We used two well-established biomass burning emission datasets to interpret and validate TROPOMI-inferred combustion characteristics and efficiencies: the Global Fire Emission Database version 4 with small fires (GFED4s; van der Werf et al., 2017) and the Global Fire Assimilation System version 1 (GFAS; Kaiser et al., 2012). Both datasets provide global fire emission fluxes for a large number of chemical species but use different methods.

GFED4s is based on the Carnegie–Ames–Stanford approach (CASA) biogeochemical model (Potter et al., 1993) to predict the amount of above- and below-ground biomass at monthly temporal resolution. The MODIS Collection 5.1 MCD64A1 500 m burned area satellite product (Giglio et al., 2013) is used to estimate the daily dry matter combustion rate at 0.25∘ × 0.25∘ spatial resolution from 2001 up to 2016. GFED4s also includes 1 × 1 km2 thermal anomalies (active fire counts) from Terra and Aqua MODIS and 500 × 500 m2 surface reflectance observations, providing a statistical estimate of the burned area associated with small fires (Randerson et al., 2012; van der Werf et al., 2017). The GFED4s flux estimates from 2017 onward (used in this study) are not directly derived from the burned area product because the underlying MODIS algorithm was upgraded from Collection 5.1 to Collection 6. Instead, flux estimates are simply derived from MODIS active fire detections and their FRP and the climatological ratio between them derived from the overlapping 2003–2016 period. The GFAS product calculates emissions by assimilating FRP observations from the MODIS Terra and Aqua satellites and is tuned to match the dry matter combustion rate of GFED3 per biome (Kaiser et al., 2012). The version we used provides daily emissions at 0.1∘ × 0.1∘ spatial resolution.

Both biomass burning products are combined with EFs to translate the derived dry matter combustion rate to specific trace gas and aerosol emissions. These EFs are based on a large number of trace gas measurement campaigns in the field, in the air or in the laboratory and are subdivided for dominant biome/burning categories without specifying any variability in space and time. GFAS uses the older EF dataset compiled by Andreae and Merlet (2001) with additional updates from the literature and EFs of peatland fires from Christian et al. (2003). In this dataset boreal and temperate forest fires form together a single category named extratropical forest fires (ETFs). GFED4s uses EFs largely based on the dataset compiled by Akagi et al. (2011). This dataset is based on trace gas measurements from fresh smoke sampled in close proximity to the fire source and cooled to ambient temperature but with minimal photochemical processing. This provides a better representation of the initial emissions without chemical disturbances (to aid assessment of biomass burning in atmospheric chemistry models). The Akagi et al. (2011) dataset makes a distinction between boreal and temperate forest fires. For boreal fires they used the average of airborne and ground-based measurements that is roughly equivalent to assuming 70 % of dry matter consumption originates from smoldering combustion. Therefore, the EFs for the boreal latitudes are relatively high for carbon monoxide (EFCO: 127.0 g kg-1) and low for nitrogen oxides (EFNOx: 0.9 g kg-1). The EFs for the temperate fires are 88.0 and 1.9 g kg-1, respectively, for CO and NOx and represent a larger fraction of flaming combustion similar to the ETF category used in the Andreae and Merlet (2001) dataset. Other variations in EFs between Akagi et al. (2011) and Andreae and Merlet (2001) are due to variations in the averaging and weighting methods of the measurements. In addition, GFED4s includes sub-grid cell partitioning of burned area to account for different fire types within a grid cell, which affects the grid-average emissions of CO and NOx. Because NO is usually the most abundant N species emitted into the atmosphere and because NO and NO2 are rapidly interconverted in the atmosphere, both datasets report EFs for NOx as NO. Henceforth, the EFs are reported in units of mmol kg-1 and mol kg-1 for EFNOx and EFCO, respectively, to make ratios of EFs of similar magnitude to the ratios of column densities measured by TROPOMI (see Sect. 2.2). Table 1 shows an overview of EFs of CO and NOx used by GFAS and GFED4s.

The spatial stratification of the different biomass burning categories is apparent in the ratio between NOx and CO EFs. This ratio exhibits a distinct “fingerprint” that carries information on combustion efficiency, combustion practice, and fuel type. In this study, we call this dimensionless metric the emission factor ratio (EFR = EFNOx / EFCO). EFR is a relative measure of how many millimoles of NOx are released into the atmosphere for each mole of CO. This metric is a proxy for the modified combustion efficiency (MCE) parameter that is often used in fire emission quantification studies but is more difficult to derive from space given the relatively small departures of CO2 concentrations over biomass burning regions from background conditions. The MCE is defined as ΔCO2 / (ΔCO + ΔCO2) to indicate combustion efficiency of a fire by measuring the amount of excess in CO2 in comparison to total emitted C from CO2 and CO (Yokelson et al., 1999). Table 1 gives the EFR for the different combustion types based on the ratio between the EFs of NOx and CO used by GFED4s and GFAS. Figure 1a and b show the spatial distributions of EFR in both datasets. For GFED4s, we subdivided EFR into three different categories: high EFR above 50 for savanna fires, EFR between 10 and 50 for temperate forest fires, tropical deforestation fires and agricultural waste burning, and EFR lower than 10 for boreal and peatland fires. High EFR is thus related to the flaming type of combustion that is hotter and more efficient as it produces relatively less CO alongside CO2 and relatively more NOx by combustion of N in the biomass itself. Conversely, low EFR is generally related to slow smoldering type of combustion. The EFR categories are similar for GFAS; however, due to differences between the EFs datasets (for reasons discussed in the previous paragraph), EFRs are classified differently: high EFR above 30 for savanna fires, EFR between 10 and 30 for ETF fires, tropical deforestation fires and agricultural waste burning, and EFR lower than 10 for peatland fires.

Highlighted in Fig. 1 are various regions of interest studied in this paper with strong seasonal occurrences of biomass burning. Regions that have been selected for detailed analysis using TROPOMI are 2 boreal fire regions in North America, 1 boreal fire region in Siberia, 5 savanna fire regions on the African continent, 1 savanna fire region in Australia, 2 peatland fire regions in Indonesia, and 15 regions in South America to more specifically study spatial gradients in combustion efficiency between tropical deforestation and savanna fires.

The TROPOMI instrument was launched on 13 October 2017 onboard the Sentinel-5 Precursor satellite to monitor the chemical composition of the atmosphere (Veefkind et al., 2012). It measures a range of trace gases at unprecedented spatial resolution with a daily global coverage. Section 2.2.1 and 2.2.2 provide further details about the TROPOMI operational level-2 column density data products of carbon monoxide (XCO) and nitrogen dioxide (XNO2). Figure 2 shows a few examples of monthly and daily average maps of XCO and XNO2 for a number of biomass burning regions together with CO emissions from GFED4s (Sect. 2.1). Enhancements in XCO and XNO2 correspond well to local fire emissions based on an independently derived burned area. Note that the chemical lifetime of NOx is much shorter than for CO (minutes to hours vs. weeks to months). The main chemical driver during daytime is the photochemical balance between NO2 photolysis and NO oxidation by ozone converting NO into NO2 and makes NO2 a robust measure for NOx. The NOx lifetime is limited by the conversion of NO2 to HNO3 in reaction with hydroxyl (OH) radicals. The short chemical lifetime results in a precise alignment between the enhancements of XNO2 and the location of fire emissions, while enhancements of XCO are more affected by atmospheric transport due to its longer chemical lifetime. These differences in lifetime can cause biases in the joint analysis of XNO2 and XCO and its ratio. That limits our ability to make direct quantitative comparisons between EFs and column densities. Nonetheless, assuming the lifetime does not vary greatly from fire to fire and from region to region, it is probable that it does not affect our ability to detect variations in fire characteristics around the world. This limitation is further discussed in Sect. 4 of the paper.

For our analysis, we selected important hotspots of biomass burning according to the GFED4s database. To prevent contamination with urban trace gas emissions, large population centers were avoided. The regions are outlined in Figs. 1 and 2. Within each region we collected all the available XCO and XNO2 data that passed the filters explained in Sect. 2.2.1 and 2.2.2 for up to 3 consecutive months depending on the timing and duration of the fire season.

To derive the regional enhancements in XCO and XNO2 relative to the background, ΔXCO and ΔXNO2, respectively, we used two main sampling methods depending on the scale and severity of the fires in the region. A statistical bulk method (SBM) is used for regions that are characterized by extensive seasonal burning over a large area and where upwind background levels of XCO and XNO2 are difficult to define. A local sampling method (LSM) is used for regions where we could identify local fires and plumes of enhanced trace gas abundance for which the wind direction and background column density upwind of the fires could be determined. Each method is discussed in more detail in Sect. 2.3.1 and 2.3.2, respectively. With daily estimates for ΔXNO2 and ΔXCO we were able to derive a new dimensionless metric: the mole density ratio (MDR = ΔXNO2 / ΔXCO). The MDR is the atmospheric equivalent of EFR and provides a remotely sensed proxy for biomass combustion efficiency.

To evaluate the methodology of the joint analysis of TROPOMI XCO and XNO2, and in particular the SBM sampling technique, we used the Weather Research Forecasting model version 4.0 coupled with chemistry (WRF-Chem). The main purpose was to investigate whether the sampling techniques can provide estimates of XCO, XNO2 and MDR that are distinctly different between four combustion types. Synthetic WRF-Chem simulations were performed using a single domain located over the northern part of South America stretching over 6000 km in the east–west direction and 3900 km in the north–south direction (see Fig. 4). We used a horizontal resolution of 30 × 30 km2 with 32 vertical levels. We chose the “tropical” suite of physics options that includes the Yonsei University (YSU) scheme for planetary boundary layer physics (Hu et al., 2013), WSM six-class scheme for microphysics (Hong and Lim, 2006), Tiedtke scheme for cloud physics (Tiedtke, 1989), and rapid radiative transfer method (RRTM, Mlawer et al., 1997) for shortwave and longwave radiation. We included gas-phase chemistry mechanisms from the Regional Acid Deposition Model version 2 (Stockwell et al., 1990) but without aerosol chemistry. Boundary and input meteorological fields for September 2018 came from the NCEP FNL Operational Global Analysis dataset prepared on a 1∘ × 1∘ grid every 6 h. Boundary and initial trace gas concentrations were taken from the CAMS model and were interpolated to WRF vertical levels. GFED4s (see Sect. 2.1) provided the biomass burning dry matter combustion rate based on real fire events for South America in September 2018. These were multiplied by a set of EFs to acquire a synthetic estimate of biomass burning emissions for the entire domain that is associated with a single combustion type: either boreal fires, mixed peat fires, tropical deforestation fires or savanna fires. For the EFs of the mixed peat fires we assumed 60 % is combusted by tropical deforestation (to mimic overstory consumption) and 40 % is combusted by peat soils. As a consequence, each of these four emission estimates carried a different (but in space constant) EFR between NOx and CO over the entire domain for each of the respective fire types. In addition, we used a spatial characterization of injection height profiles based on space-based stereo-height information from smoke plumes (Martin et al., 2018).

WRF-Chem was executed four times for September 2018 under the exact same conditions and settings described above, except that we used for each run one of the four modified biomass burning emissions. Each simulation provided hourly three-dimensional fields of CO and NO2 concentrations for the entire month. Close to the center of the domain we collected within a 5∘ × 5∘ region each day at 14:00 local time (half an hour later than the actual TROPOMI overpass) all CO and NO2 data. These were translated into XCO and XNO2 column mole densities using a daily mean estimate of TROPOMI's averaging kernel (AK) of the two respective species, derived from September 2018 data over the same 5∘ × 5∘ collection region. This ensured realistic differences in column sensitivity for simulated XCO and XNO2, even under cloudy conditions. The collection of column densities was used to derive daily ΔXCO, ΔXNO2, and MDR using the three SBM sampling methods discussed in Sect. 2.3.1.

By combining ΔXCO and ΔXNO2 data from all investigated regions we can identify a four-way split in regional combustion activity and efficiency (see Fig. 6a and Table 2). According to the TROPOMI data, there is a group of boreal and peatland regions that emit relatively much CO per gram of NOx in comparison to the other regions and that are in the literature typically characterized as smoldering fires. For these regions we determined MDR values ≤ 1.4. At the opposite end of the spectrum are the savanna regions on three different continents that emit relatively much NOx generated from nitrogen in the biomass and possibly from thermal decomposition of atmospheric N2 at very high combustion temperatures. The MDR values for savanna fires are much higher than for the boreal and peatland fires and range between 3.6 and 6.2 among the different savanna regions and sampling methods (SBM for South America and Africa, LSM for Australia). In between these two extremes lie four different South American deforestation regions with MDR values that range between 1.6 and 2.5, which is less efficient than the savanna combustion but still more efficient than boreal and peat combustion. In particular for the boreal regions we observe a 3 to 6 times larger increase in ΔXCO relative to ΔXNO2 than for the deforestation and savanna regions. It consistently translates to a much lower monthly average MDR (< 1) for the boreal fires on two different continents, with higher MDR values over boreal North America compared to boreal Siberia.

A similar four-way split in regional combustion characteristics is shown in Fig. 6b using the four synthetic WRF-Chem simulations, each driven by a different set of modified biomass burning emissions to mimic different ecosystems with frequent occurring fires (see Sect. 2.4). The estimates in ΔXCO and ΔXNO2 were derived with the SBM and, while day-to-day variability is quite substantial just like in the real TROPOMI data in Fig. 6a, it demonstrates that the SBM can provide a robust monthly average estimate of MDR. The MDR estimates and the relative differences between the four different fire types compare quite well with the actual derived signals in TROPOMI. The simulated savanna fires have consistently the highest MDRs, which are about twice as high as the simulated tropical deforestation fires and 3 to 4 times as high as the simulated peatland and boreal fires, respectively.

The pattern of combustion signatures in MDR determined with TROPOMI across the different fire types compares well with the spatial patterns in GFED4s EFR (see Fig. 7a). The Pearson correlation coefficient between estimates of MDR and EFR across the five continental areas implies that a strong linear relationship exists (r=0.90). It demonstrates that the regional combustion efficiency that is detected with TROPOMI is generally consistent with the worldwide spatial distribution of EFs used by GFED4s (mostly based on Akagi et al., 2011). For the savanna fires we have consistently the highest estimates for MDR and EFR. The MDR and EFR for tropical deforestation fires are about half those of savanna fires. The peatland fires in turn have 3 times lower MDR and EFR. The lowest values in MDR and EFR (4 times lower than savanna fires), which suggest the least efficient type of biomass combustion, are shown for the North American and Siberian boreal regions. Note that the retrieved linear relationship in Fig. 7a between GFED4s EFR and TROPOMI MDR is similar to the linear relationship and slope found between EFR and MDR based on the four synthetic WRF-Chem simulations. Even with the aforementioned caveat that these simulations are simple in design, it does demonstrate quite convincingly that satellite and ground-based measurements of trace gas ratios are related to one another through a simple linear relationship.

The pattern of combustion signatures in MDR is somewhat different from EFR of GFAS (see Fig. 7b), which is mostly based on the older set of EFs from Andreae and Merlet (2001). The Pearson correlation coefficient between MDR and EFR is lower (r=0.49) and the slope across the different fire regions is less steep. One reason is that the boreal fire characteristics in GFAS were lumped together with the temperate fires into a single category called extratropical fires, which reflects a much smaller smoldering combustion component and thus a higher EFR (see Table 1). As a consequence, Siberian boreal NOx emissions are in GFAS approximately 4 times larger than in GFED4s. This confirms the findings in the CAMS validation activity: comparisons with GOME-2 indicated largely overestimated boreal NO2 concentrations in the CAMS forecasts driven by GFAS (Ramonet et al., 2019). Another reason for the mismatch is that the combustion efficiency of savanna fires in GFAS is on the same order of magnitude as the combustion efficiency in the boreal regions, which seems less realistic given the current body of EFs measurements from savanna ecosystems that claim the contrary. For peatland fires in Borneo, Indonesia, GFAS assumes a much larger fraction of smoldering combustion than GFED4s also. This large fraction of smoldering combustion would probably be more accurate during the drier years, e.g., during El Niño, than for the relatively wet year 2018.

The joint analysis of ΔXCO and ΔXNO2 column densities in Fig. 8a–c shows a clear distinction between the deforestation and savanna regions during the 3-month dry season. For the deforestation burning region we observe much larger increased levels of ΔXCO (up to 0.030 mol m-2) relative to ΔXNO2, mostly in September, indicating a substantial fraction of smoldering combustion later in the season. This is consistent with ground-based measurements of tropical forest fires that usually show a persistent smoldering phase that can continue for days, in particular when woody debris is ignited that is piled together (Carvalho et al., 2001; Morton et al., 2008). The day-to-day variability in ΔXCO is large but remains quite proportional to the variability in ΔXNO2 in September, indicating that the relative amount of smoldering and flaming combustion remains relatively constant throughout the month. For the deforestation region ΔXCO and ΔXNO2 are correlated in September with Pearson correlation coefficients of r=0.82, r=0.84, and r=0.68 for, respectively, the SBM, SBM_alt1 and SBM_alt2. The estimates of ΔXNO2 are quite similar for the deforestation and the savanna region (between 0.005 and 0.05 mmol m-2). However, lower ΔXCO values (up to 0.010 mol m-2) point to a much cleaner combustion of savanna biomass, which is common for savanna fires, where the flaming phase typically dominates (Andreae and Merlet, 2001). In contrast to deforestation fires, ΔXCO and ΔXNO2 for savanna fires are less correlated (r<0.66); i.e., the trace gases do not change consistently on a day-to-day timescale.

In September, at the peak of the fire season, the monthly average MDR is significantly different for the two fire types, irrespective of the sampling method used (see Fig. 8c). The MDR estimates for deforestation fires range between 1.06 and 1.55, whereas the MDR estimates for savanna fires are higher, ranging between 2.97 and 3.17. The separation between deforestation and savanna fires remains also quite robust to the different bulk sampling methods used (SBM, SBM_alt1 or SBM_alt2). For each fire type, the monthly average ΔXCO and ΔXNO2 estimates of the three sampling methods lay well within the 1σ uncertainty level of each method (the standard deviation of day-to-day variability in ΔXCO and ΔXNO2). The differences in MDR between the SBM and SBM_alt1 are quite small in August and September (within 15 % for both fire types), indicating that the SBM is not so sensitive to a possibly non-Gaussian shape of the distribution. Much larger differences in MDR exist between the SBM and SBM_alt2 (up to 35 % for deforestation fires), with much larger day-to-day variability in MDR using SBM_alt2. It indicates the background estimates of XCO and XNO2 in the adjacent regions are not necessarily consistent with the background densities derived with the Gaussian fit. These uncertainties can be attributed to the incorrect assumption that the sampled trace gases are Gaussian distributed or to the somewhat arbitrary choice of the background regions for SBM_alt2, which do not necessarily characterize the true background of each day in September. As discussed in Sect. 2.3.1, it is not our goal to determine the best possible background estimate, which is difficult for these kinds of regions that are continuously surrounded by fires. Instead, we opted for a mathematical method that is consistent in application for both trace gases and provides a reasonable proxy for regional fire-induced trace gas enhancements.

The influence of instrument precision on the ΔXCO, ΔXNO2 and MDR estimates was also quantified. The instrument precision of XCO under relatively clear-sky conditions is primarily a function of surface albedo (Landgraf et al., 2016). The low albedo of the tropical deforestation region yields an average XCO precision of 0.0014 mol m-2, which is 2 times larger than for the savanna region with an average XCO precision of 0.0007 mol m-2. The precision of XNO2 is around 0.011 mmol m-2 for both regions and is dominated by air mass factor uncertainties under polluted conditions (Lorente et al., 2017). The contribution of instrument/retrieval precision to the uncertainty in ΔXCO, ΔXNO2 and MDR was estimated using a synthetic distribution of daily measurements for both trace gases in September 2018. This was done as follows: assuming no systematic errors or biases, each single TROPOMI measurement was randomly perturbed by its precision value or decreased by its precision value or remained unchanged. Subsequently, the SBM was applied on this distribution, yielding an uncertainty range around ΔXCO, ΔXNO2 and MDR values due to instrument or retrieval noise (see Fig. 9). While the XNO2 uncertainty is quite substantial in comparison to XCO, their contribution to ΔXNO2, and in the resulting MDR, it is actually quite small on a monthly timescale. There is barely any overlap in the range of monthly MDR estimates of the deforestation and savanna fires, which means that differences in combustion characteristics easily exceed the precision. Only for smaller ΔXCO and ΔXNO2 values will TROPOMI's precision become a more limiting factor in terms of signal-to-noise ratio and could explain the lower correlation between daily ΔXCO and ΔXNO2 noted before for savanna fires.

The September daily estimates of ΔXCO and ΔXNO2 are shown in Fig. 10a for the deforestation region along with the smoothed estimates of the MDR and the smoothed GFED4s and GFAS CO fire emissions. In September, when the enhancements of the two trace gases and the CO emissions are at their maximum, the MDR is also at its lowest level around 1.5, indicating persistent, less efficient combustion in the region (relatively less NO2 and more CO release into the atmosphere). The same set of daily estimates are also shown for the savanna region in Fig. 10b. The peaks in ΔXCO and ΔXNO2 correspond to peaks in the GFAS CO emissions, and the MDR is quite constant between 3.0 and 3.5, i.e., about twice as high as the MDR for the deforestation region.

It is worthwhile noting that the GFAS and GFED4s emission products do not necessarily align well in Fig. 10a and b. As mentioned in Sect. 2.1, these two products use different methods and data products to derive fire emission estimates. The difference between GFAS and GFED4s reflects the uncertainty in the amount of CO emitted. This is apparent for the savanna region, where the daily GFAS CO emissions are on average a factor of 4 larger than the GFED4s emissions. The 2018, GFED4s estimates are derived using active fire detections and their FRP and a simple relationship based on the climatological FRP and GFED4s ratio based on the 2003–2016 period for each 0.25∘ × 0.25∘ grid cell. GFAS, on the other hand, used the FRP associated with those active fires and is tuned to match GFED3 emissions, not for each grid cell, but for each biome (Kaiser et al., 2012). In addition, the used EFs between the two products are different (see Table 1). For the large deforestation fires the estimates of CO emissions from GFAS and GFED4s are more similar, although the timing is somewhat different (see Fig. 10a).

We also demonstrate the detection of subcontinental-scale gradients in biomass burning efficiency from space. The September average MDR (derived with the standard SBM) is shown in Fig. 11a for 15 5∘ × 5∘ regions south of the Amazon River that cover the two main biomes: tropical rainforest (dominated by deforestation fires) and Cerrado savanna (dominated by savanna fires). For the same 15 regions we also derived the EFR from the monthly average GFED4s NOx and CO fire emissions shown in Fig. 11b. These emissions are essentially based on the Akagi et al. (2011) EF database that is compiled from ground and airborne measurements. Although the relationship is not perfect, MDR responds fairly linearly to EFR (r=0.59 shown in Fig. 11c). A very similar west–east relationship also exists between the 15 EFR estimates from GFAS and MDR (r=0.61, not shown). However, GFAS EFR values are generally lower across the entire domain, which can be traced back to differences in EFs (see Table 1). In general, we see lower MDR and EFR values over the western part of the domain (where deforestation fires dominate) and higher MDR and EFR values over the eastern part of the domain (where savanna fires are more prevalent). The relationship becomes much more significant (r=0.89) if three outliers are excluded from the analysis. One of these outliers represents a mountainous region in the southwestern corner of the domain, where fire activity was much lower than elsewhere in the domain. The other two outliers represent regions that are located in between the tropical rainforest and the Cerrado savanna (highlighted in Fig. 11a and b). It is possible that the MDR and EFR do not align well at these locations because the biome-specific EFs are not representative of a more complex transition region. One would expect here more diversity in burning practices, vegetation types and climate conditions, resulting in a mixture of different burning characteristics that are not accounted for in the EF in either GFED4s or GFAS, which are based on coarser-resolution land cover data. Another factor that plays a role is atmospheric transport, as it affects the column mole densities that are measured downwind of the dominant fire type. The wind direction was predominantly from the east in September (see Fig. 4) and could as well carry the savanna-like combustion characteristics from the easternmost regions more towards the west (see Fig. 11a).