An evaluation of advanced baseline imager fire radiative power based wildfire emissions using carbon monoxide observed by the Tropospheric Monitoring Instrument across the conterminous United States

The GOES-R ABI detects fires at a nominal spatial resolution of 2 km (at nadir). GOES-16, the first GOES-R satellite positioned at 75.2° W, has been operationally running since December 2017. GOES-16 ABI scans the CONUS every 5 min with a pixel area increasing from 4.8–16.8 km² as satellite view zenith angle (VZA) varies from 29–72°. GOES-16 ABI detects fires using the 3.9-μm and 11.2-μm bands and provides for each fire pixel the detection time, geolocation, fire temperature, fire area, FRP (MW per pixel), and fire mask quality flag (Schmidt et al 2013). FRP is calculated using radiances of a fire pixel and its ambient background non-fire pixels in the 3.9-μm band (Schmidt et al 2013). The fire detection performance of GOES-16 ABI has been evaluated using the 30-m Landsat-8 and 375-m VIIRS (the Visible Infrared Imaging Radiometer Suite) active fire data and ground fire records (Hall et al 2019, Li et al 2020), which shows that GOES-16 ABI is able to confidently detect fires with FRP >34.5 MW (Li et al 2020). Compared with the 750-m VIIRS FRP, GOES-16 ABI FRP is relatively larger in individual fires but generally comparable at a regional scale across the southeastern CONUS (Li et al 2020). This study obtained GOES-16 ABI active fire data during July 2018—October 2019 from NOAA Comprehensive Large Array-data Stewardship System (CLASS) (https://www.avl.class.noaa.gov/; last accessed on 3 April 2020).

Sentinel-5P TROPOMI observes global CO at a nominal spatial resolution of 7 km. Sentinel-5P crosses the equator at ∼13:35. TROPOMI senses earth reflected radiances over a scan angle range of ±54° with a swath width of ∼2600 km (Veefkind et al 2012). TROPOMI CO is retrieved using radiances in the shortwave infrared bands (SWIR, 2305–2385 nm) based on the Shortwave Infrared CO Retrieval (SICOR) algorithm (Vidot et al 2012, Landgraf et al 2016). The ground pixel size of SWIR bands is generally consistent (7 km) in the along-track direction but varies from 7 km at nadir to 34 km at the scan edge in the across-track direction (Veefkind et al 2012). The 7-km TROPOMI Level 2 CO product provides for each pixel observation time, coordinates (longitude and latitude), satellite VZA, solar zenith angle, CO total column density (mol m⁻²), and retrieval precision. The vertical sensitivity of the CO total column density retrievals depends on the total column averaging kernels that are sensitive to clouds (Borsdorff et al 2018a). The averaging kernel of the CO total column density across the globe in clear-sky conditions has been demonstrated to be close to 1.0 at all altitudes, which suggests that the CO total column retrievals represent the true CO total columns very well (Borsdorff et al 2018a). The validation also shows that TROPOMI CO data agree with ground-based CO observations and model-simulated CO data well (Borsdorff et al 2018b, 2018c).

The TROPOMI sensor also senses aerosols using the ultraviolet bands (UV). The ground pixel size of TROPOMI UV bands is 7 km in the along-track direction and varies from 3.5 km at nadir to 15 km at the scan edge in the cross-track direction, which is ∼2 times finer than the SWIR bands (Veefkind et al 2012). The TROPOMI Level 2 aerosol index (AI) product calculates AI based on the 354/388 nm and 340/380 nm pair bands (Apituley et al 2018). A positive AI indicates the presence of absorbing aerosols (i.e. smoke from fires, and volcanic ash, and dust). Thus, this study obtained the TROPOMI Level 2 CO and AI products for the period from July 2018–October 2019 from the Copernicus Sentinel-5P pre-operations data hub (https://s5phub.copernicus.eu/; last accessed on 3 April 2020).

The true-color (RGB) images from Terra MODIS and VIIRS onboard S-NPP (the Suomi National Polar-Orbiting Partnership) are obtained to identify visually interpretable fresh smoke plumes. The daily global true-color composites based on MODIS and VIIRS surface reflectance are available on the NASA (National Aeronautics and Space Administration) Worldview tool (https://worldview.earthdata.nasa.gov/).

The FINN, GFAS, and QFED emissions inventories provide daily global fire emissions, including CO, based on the 1-km MODIS active fire data (Wiedinmyer et al 2011, Kaiser et al 2012, Darmenov and da Silva 2015). The FINN inventory estimates emissions at the fire pixel level using the conventional bottom-up method, with the parameters-burned area and fuel consumption calculated in each MODIS fire detection (Wiedinmyer et al 2011). The GFAS calculates emissions using MODIS FRP and the land cover specific FRE combustion factors that are derived by relating MODIS FRP to the dry matter combustion ratio from the Global Fire Emission Database (GFED3.1) (Kaiser et al 2012). The QFED estimates emissions using MODIS FRP and emission coefficients that are derived in two steps: first obtain initial factors by relating MODIS FRP to GFED emissions in four biomes, and then adjust the initial factors by comparing the model-simulated AOD with MODIS AOD (Darmenov and da Silva 2015). This study obtained the FINN v1.5 (http://bai.acom.ucar.edu/Data/fire/; last accessed on 18 May 2020), the GFAS v1.2 (https://apps.ecmwf.int/datasets/data/cams-gfas/; last accessed on 18 May 2020), and the QFED v2.5r1 (https://portal.nccs.nasa.gov/datashare/iesa/aerosol/emissions/QFED/v2.5r1/0.1/; last accessed on 18 May 2020) emissions data for the period from June to September 2018.

Fire events with fresh smoke plumes were selected based on active fire data from GOES-16 ABI and natural color imageries from both Terra MODIS and S-NPP VIIRS. Specifically, the ABI fire detections with valid FRP were first used to locate potential fire events. Then for each fire event the VIIRS natural color reflectance composite was employed to examine if the fire released any visually interpretable smoke plumes that contrast sharply with background and are not contaminated by clouds and smoke from other fire events. Meanwhile, Terra MODIS natural color reflectance was used to determine the age of smoke plumes. In this study, we define that a smoke plume is fresh if it is released within a few hours (i.e., 3 h) of TROPOMI overpass. Because Terra MODIS overpasses (10:30) the same area ∼3 h earlier than S-NPP VIIRS (13:30) and Sentinel-5P TROPOMI (13:35), a fire event is selected if the Terra MODIS reflectance composite shows no smoke plumes (or smoke with very limited size) associated with the fire but VIIRS reflectance shows very clear smoke plumes. Figures 1(a)–(f) shows the selection process of two fires in northern Colorado on 22 September 2018, where the associated two smoke plumes released after Terra MODIS overpass were observed by S-NPP VIIRS and further confirmed by ABI active fire detections and TROPOMI aerosol index data.

Zoom In Zoom Out Reset image size

The selected fire events were further refined using TROPOMI CO. During data exploration, we found that the TROPOMI Level 2 CO product could fail to provide valid CO retrievals for a few smoke pixels close to fire fronts over very large and intense fires. Thus, any selected fires with invalid CO observations were excluded. As a result, a total of 41 wildfires were finally extracted and they are mainly forest fires (figure 1(g)). For each selected fire, TROPOMI AI data were used to discriminate smoke plumes from the background, and CO observations inside the boundaries of a smoke plume were extracted as smoke CO (figures 1(e) and (f)). The smoke plume boundaries were then expanded outward by a distance of at least three CO observations (∼21 km = 7 km × 3 pixels), and the CO observations located at the region between smoke plume boundaries and the expanded boundaries were extracted as the background non-smoke CO (figure 1(d)). As the background pixels were not contaminated by clouds or smoke from other fires, the CO total column retrievals in the clear-sky observation conditions are able to well represent the true CO total columns (Borsdorff et al 2018a). Because the TROPOMI Level 2 CO data measure vertically integrated CO total column above land/water surface, TROPOMI CO over either smoke or background pixels also contains CO from other sources, including oxidation from methane and biogenic volatile organic compounds (VOCs), etc.

The total mass of fire smoke CO emissions was calculated using TROPOMI CO observations for each of the selected 41 fires. Because smoke plumes of the selected fires were just released within a few hours of the TROPOMI observation time, we assume that oxidation of the CO emissions during such a short period is negligible because atmospheric CO has a lifetime of about two months at middle latitudes (Khalil and Rasmussen 1990). Following the method used by (Heymann et al 2017), for each selected fire we first calculated background CO density by averaging the CO total column density of all the extracted non-smoke background CO pixels to account for non-fire contributions to the observed CO. Then the total mass of the fire-released CO emissions is calculated using TROPOMI CO observations as:

$$\begin{equation}{\boldsymbol{C}}{{\boldsymbol{O}}_{\boldsymbol{t}}} = \sum\nolimits_{{\boldsymbol{i = 1}}}^{\boldsymbol{n}} {\left[ {\left( {{\boldsymbol{\rho }}_{{\boldsymbol{sm}}}^{\boldsymbol{i}} - {{\boldsymbol{\rho }}_{{\boldsymbol{bg}}}}} \right) \times {{\boldsymbol{A}}^{\boldsymbol{i}}} \times {\boldsymbol{M}}} \right]} \end{equation} \tag{ 1 }$$

where CO_t is the fire-released CO total mass (g) observed by TROPOMI, n is the number of TROPOMI CO pixels inside the associated smoke plume boundaries for a given fire, ${\boldsymbol{\rho }}_{{\boldsymbol{sm}}}^{\boldsymbol{i}}$ is the observed CO total column density (mole m⁻²) of the ith TROPOMI smoke CO pixel, ${{\boldsymbol{\rho }}_{{\boldsymbol{bg}}}}$ is the mean CO total column density (mole m⁻²) of the non-smoke background pixels, ${{\boldsymbol{A}}^{\boldsymbol{i}}}$ is pixel area (m²) of the ith smoke CO pixel, and M is CO molecular mass (28.01 g mol⁻¹).

To examine the enhancement of CO over each fire smoke plume against the non-smoke background, we also calculated the dry-air CO total column mixing ratio (in units of parts per billion volume, abbreviated as ppbv) based on the TROPOMI CO total column density and the co-located surface pressure by following the method in Borsdorff et al (2018a). For the sake of simplicity, the CO column mixing ratio is referred to as the dry-air CO total column mixing ratio.

The total mass of the CO emissions was also calculated based on 5-min FRP retrievals from GOES-16 ABI using an FRE biomass combustion factor. (Wooster et al 2005) derived an FRE biomass combustion factor of 0.368 kg MJ⁻¹ in a field experiment that has been commonly used to estimate regional-to-global biomass consumption and emissions (Zhang et al 2012, Roberts et al 2015). (Li et al 2018) reported an FRE combustion factor of 0.320 kg MJ⁻¹ for the CONUS wildfires based on FRP from MODIS and GOES and the Landsat-based biomass consumption. For each selected fire, the total mass of the CO emissions was calculated using these two combustion factors separately:

$$\begin{equation}{\boldsymbol{C}}{{\boldsymbol{O}}_{\boldsymbol{f}}} = {\boldsymbol{FRE}} \times {\boldsymbol{B}} \times {\boldsymbol{F}}\end{equation} \tag{ 2 }$$

where $C{O_f}$ is FRP-based CO total mass (g), FRE is the ABI FRP-based fire radiative energy (MJ), B is the FRE biomass combustion factor, and F is the CO emission factor (89 g kg⁻¹ for temperate forests; Akagi et al 2011). By assuming that fire intensity is generally consistent within 5 min, the total radiative energy FRE is calculated using 5-min ABI FRP as:

$$\begin{equation}{\boldsymbol{FRE}} = \sum\nolimits_{{\boldsymbol{k = 1}}}^{\boldsymbol{q}} {\boldsymbol{F}} {\boldsymbol{R}}{{\boldsymbol{E}}_{\boldsymbol{k}}} = \sum\nolimits_{{\boldsymbol{k = 1}}}^{\boldsymbol{q}} {\int\nolimits_{{\boldsymbol{t1}}}^{{\boldsymbol{t2}}} {{\boldsymbol{FR}}{{\boldsymbol{P}}_{\boldsymbol{k}}}} } \left( {\boldsymbol{t}} \right){\boldsymbol{dt}}\end{equation} \tag{ 3 }$$

where, given a fire with a total number of q ABI fire detections, FRE_k is the fire radiative energy (MJ) observed at the kth ABI fire pixel during a burning period from t1 to t2. t1 is set as the observing time of the first valid GOES-16 ABI fire observation at the kth fire pixel and t2 is the TROPOMI overpass time. FRP_k(t) is ABI FRP at observing time t (t ∈ (t1, t2)).

Then, the ABI FRP-based CO total mass estimates were compared statistically with the observed TROPOMI CO total mass estimates over the selected 41 fires using the ordinary least squares (OLS) regression.

Finally, the CO emission coefficient that converts FRP to the mass of the CO emissions was also derived from ABI FRE and TROPOMI CO observations using the following formula:

$$\begin{equation}{\boldsymbol{C}}{{\boldsymbol{O}}_{\boldsymbol{t}}} = {\boldsymbol{FRE}} \times {\boldsymbol{Ce}}\end{equation} \tag{ 4 }$$

where CO_t is the fire-released CO total mass (g) observed by TROPOMI, FRE is the ABI FRP-based fire radiative energy (MJ), and Ce is CO emissions coefficient (g MJ⁻¹). The CO emission coefficient was statistically calculated from 41 fire events using the OLS regression.

The CO emissions across the CONUS were estimated using ABI FRP and the derived CO emissions coefficient (Ce) at a grid resolution of 0.25° for the four months from June to September 2018. The result was compared with the CO emissions from the FINN, GFAS, and QFED fire emissions inventories. The four months from June to September cover the typical fire season in the west CONUS where large wildfires are usually widespread (Zhang et al 2014b, Balch et al 2017). Specifically, for each ABI fire pixel, the daily FRE was first calculated by integrating the 5-min ABI FRP and then applied to calculate the daily CO emissions using the equation (4). The daily pixel-level CO emissions were further aggregated in a 0.25° grid and at an interval of one month. Finally, the ABI FRP-based monthly CO emissions were compared with the CO emissions from the FINN, GFAS, and QFED inventories.

Smoke plumes show significantly higher CO concentration than the non-smoke background (figure 2). Over the selected 41 fires, the CO column mixing ratio of smoke plumes on average was 47 ppbv higher than that of the background, which represents a CO mixing ratio enhancement of 51%. The largest enhancements of 164% (135 ppbv) and 261% (190 ppbv) were seen over two fires in Idaho State. The CO enhancements were in similar ranges to the ground- and airborne-measured CO enhancement over wildfires in the western CONUS and North America boreal regions (Val Martín et al 2006, Ditas et al 2018). The mean CO mixing ratio enhancement was 79% in the western CONUS while it was 27% in the eastern CONUS. The large difference is partly due to the difference in the background CO between the two regions. The mean background CO column mixing ratio in the eastern CONUS (113 ppbv) is approximately 138% of that in the western CONUS (82 ppbv), which is consistent with the pattern based on long-term MOPITT CO observations (Deeter et al 2012, Jiang et al 2018). This is primarily attributed to the much larger contributions from anthropogenic and biogenic CO sources in the eastern CONUS (Hudman et al 2008, Jiang et al 2018).

Zoom In Zoom Out Reset image size

The ABI FRP-based CO estimates agree well with the TROPOMI CO estimates (figure 3). Over the selected 41 fires, the total mass of the CO emissions estimated via the two FRE biomass combustion factors (equation (2)) was 0.001–6.78 Gg and 0.001–5.90 Gg, which has a similar magnitude to the TROPOMI CO based total mass: 0.015–6.21 Gg. The ABI FRP-based CO was significantly correlated to the observed TROPOMI CO (R² = 0.94, p-value < 0.001). When an FRE biomass combustion factor of 0.368 kg MJ⁻¹ was applied, the FRP-based CO was overall slightly larger than TROPOMI CO by up to 18% at a 95% confidence interval, with an RMSE (root mean square error) of 0.3. For some relatively small fires, the FRP-based CO was smaller than TROPOMI CO, as indicated by the small negative bias. Similarly, the FRP-based CO via an FRE combustion factor of 0.320 kg MJ⁻¹ was overall slightly smaller than TROPOMI CO by up to 13% at a 95% confidence interval, with an RMSE of 0.3. Although the ABI and TROPOMI sensors provide unprecedented spatiotemporal resolution observations of fires and smoke emissions, the accuracy of the associated FRP and CO data could influence CO estimates for individual fires. However, the very good agreement between the ABI FRP-based CO and TROPOMI CO estimates over the 41 fires across the CONUS confirms the robustness of the FRP-based method for BBE estimation via FRE biomass combustion factors. The small difference (−13% – +18%) between the FRP-based CO and TROPOMI CO suggests that for the CONUS forest wildfires the uncertainty of the FRP-based BBE via the two FRE combustion factors is limited within 18%. This differs from the previous finding that BBE estimated via FRE combustion factor are largely underestimated and an increase of BBE by a factor of several times is needed to reduce the underestimation (Kaiser et al 2012, Darmenov and da Silva 2015).

Zoom In Zoom Out Reset image size

Figure 4 shows a CO emission coefficient of 29.92 ± 2.39 g MJ⁻¹ based on ABI FRE and TROPOMI CO over the selected 41 fires. The coefficient of determination indicates that 94% of the variation in CO could be explained by ABI FRE, which also implies that the derived CO emission coefficient can be used to calculate the CO emissions from FRP without bias and with limited uncertainties.

Zoom In Zoom Out Reset image size

CO emission coefficient has been derived in several previous studies to directly estimate BBE using satellite-based FRP. In the published literature, CO emission coefficient has been derived directly by correlating lab-measured rate of the CO emissions to FRP (or mass of CO to FRE; Freeborn et al 2008) or indirectly inferring from the TPM (total particulate matters) emission coefficient based on the ratio of CO and TMP emission factors (Mota and Wooster 2018). Our CO emission coefficient is very close to the lab-based CO emission coefficient (33.71 g MJ⁻¹; Freeborn et al 2008). However, the CO emission coefficient converted from the AOD-derived TPM emission coefficient is much larger. (Mota and Wooster 2018) reported a CO emission coefficient of 142 g MJ⁻¹ based on MODIS AOD and FRP from SEVIRI (Spinning Enhanced Visible and Infra-red Imager) in open-canopy forest fires in Africa. Similarly, (Ichoku and Ellison 2014) and (Lu et al 2019) separately showed a TPM emissions coefficient of 40 g MJ⁻¹ and 21.73 g MJ⁻¹ for the CONUS forest fires based on MODIS FRP and AOD data, which corresponds to a CO emission coefficient of 193 g MJ⁻¹ and 105 g MJ⁻¹. One potential issue in deriving AOD-based CO emission coefficients is the uncertainties of the smoke mass extinction coefficient that converts AOD to TPM. The mass extinction coefficient varies between 2.22 and 7.6 m² g⁻¹ as smoke ages, which has been discussed in (Mota and Wooster 2018). Moreover, previous studies also showed that the BBE estimated via the AOD-based coefficients is larger than BBE estimates calculated via FRE biomass combustion factor by a factor of several times (Zhang et al 2014a, Li et al 2019, Carter et al 2020). As a result, it is reasonable to deduce that combining ABI FRE and TROPOMI CO could provide a much reliable CO emission coefficient.

Figure 5 shows the spatial patterns of the ABI FRP-based monthly CO emissions across the CONUS from June to September 2018. During this period, fires occurred in most regions of the CONUS but the fire-released CO emissions varied significantly spatially. Fires in the western CONUS accounted for 82%, 92%, 95%, and 93% of the monthly total CO emissions across the CONUS in the four months, respectively. In the same period, however, fires in the eastern CONUS released a relatively very limited amount of the CO emissions, which were mainly contributed by fires across the southeastern CONUS. This is consistent with the previous finding that the majority portion of the CONUS emissions is from fires across the western CONUS during the summer months (Urbanski et al 2018, Li et al 2019).

Zoom In Zoom Out Reset image size

The ABI FRP-based CO emissions provide a good reference for evaluating the CO emissions from the existing fire emissions inventories. Figure 6 shows the monthly total CO emissions based on ABI FRP and from the FINN, GFAS, and QFED emissions inventories across the CONUS. In the four months from June to September, the ABI FRP-based monthly CO emissions were 294 (±23) Gg, 702 (±56) Gg, 934 (±75) Gg, and 455 (±36) Gg, respectively. In comparison with the three inventories, the ABI FRP-based monthly CO emissions on average were very similar to the FINN and GFAS CO, with a mean difference of less than 13%, but differed from the QFED CO by a factor of approximately four. Although the mean of the ABI FRP-based monthly CO emissions (596 Gg) was comparable with that of the GFAS CO (670 Gg), the ABI FRP-based CO and the FINN and QFED CO peaked in August while the GFAS CO peaked in September and was the smallest in August among all the four CO datasets. Further, previously published studies also found that the QFED emissions were much larger than the other emissions inventories across the CONUS and other regions, with a difference of 2–6 times (Li et al 2019, Carter et al 2020, Pan et al 2020). The large discrepancies are most likely associated with the fact that the QFED emissions are significantly enlarged by the MODIS AOD based scaling factors (Carter et al 2020).

Zoom In Zoom Out Reset image size

Potential uncertainties in the ABI FRP-based CO emissions estimates could arise from the omission errors in the ABI fire detection and the application of the single CO emission coefficient. First, the GOES-16 ABI sensor could miss very small and/or cool fires at large view angles in the northwestern CONUS because of the decrease in pixel resolution although its high temporal resolution significantly improves the fire detection capability (Li et al 2020). Luckily, the portion of small fires in the northwestern CONUS is much smaller than that in the southeastern CONUS (Randerson et al 2012) where the ABI sensor has a fire detection capability comparable to the VIIRS sensor (Li et al 2020). Thus, the uncertainty of the ABI FRP-based CO due to the omission errors of the ABI fire detections could be relatively small. Second, the CO emissions across the CONUS in this comparison were calculated using the single CO emission coefficient that was derived based on the TROPOMI CO over mostly forest fires (figure 4). For other types of fires (e.g. agricultural burnings), the CO emission coefficient could be different, which could introduce certain uncertainties in the CO emissions estimates. However, this effect could be limited in this comparison because fires, especially large forest wildfires, in the western CONUS contributed to the majority portion (82% in June and >90% in the following three months) of the CO emissions across the CONUS.