern Africa, the OMI-based inversion over southern Australia enhances the emissions where and when ${\mathit{\gamma }}_{\text{SM}}$ reaches its lowest values (Figs.  and ). As discussed above, the poor performance of the parameterization could be partly due to misrepresentations of driving variables (soil moisture content) or soil characteristics (wilting point, rooting depth). The use of satellite-derived soil moisture or solar-induced fluorescence could be a promising way for improving the soil stress estimation in the future.

Global distribution of annual isoprene emission trends over 2005–2013 according to the a priori (left) and top–down inventory (right) expressed in percentage year${}^{-\mathrm{1}}$.

[Figure omitted. See PDF]

Midlatitude emissions

In Europe, the fire season peaks in summertime and a secondary peak is also recorded in spring, mainly due to emissions from agricultural waste burning (Fig. ). The optimized fluxes lie generally close to the a priori except in 2006 and 2007, when the OMI observations point to higher fluxes (by 40–50 %) than in GFED4s during the emission peak. The strong fluxes in April–May 2006 and in summer 2007 were due to numerous agricultural fires that occurred in the Baltic countries, western Russia, Belarus, and Ukraine and to intense biomass burning in southern Europe. The increase in the top–down estimates in 2007 is in line with the reported increase based on IASI CO columns . The top–down estimate agrees well with GFED4s during the devastating fires in the Moscow area in July–August 2010, whereas previous studies reported values which were a factor of 2 , 3 , and 10 higher than the older GFED3 inventory , which was about 60 % lower than GFED4s in this region.

Regarding isoprene fluxes over Europe, the satellite observations suggest an average increase by 15 % in western Europe (from 2.9 to 3.4 Tg year${}^{-\mathrm{1}}$) and by 33 % in eastern Europe (from 3.9 to 5.2 Tg year${}^{-\mathrm{1}}$), whereas the inferred increase is significantly stronger during extremely hot summers, like in 2007 and 2010. Indeed, in July 2007 Greece experienced the hottest summer on record since 1891 , with temperature anomalies of $+$5 ${}^{\circ }$C compared to the 1961–1990 mean, and in July 2010, the hottest summer since 1500 was recorded in western Russia, with temperature anomalies of $+$6 ${}^{\circ }$C with respect to the 1961–1990 mean (http://www.ncdc.noaa.gov/temp-and-precip).

The concurrence of pyrogenic and isoprene emissions in the mid- and high latitudes of the Northern Hemisphere during summertime is, however, an inherent difficulty in the derivation of top–down emissions by inverting for HCHO columns. HCHO being an intermediate compound in the oxidation of both pyrogenic and biogenic hydrocarbons, it cannot be excluded that the HCHO column enhancements associated with higher isoprene emissions have in reality a pyrogenic origin and vice versa. The inversion scheme relies strongly on the a priori emission distributions and errors in the retrievals, and, thereby, errors in the geolocation of fire hot spots in the bottom–up inventories could propagate as errors in the source attribution, in particular for intense fire events associated with summer heat waves.

In the southeastern US, a major isoprene-emitting region, the top–down fluxes are systematically reduced compared to the initial inventory, by 35 % on average, with the strongest decrease (50 %) inferred in 2013. Similarly to the a priori, the a posteriori estimates peak in 2011 and are lowest in 2013. This variability is primarily related to temperature changes, with recorded temperature anomalies of $+$3 ${}^{\circ }$C in 2011 and $-$1.5 ${}^{\circ }$C in 2013 with respect to the 1961–1990 mean (http://www.ncdc.noaa.gov/temp-and-precip/). The use of GOME-2 HCHO columns to constrain the inversion in 2010 results in an excellent agreement with the OMI-based fluxes (Figs.  and ), whereas both optimizations suggest a slightly modified seasonal profile, with a primary peak in June and a secondary one in August. The need for lower emissions in the southeastern US compared to the MEGAN model has been put forward by past studies based on satellite observations of HCHO from GOME, SCIAMACHY, and OMI sensors . The MEGAN-MACC and GUESS-ES estimates are in very good agreement with the a posteriori fluxes in terms of magnitude, although in some years the peak emission is delayed by 1 month (Fig. ).

Emission trends

The global distribution of isoprene emission trends over 2005–2013 according to the bottom–up emission inventory and as suggested by the inversion of satellite data is displayed in Fig. . Although deriving long-term emission trends from satellite data may be very useful for diagnosing global and regional change, particular caution is required when interpreting the results, since physical changes in the satellite instruments over time may result in artificial drifts in the observations. In the case of OMI HCHO columns, special efforts were made to reduce the effects of the row anomaly issue (http://projects.knmi.nl/omi/research/product/rowanomaly-background.php) on the retrieved HCHO columns, in order to ensure the suitability of the data for addressing trend studies . Nevertheless, it appears difficult to avoid any time-dependent instrumental effect possibly impacting the interannual variability of emissions reported in this section.

Amazonia experienced a rapid decline in pyrogenic emissions, estimated to be $-$7 % year${}^{-\mathrm{1}}$ in the a priori GFED4s inventory and $-$8 % year${}^{-\mathrm{1}}$, in the OMI-based emissions as a result of the trend in OMI columns calculated during the dry season ($-$3.2 % year${}^{-\mathrm{1}}$ in August–September). This trend in HCHO columns was attributed to a strong decline in deforestation rates in the Amazon and especially in Mato Grosso and Rondônia, where the cover loss in evergreen broadleaf forests decreased by more than 80 % between 2002 and 2009 . The isoprene emission trend over Amazonia, which is close to negligible ($-$0.2 % year${}^{-\mathrm{1}}$) in the a priori inventory (Figs.  and  ), becomes negative after optimization ($-$2.1 % year${}^{-\mathrm{1}}$). The derivation of biogenic emission trends in this region is made difficult by the magnitude and strong interannual variability of biomass burning. However, a decline in isoprene emissions is supported by the negative trend ($-$1.3 % year${}^{-\mathrm{1}}$) in the observed HCHO columns during the wet season (November–April), when biomass burning plays only a very minor role. This result is difficult to interpret. Recent findings based on satellite surface reflectance data (more precisely, normalized difference vegetation index, or NDVI, data) point to diminished vegetation greenness since 2000 due to a precipitation decline across large parts of Amazonia, especially northern Brazil . However, most of these changes occurred between 2000 and 2005, whereas the precipitation rates and NDVI values were comparatively more stable afterwards, and the leaf area index from MODIS Collection 5 (MOD15A2 composite, http://modis.gsfc.nasa.gov/data/dataprod) either increased (by less than 0.5 % year${}^{-\mathrm{1}}$) or showed no trend over 2005–2013 over most of Amazonia (see Fig. S5).

Over northern Africa, during the fire season (November–February) a decreasing trend of about 3 % year${}^{-\mathrm{1}}$ over the study period is derived for the OMI-based fire fluxes (Fig. ), close to the GFED4s trend ($-$3.2 % year${}^{-\mathrm{1}}$), whereas the corresponding trend of FINN ($-$2.4 % year${}^{-\mathrm{1}}$) is somewhat weaker. This trend is most likely related to negative trends observed in burned area in northern Africa , owing to land use changes (conversion of savannah into cropland) and to changes in precipitation, driven by the El Niño–Southern Oscillation .

In Siberia, the strongly positive isoprene emission trend of the bottom–up inventory (3.8 % year${}^{-\mathrm{1}}$) (Fig. ) is a result of the warming temperature trends in this region (0.12 ${}^{\circ }$C year${}^{-\mathrm{1}}$ over 55–75${}^{\circ }$ N, 40–120${}^{\circ }$ E, based on ECMWF ERA-Interim temperature data over 2005–2013). The model incorporates both the direct effect of warming on the MEGAN temperature response of the emissions and the indirect effect through the increase in leaf area index (LAI), which reaches 3 % year${}^{-\mathrm{1}}$ in northern Siberia (Fig. S5). The inversion leads to an even higher trend (4.2 % year${}^{-\mathrm{1}}$), induced by the strongly positive trend in the HCHO observations (4.2 % year${}^{-\mathrm{1}}$) over 2005–2013 in this region, suggesting a stronger response of isoprene emissions to warming. This result is in line with reported ecosystem measurements in the Arctic exhibiting a higher emission response of biogenic emissions than observed at more southern latitudes . Higher temperatures may also favor the extension of forests inducing even higher isoprene emissions . According to MODIS land cover data, the forest fraction in this region has increased from 31 % in 2005 to 36 % in 2012 .

Opposite a priori isoprene trends are calculated in western and eastern Europe over the study period, $-$2.5 and 3.2 % year${}^{-\mathrm{1}}$, respectively, mostly related to the temperature and solar radiation trends. The OMI observations corroborate these trends, showing a negative trend in western Europe ($-$1.1 % year${}^{-\mathrm{1}}$) and a positive trend in eastern Europe (0.4 % year${}^{-\mathrm{1}}$). The calculated trends after optimization are moderately enhanced: $-$3.3 and 3.9 % year${}^{-\mathrm{1}}$ in western and eastern Europe, respectively. Besides climate parameters, land use changes may also contribute to the increasing column and emission trend in eastern Europe. Based on MODIS land cover data , the forest fraction increased at a faster pace in eastern than in western Europe (1.1 % year vs. 0.9 % year${}^{-\mathrm{1}}$) and the crop fraction decreased more rapidly ($-$0.5 % year${}^{-\mathrm{1}}$) in eastern than in western Europe ($-$0.4 % year${}^{-\mathrm{1}}$).

Over the southeastern US, the slightly negative trend in the summertime isoprene fluxes in the a priori ($-$0.3 % year${}^{-\mathrm{1}}$) becomes much more pronounced after inversion ($-$4 % year${}^{-\mathrm{1}}$), induced by the downward trend in the OMI HCHO columns ($-$2.5 % year${}^{-\mathrm{1}}$) over 2005–2013 . Except for this trend, the interannual variability of the top–down emissions over this region is similar to the a priori (Fig. ). The long-term decline could be in part an artifact resulting from the well-documented downward trend in NO${}_x$ abundances over the United States , which could significantly decrease formaldehyde production over time if the yield of HCHO per isoprene molecule is substantially lower at low NO${}_x$ level than at high NO${}_x$. The ground-level NO${}_{\mathrm{2}}$ concentrations have decreased by as much as a factor of 2 over the eastern US based on OMI and in situ measurements between 2005 and 2012 . The NO${}_x$ dependence of the HCHO yield is taken into account in the calculations presented in this study, but the modeled decrease in planetary boundary layer (PBL) NO${}_x$ level in the eastern US is lower (ca. $-$30 %) than observed ($-$50 %) during 2005–2012. Furthermore, the low-NO${}_x$ oxidation mechanism remains incompletely characterized, especially regarding the further degradation of primary oxidation products, leaving open the possibility of a significant overestimation of the HCHO yield at low NO${}_x$, even though a recent analysis of airborne measurements over the southeast US indicated that state-of-the-art oxidation mechanisms can reproduce the NO${}_x$ dependence of prompt HCHO formation inferred from the measurements . If confirmed, an overestimation of the HCHO yield at low NO${}_x$ could also help to explain the negative trend in top–down isoprene emission over western Europe (Fig. ). More importantly, it would imply a general underestimation of our top–down emissions in low-NO${}_x$ environments and tropical forests in particular.

In south China, the negative summertime trend ($-$0.7 % year${}^{-\mathrm{1}}$) in HCHO columns drives a change in the sign of the 2005–2013 isoprene trend: from 0.1 % year${}^{-\mathrm{1}}$ in the a priori to $-$1.6 % year${}^{-\mathrm{1}}$ in the OMI-based fluxes. The very small a priori emission trend results from a combination of compensating effects: on the one hand, declining trends in the ERA-Interim photochemically active radiation (PAR) ($-$0.33 % year${}^{-\mathrm{1}}$) and temperature ($-$0.03 K year${}^{-\mathrm{1}}$), and on the other hand, an increasing trend in leaf area index (1 % year${}^{-\mathrm{1}}$, cf. Fig S5) and a decline in crop extent in south China, suggested by the land use database of and supported by MODIS land cover data . However, a recent land cover database suggests that the extent of crops has increased in eastern China in the last 30 years . In addition, the declining trend in PAR was also derived from ERA-Interim data complemented by surface radiation measurements . The crop expansion and declining PAR were proposed to cause a negative isoprene trend in south China and likely explain the observed negative trend in HCHO.

Conclusions

Global distributions of pyrogenic and biogenic VOC fluxes between 2005 and 2013 were derived using the adjoint inversion scheme built on the IMAGESv2 global CTM and HCHO column abundances retrieved from the OMI sounder. The inversion suggests a moderate decrease (ca. 20 %) in the global average emissions of both pyrogenic and biogenic VOCs relative to the a priori emissions used in the model. The main findings of this study are presented below.

• The global top–down fire fluxes exhibit strong interannual variability, ranging between ca. 1400 Tg C year${}^{-\mathrm{1}}$ (2011) and 2000 Tg C year${}^{-\mathrm{1}}$ (2007). The a priori interannual variability is generally well preserved, but the inferred estimates are ca. 250 to 450 Tg C lower than the a priori, depending on the year, with the largest decreases suggested over Africa, South America, and southeast Asia (23 %). The top–down emissions are better correlated with MODIS than GFED4s fire counts in regions with small fires, indicating that the associated emissions may be too low in GFED4s and that they can be derived by the OMI-based inversion.

• The inversion suggests (i) important fire flux decreases (15–30 %) in Amazonia during years with strong a priori emissions, (ii) about a 50 % emission decrease during the peak fire season in northern and southern Africa, (iii) generally increased emissions in Indochina and decreased fluxes in Indonesia during intense fire events related to El Niño years, (iv) a significant flux reduction during the major bushfires in Australia, and (v) flux increases during the devastating fires in southern Europe in 2007.

• Changes in fire seasonal patterns are suggested, in particular in southeast Asia and Africa. In southeast Asia, the seasonal amplitude is reduced after inversion, with enhanced emissions due to aboveground vegetation burning in March and weaker emissions due to Indonesian peat burning in August. The inversion suggests generally increased fluxes due to agricultural fires over Africa and decreased emission maxima due to natural fires.

• Significant reductions in isoprene fluxes are inferred in tropical ecosystems (30–40 % in Amazonia and northern Africa), suggesting overestimated basal emission rates in these areas. The top–down fluxes generally increase over Eurasia, especially during heat waves in summer (e.g., western Russia in 2010), suggesting a possibly stronger emission response to high temperatures than currently assumed.

• The inversion suggests large isoprene emission increases (up to 100 % locally) over areas most affected by the soil moisture stress parameterization in MEGAN, in particular in southern Africa and southern Australia. The inferred isoprene increments present a strong interannual correlation with $\mathrm{1}/{\mathit{\gamma }}_{\text{SM}}$, i.e., the factor by which isoprene emissions are reduced due to soil moisture stress in MEGAN ($r\ge 0.7$), indicating that the soil moisture parameterization leads to overly decreased isoprene fluxes.

• The isoprene emission trends are found to be often enhanced after inversion. Positive trends in top–down isoprene emissions are inferred in Siberia (4.2 % year${}^{-\mathrm{1}}$) and eastern Europe (3.3 % year${}^{-\mathrm{1}}$), likely reflecting forest expansion and the warming trend. Negative trends are derived in Amazonia ($-$2.1 % year${}^{-\mathrm{1}}$), south China ($-$1 % year${}^{-\mathrm{1}}$), the United States ($-$3.7 % year${}^{-\mathrm{1}}$), and western Europe ($-$3.9 % year${}^{-\mathrm{1}}$). The top–down trends should be considered with caution due to possible drifts in the satellite data. In several instances, however, they are supported by independent evidence from literature studies. Trends in NO${}_x$ emissions may play a significant role given their possibly large influence on formaldehyde yields, which remain imperfectly characterized and deserve more attention, especially at low NO${}_x$.

For simplicity and to avoid excessive computational costs, a detailed error assessment of the a posteriori emission estimates is not addressed in this work. Nevertheless, sensitivity inversions conducted in an earlier study, also based on OMI columns for 2010, have shown that the inferred fluxes were generally weakly dependent on the choice of key model and inversion parameters and lay within 7 % of the standard inversion results . Recent developments in the representation of vertical profiles of smoke released by open fires , in the partitioning of burned biomass into emitted trace gases , and in the spatiotemporal variability of emission factors indicate additional sources of uncertainty that could impact the top–down fluxes and should therefore be carefully assessed in future studies.

Data availability

The OMI HCHO column data are publicly accessible at http://h2co.aeronomie.be (De Smedt and Van Roozendael, 2016). The OMI-based inventories are freely available at the GlobEmission web portal (http://www.globemission.eu; Stavrakou et al., 2016b) and at the BIRA-IASB emission portal (http://emissions.aeronomie.be; Stavrakou et al., 2016c). The MEGAN-MOHYCAN bottom-up inventory is also available at http://emissions.aeronomie.be.

The Supplement related to this article is available online at doi:10.5194/acp-16-10133-2016-supplement.

Acknowledgements

This research was supported by the Belgian Science Policy Office through the PRODEX projects ACROSAT, by the European Space Agency (ESA) through the GlobEmission DUE project (2011–2016), and by the MARCOPOLO project (2014–2016) funded by the European Commission within the Seventh Framework Programme (grant agreement: 606593). Edited by: A. Pozzer Reviewed by: two anonymous referees