FINNv2.5 first determines the burned area from daily satellite detections of active fires. FINNv2.5 uses MODIS detections (nominal 1 km2 resolution; Giglio et al., 2006), as in FINNv1, and adds the option to use active fire detections at 375 m resolution from the Visible Infrared Imaging Radiometer Suite (VIIRS; Csiszar et al., 2014), on board the Suomi National Polar-orbiting Partnership (Suomi-NPP) satellite, alone or in combination with MODIS active fire data. The use of VIIRS 375 m detections is a major advancement from the use of MODIS-only fire detections, as this product better captures small fires. VIIRS detections are available from 2012 to the current year.

The MODIS Collection 6 (MCD14DL) and VIIRS active fire products are obtained from NASA's Fire Information for Resource Management System (FIRMS) data portal (https://firms.modaps.eosdis.nasa.gov/download/, last access: 19 June 2022). The MODIS product provides the location, overpass time (Coordinated Universal Time, UTC), and confidence of daily fire detections. Data confidence in the MODIS product is specified by a numerical scale of 0 % to 100 %. Detections with a confidence specification of less than 20 % are eliminated from our calculation in FINNv2.5, as was done in earlier FINN versions. Daily global coverage is not accomplished at latitudes between approximately 23.5∘ N and 23.5∘ S, due to the observational swath width. To account for the lack of daily observations, fire detections in these equatorial regions only are counted for a 2 d period; each fire is assumed to continue into the next day. Regardless of whether the detection is from MODIS or VIIRS, the intent is to repeat the fire at the same location, since there is not a clear alternative at this time in these global regions, as described by Wiedinmyer et al. (2011).

With its improved spatial resolution of 375 m, the VIIRS product provides a more sensitive detection of the fires of relatively small areas, fully global coverage, improved mapping of large fire perimeters, and improved nighttime performance relative to MODIS fire detections (Schroeder et al., 2014). The higher detection rates of small fires can be particularly important for areas of the world such as southeast Asia, where burning of agriculture is common, and in the southeastern USA, where there is a large amount of managed burning. Detection confidence is provided by the VIIRS product and is specified by three categories, namely low, nominal, and high. In the FINNv2.5 preprocessor, detections with a confidence specification identified as low are eliminated from the analysis. We only include data attributed to thermal anomalies from vegetation fires (type = 0; i.e., other thermal anomaly types associated with active volcanos or other static land sources are filtered from the product).

The processing of the two simultaneous fire products in FINNv2.5 does not lead to double-counting of the fires; the FINNv2.5 method determines the spatial union of all adjacent detections for a given day as the daily burned area of a fire, as described below in Sect. 2.2. The identity of the sensor is not relevant for the determination of the burned area, as long as the pixel size for each detection is correctly represented (i.e., 0.14 km2 for VIIRS and 1 km2 for MODIS). FINNv2.5 and earlier versions do not account for the effects of burning from earlier days for fire detections at the same location; i.e., the fraction burned is consistent with unperturbed vegetation on the first day, regardless of the persistence of fire activity, and emissions scale directly with the detection of a thermal anomaly in the same location over multiple days.

The active fire products report the time of acquisition in Coordinated Universal Time (UTC). In contrast to previous versions of the model, the FINNv2.5 preprocessor uses local time in the specification of the date of a fire detection in order to facilitate comparisons of emissions estimates with observational data. 2Local time=UTC+Nearest_Integer(Longitude/15)

FINNv1 estimates burned area for each fire pixel identified individually, and the nominal pixel size for the MODIS fire detections, 1 km2, is assumed per detection. Spatially overlapping detections are eliminated from further analysis. It was recognized that for large fires in forested regions, an array of multiple discrete detections is typically reported, and an estimate of a contiguous area that represents the total area burned by a fire is needed. We improved the burned area estimate in FINNv2 to better represent the area associated with each fire.

A fire event in the western USA is shown in Fig. 1a to illustrate the new approach for estimating the area burned. For FINNv2.5, each reported active fire detection (Fig. 1b) is assigned a square area of 0.14 km2 from VIIRS or 1 km2 from MODIS (Fig. 1c), based on the nominal horizontal resolution of the data (denoted as the instrument resolution square). Detections determined to be in proximity with one another are aggregated by two different approaches, depending on the land cover type and forest cover. Initially, it is assumed that multiple detections by adjacent pixels in a satellite sensor array are part of a larger fire, and these detections are merged. The scan and track sizes of the satellite pixel are provided by the fire detection product and define the actual resolution of the fire detection. The scan and track sizes for each fire detection are used for identifying groups of records that represent contiguous or overlapping detections. A rectangle with easterly and northerly sizes equal to 110 % of the scan and track sizes is established for each detection (denoted as the detection rectangle; Fig. 1d), with the objective of identifying adjacent neighboring detections but not for direct application to the burned area estimation. Fire detections are identified as being from one larger fire when any of the satellite detection rectangles overlap. To minimize an overestimation of the burned area, a convex hull is generated between corresponding pairs of detection rectangles that directly intersect. The union of pairwise convex hulls from a cluster forms an extended fire polygon that represents the tentative estimated burned area for a single fire event or group of nearby fires for the day (Fig. 1d). This approach effectively fills any gap between instrument resolution squares.

For each of the extended polygons, the MOD44B v006 MODIS/Terra VCF annual product (https://lpdaac.usgs.gov/products/mod44bv006/, last access: 13 July 2022) is overlaid (Sect. 2.3), and the average tree cover fraction is determined (Fig. 1e). For forested areas with tree cover ≥50 %, as determined by the VCF product, the merged polygons are accepted as the final burned area estimate. Otherwise, the merging is not used, and instead, an alternative, more conservative, approach is applied to determine the burned area for the region. This alternative approach is used to prevent overestimation of emissions in regions with many small fires, as in the savanna fires in sub-Saharan Africa. The alternative polygon aggregation is achieved by aggregating nearby detections only when the instrument pixels themselves are intersecting (Fig. 1c) and therefore not with the extended detection footprints (Fig. 1d). The result is an aggregation algorithm that is repeated with a smaller set of detections to determine the alternative conservative set of polygons (denoted as the conservative fire polygon). The final burned area polygons in Fig. 1f show examples of a composite of polygons based on these two different aggregation approaches. Note that at the bottom of Fig. 1f there is a region with less than 50 % tree cover that has a smaller final burned area polygon estimate than the extended polygon (blue). In contrast, the polygon in the center of Fig. 1f, which is a forested area with more than 50 % tree cover, uses the extended polygon determined in Fig. 1d (further information about this aggregation is given in Sect. S1 in the Supplement).

Subsequently, the final burned area polygons are subdivided using a Voronoi tessellation algorithm in order to develop emissions estimates by land cover classification, as described in Sect. 2.3 and 2.4. Each of the undivided final burned area polygons are assigned a unique fire ID to enable users to group emissions estimates from a presumed single fire event.

The NASA MODIS VCF product provides estimates of the percentage of bare surface, herbaceous, and forested cover at a horizontal resolution of 250 m (Fig. S1). For each fire area, the subdivided polygons described in Sect. 2.2 (Fig. 1g) are overlaid on the vegetation cover data from the MOD44B v006 MODIS/Terra VCF annual product (https://lpdaac.usgs.gov/products/mod44bv006/, last access: 13 July 2022; Fig. 1e). The VCF data for the prior year are chosen, so that the VCF before any land cover changes due to fire are used in the emissions estimation process. The VCF raster is clipped to the geometry of the fire polygon, and the averages of the VCF tree, herbaceous, and bare cover are calculated for each fire polygon.

FINNv2.5 uses the Terra and Aqua combined MODIS LCT MCD12Q1 Version 6 data product with the International Geosphere–Biosphere Programme (IGBP) classification scheme (https://lpdaac.usgs.gov/products/mcd12q1v006/, last access: 13 July 2022) as its default land cover information. Figure 2 shows the global distribution of land cover applied as the default in FINNv2.5. Each subdivided polygon (for example, in Fig. 3) is assigned the factional coverage of 1 of 16 land cover classifications (Table 1). Similar to the application of the VCF information, land use data from the previous year are used.

Use of the LCT and VCF products in FINNv2.5 is an improvement on FINNv1. FINNv1 used one static map of LCT and VCF (from 2002) for any year processed. FINNv2.5 employs year-specific MODIS LCT and VCF maps that change annually. Furthermore, the specific vegetation assignments for each subdivided polygon enable different vegetation types and coverage to be represented across larger fires. These input data and processes enable better representation of the vegetation that is burned.

All fire polygons are assigned to 1 of 13 global regions (Wiedinmyer et al., 2011) used to assign fuel loadings (Sect. 2.4). This completes the first component of the FINNv2.5 modeling framework and results in a file of daily burned areas and associated land cover information.

The next step of the model framework is the emissions calculation. In this step, the daily burned area and associated vegetation information (described above) are assigned associated fuel loadings. Using the same process described by Wiedinmyer et al. (2011), where the biomass burned is assigned based on land cover type and global region (B), the fraction of the biomass that is burned (FB) is assigned as a function of tree and herbaceous cover, emissions factors (EFs) are determined based on land cover, and daily pollutant emissions estimates are calculated following Eq. (1). Overall, the emissions calculation process follows this framework, as described by Wiedinmyer et al. (2011), with the following exceptions.

Similar to earlier FINN versions (Wiedinmyer et al., 2011), the 16 IGBP land cover classifications of the LCT product are mapped to consolidated vegetation types, depending on the land cover class and latitude that distinguish tropical, temperate, and boreal forests (Table 1). The consolidated vegetation types used in FINNv2.5 are grassland and savanna, woody savanna or shrubs, tropical forest, temperate forest, boreal forest, temperate evergreen forest, and crops.

The fuel loading, or the potential maximum amount of biomass available to be burned (B(x) in Eq. 1), is assigned by generic vegetation type and global region (Table 2). Selected values were updated for FINNv2.5 from earlier versions of FINN, based on van Leeuwen et al. (2014). The fuel loading for crops was updated to 902 g m-2, based on an average from the literature (Akagi et al., 2011; van Leeuwen et al., 2014; Pouliot et al., 2017). Specific crop types are not identified in the version described here.

For North America, FINNv2.5 utilizes the fuel loadings for coarse/woody and herbaceous vegetation by land cover type derived from the Fuel Characteristic Classification System (FCCS) of the U.S. Department of Agriculture Forest Service (https://www.landfire.gov/fccs.php, last access: 23 December 2018), as described by McDonald-Buller et al. (2015). These fuel loadings (Table 3) have priority over the regional default fuel loadings (Table 2).

Emissions factors are assigned based on the generic vegetation type. Since the original release of FINNv1 in 2011, there have been many studies to measure emissions factors from wildland fires. We have updated the emissions factors from FINNv1.5 with the results from recent publications (Table 4).

Wiedinmyer et al. (2011) provided mappings from the total mass NMOG emissions values calculated by the FINN model to the surrogate species of three chemical mechanisms commonly used in chemical transport models, namely MOZART (Model for Ozone and Related chemical Tracers), SAPRC99 (Statewide Air Pollution Research Center Mechanism), and GEOS-Chem (Goddard Earth Observing System with Chemistry). The mapping of NMOG emissions to the MOZART-T1 chemical mechanism was created for FINNv2.5, based on recent published emissions data and updates in the chemical mechanisms (Table 5). To convert the total NMOC to the individual mechanism species, the total NMOG mass estimated by FINN should be multiplied by the mapping value (mole species in kg NMOG-1; e.g., Table 5) to assign the molar emissions of the surrogate species. The mappings to the SAPRC99 and GEOS-Chem have not been updated and are the same as those described by Wiedinmyer et al. (2011).

The FINNv2.5 model was run in two ways to produce emissions for evaluation and assessment, namely (1) for comparison with the previous version of FINN (FINNv1.5) using MODIS-only fire detections and calculated starting in 2002 (FINNv2.5(MODIS)) and (2) by using both MODIS and VIIRS fire detections and calculated starting in 2012 (FINNv2.5(MODIS + VIIRS). The FINNv2.5 emissions files are freely available for use by the community (Wiedinmyer and Emmons, 2022). Results through 2019 are presented in this work.

Estimates from several versions of FINN are compared to other emissions inventories, such as the Global Fire Emissions Database (GFED; van der Werf et al., 2017), Fire Energetics and Emissions Research (FEER; Ichoku and Ellison, 2014), the Global Fire Assimilation System (GFAS; Kaiser et al., 2012), and the Quick Fire Emissions Dataset (QFED) version 2.5 (Darmenov and da Silva, 2015; Fig. 4). The comparisons are done by the global region that follows (Giglio et al., 2010; Fig. 4).

Figure 4 compares the annual averaged (2012–2019) emissions of key pollutants from several versions of FINN and the other inventories by region. For all emitted species, FINNv2.5(MODIS + VIIRS) global emissions are higher than, and approximately double, those predicted by FINNv1.5. This is the case, even when only MODIS fire detections are considered. The increase in emissions from previous versions is primarily due to the new processing of the area burned. In previous versions, the fire area was determined from a satellite detection pixel only; the updated version here also includes the composite of many detections into larger areas of fire activity (Sect. 2.1). The inclusion of VIIRS into the FINNv2.5(MODIS + VIIRS) inventory globally adds approximately 25 % above the FINNv2.5(MODIS) processing for all emitted species. Further, emissions of NMOGs and the individual species that make up NMOGs (e.g., CH2O and C2H6 in Fig. 4) are increased significantly due to the use of updated emissions factors from recent field campaigns. Previous studies have shown low biases in FINN regional and species-specific estimates; for example, CO in the western USA (Pfister et al., 2011) and Australia (Desservettaz et al., 2022) and particulate carbon in North America (Carter et al., 2020). The updated version is expected to correct some of these prior biases.

FINNv2.5(MODIS + VIIRS) emissions estimates are overall at the higher end of the range of annual global total emissions provided by our sample of other commonly used emissions inventories, likely due to a combination of the aggregated burned areas and the fact that FINNv2.5(MODIS + VIIRS) includes fire information from VIIRS, which captures more small fires. However, depending on the pollutant emitted, a comparison across different emissions inventories shows varied results. For example, both CO2 and CO global annual emissions from FINNv2.5 (MODIS and MODIS + VIIRS) are higher than QFED, but black carbon (BC) and ammonia (NH3) are lower in FINNv2.5 than QFED. The primary drivers of these differences are the assumed fuel type burned and associated emissions factors. This difference in emissions amounts between inventories is more variable when looking regionally and year to year (e.g., Figs. 5 and S4).

In general, the year-to-year variabilities in the annual fire emissions are consistent between different inventories, and mainly the magnitudes of the emissions differ (Figs. 5, S5, and S6). FINNv2.5(MODIS + VIIRS) is often among the inventories that produces the highest CO emissions in all 14 global regions. Some notable exceptions are boreal North America and boreal Asia, where GFAS and sometimes GFED estimate higher emissions. This is likely due to the representation of smoldering peat fires in the high latitudes that are represented as a specific vegetation type in the GFED and GFAS inventories but not in FINNv2.5. (Note that Kiely et al., 2019, developed a parameterization used in FINNv2 to represent regional peat emissions in Indonesia, but this was not included in FINNv2.5.) Similarly, in equatorial Asia, GFED and GFAS estimate the highest CO emissions for the years when much of the tropical peatland burned. The magnitude of FINNv2.5(MODIS + VIIRS) CO emissions in Australia and New Zealand has increased relative to FINNv1.5, which compares better to downwind surface measurements of instantaneous mixing ratios (Desservettaz et al., 2022). However, regional FINNv2.5(MODIS + VIIRS and MODIS) emissions remain lower than three other emissions inventories (GFAS, QFED, and FEER) in Australia and New Zealand. A similar result is seen over the Middle East, which suggests a potential role of extremely dry landscapes in causing inter-inventory differences. For most other regions, FEER is the only inventory that produces emissions that are as high in magnitude, or sometimes higher, as FINNv2.5(MODIS + VIIRS).

The seasonal change in the regional fire emissions for CO in FINNv2.5(MODIS + VIIRS) is shown in Fig. 6, with other inventories in Fig. S4. Globally, fire emissions peak in August–September, with the largest emissions in Southern Hemisphere Africa and Southern Hemisphere South America. As mentioned above, GFED and GFAS show an increase in boreal North America in July and August that is not as prevalent in FINNv2.5(MODIS + VIIRS).

FINNv2.5(MODIS + VIIRS) also has an emissions peak in March, which is driven primarily by emissions in Southeast Asia. March–April is a peak fire season in the Northern Hemisphere tropics, and in mainland Southeast Asia, the season is driven primarily by small, agricultural fires. FINNv2.5(MODIS + VIIRS) uses VIIRS fire detections, which detect these small fires to a greater extent than MODIS. Most inventories show this second peak in emissions during March and April; however, it is not seen in GFED, nor is it as pronounced in the other inventories (Fig. S6). Consequently, determining the cause of different fire emissions in Southeast Asia is a target for future research. Although the magnitude of the regional emissions in FINNv2.5(MODIS + VIIRS) is 2–3 times higher than FINNv1.5, the seasonality is similar.

As shown above, the emissions estimates from the different fire emissions models can vary substantially in time and space. It is difficult to know which emissions estimates most closely represent reality. One way to assess the emissions is to use them as input to an atmospheric chemistry model and calculate pollutant concentrations that can then be compared to in situ measurements and satellite observations. To further evaluate FINNv2.5(MODIS + VIIRS) and understand its uncertainties and limitations, we performed a simulation with the Community Atmosphere Model with Chemistry (CAM-chem), a component of the Community Earth System Model (CESM2.2.0; Emmons et al., 2020). For this simulation, the temperature and winds of CAM-chem are nudged to the Modern-Era Retrospective analysis for Research and Applications, version 2 (MERRA-2) reanalysis fields. Anthropogenic emissions are from the Copernicus Atmosphere Monitoring Service (CAMS) v5.1 inventory (Soulie et al., 2023), and biogenic emissions are calculated online with the Model of Emissions of Gases and Aerosols from Nature (MEGAN) version 2.1 (described in Emmons et al., 2020). Results for the year 2018, after several years of spin-up, are shown here.

The model results with FINNv2.5 emissions are compared to the CO column density from the Measurement of Pollution in the Troposphere (MOPITT) instrument retrievals (version 8, Level 3, gridded monthly joint product; MOP03J.008; Deeter et al., 2019), and aerosol optical depth (AOD) from MODIS (Level 3 gridded monthly global product; MOD08_M3 and MYD08_M3; MODIS Atmosphere Science Team, 2017; Figs. 7 and 8). For the comparisons to the MOPITT retrievals, the model CO profiles are transformed with the MOPITT averaging kernels and an a priori profile. The joint retrieval product combines the thermal-infrared (TIR) and near-infrared (NIR) radiances to provide greater sensitivity to the boundary layer than the TIR retrievals alone.

Overall, CAM-chem driven with FINNv2.5(MODIS + VIIRS) overestimates satellite-observed AOD and CO over the Amazon basin and central Africa during the 2018 fire season (Figs. 7–9). This result suggests that FINNv2.5(MODIS + VIIRS) overestimates fire emissions over the Amazon basin and central Africa in the 2018 fire season. This overestimation could be due to a number of reasons, including inaccurate ecosystem identification (e.g., tropical forest rather than shrublands or less wooded landscapes) and/or fuel loading assignments. The model also overestimates AOD, but not CO, over Australia and northern Africa. These discrepancies are likely due to overestimated dust emissions in the model simulation rather than overestimated fire emissions from FINNv2.5(MODIS + VIIRS).

There are important fire regions in which the model predictions and observations agree. For example, the CAM-chem results using FINNv2.5(MODIS + VIIRS) simulate column-measured CO in August 2018 (Fig. 9) for the Pacific Northwest well. This is consistent with a previous study, which evaluated CAM-chem with regional refinement over the Pacific Northwest with aircraft observations during the Western wildfire Experiment for Cloud chemistry, Aerosol absorption and Nitrogen (WE-CAN; Tang et al., 2022) and found that simulated CO concentrations agreed reasonably with aircraft measurements.

Despite updates to the input data, parameters, and processing, FINNv2.5 emissions estimates remain uncertain. Uncertainties are caused by the natural variability associated with the various inputs to the model and the model assumptions and processes used to create the estimates. Uncertainties may arise due to missed fire detections caused by cloud or smoke cover, timing, and incomplete global coverage from the polar orbiting satellite paths. Furthermore, the assumed vegetation type in active fires is highly variable, and the use of different vegetation maps can introduce large changes in the predicted emissions. The assumed fuel loading, fraction burned, and resulting fuel consumption can be highly variable in space and time, whereas the model assumes best-guess average values for generic ecosystems by global region. The emissions factors used also add some uncertainty to the estimates, which is particularly highlighted in the emissions estimates of particles and reactive gases. Other issues arise when VIIRS and MODIS are used in combination to drive the emissions estimates, compared to the MODIS-only version. We recognize that the addition of a second dataset from VIIRS increases the emissions compared to those estimated using MODIS only.

Wiedinmyer et al. (2011) estimated an uncertainty that is a factor of 2 in the FINNv1 estimates. Other efforts have assigned uncertainty to fire emissions estimates (e.g., Kennedy et al., 2020; Soares et al., 2015; Urbanski et al., 2011); however, limitations in our ability to directly measure fire emissions (fluxes) prevent a comprehensive, global evaluation of existing inventories. Discrepancies between model predictions and evaluations with model output, in situ measurements, and satellite observations can help identify the processes in the models that drive the uncertainties and the regions across the globe that are the most important and uncertain. Results from the evaluation presented here suggest that high uncertainties in emissions occur in South America and southern Africa and in southeast and equatorial Asia. Emissions across boreal North America should also be assessed; wide variations in organic carbon (OC) and BC emissions in this region lead to significant uncertainty in the ability to estimate air quality and climate impacts from biomass burning (Carter et al., 2020). Furthermore, while CO and AOD are often used to assess aerosol and CO emissions estimates, the emissions of other important pollutants are more highly variable across inventories and should be further constrained.