The spatial domain of this study includes all of Alaska and Canada for our burned-area product and the ABoVE core and extended domain (hereafter the “ABoVE domain”; Loboda et al., 2019) for our combustion and burn depth product (Fig. 1). The combustion and burn depth products were not derived beyond the ABoVE domain due to a lack of field observations in eastern Canada. The temporal domain for all products is 2001–2019. Our study area includes all natural boreal and arctic vegetation within the ABoVE domain, including boreal forests, boreal wetlands, grasslands, tundra, and tundra wetlands. To determine these locations we derived a vegetation mask using the 2005 Land Cover of North America product (250 m; CCRS, 2013; Pouliot and Latifovic, 2013; Pouliot et al., 2014), MODIS land cover type with International Geosphere–Biosphere Programme (IGBP) classification (Collection 6, year 2005, 500 m; Friedl and Sulla-Menashe, 2019), the Circumpolar Arctic Vegetation Map (CAVM; Raynolds et al., 2019), and long-term climate (1970–2000, ∼ 1 km; Fick and Hijmans, 2017), all re-gridded to 500 m resolution on the MODIS sinusoidal projection (Fig. S1 in the Supplement). Boreal vegetation was distinguished from temperate using a mean annual temperature threshold of 3 ∘C, as recommended in Wolfe (1979) and implemented in Rogers et al. (2015). Pixels were designated as urban, crop, crop/natural vegetation mosaic, or water if they were represented as such in either the Land Cover of North America or MODIS land cover products. Pixels were designated as tundra if they were within the CAVM domain.

Field measurements of burn depth and combustion were derived from numerous data sources across different research groups that represent a major synthesis effort sponsored by the NASA ABoVE program (Boby et al., 2010; Dieleman et al., 2020; de Groot et al., 2009; Hoy et al., 2016; Rogers et al., 2014; Turetsky et al., 2011; Veraverbeke et al., 2015; Walker et al., 2018). Detailed descriptions of data collection methods can be found in the contributing publications. All field site information was standardized and aggregated into a single publicly available database (Walker et al., 2020a), which has been used to assess patterns and drivers of ecosystem structure and combustion across ecoregions (Walker et al., 2020b, c). Although the field database only includes measurements from boreal ecosystems, our combustion and burn depth predictions include both boreal and tundra ecosystems. Of all the pixels for which we predicted combustion and burn depth, only 0.78 % are in tundra landscapes.

The ABoVE Fire Emissions Database (ABoVE-FED) burned-area product is derived from a dNBR thresholding approach, which has previously been successfully employed for burned-area mapping in the region (Rogers et al., 2014; Veraverbeke et al., 2015; Walker et al., 2018). Our primary approach was to use Landsat imagery to separate burned from unburned pixels at 30 m. However, because Landsat imagery was not available for all regions and time periods, we used MODIS imagery to map burned pixels when necessary and upscaled our Landsat-based product to 500 m MODIS resolution. More specifically we used pre- and post-fire near-infrared (NIR) and shortwave infrared (SWIR) bands from Aqua (MYD09GA Collection 6; Vermote and Wolfe, 2015a), Terra (MOD09GA Collection 6; Vermote and Wolfe, 2015b), and Landsat 5–8, calculating dNBR as the difference in pre-fire normalized burn ratio (NBR) and post-fire NBR, where NBR is near-infrared minus shortwave infrared divided by near-infrared plus shortwave infrared.

This approach had the added advantage of accuracy; whereas a Landsat dNBR threshold tends to be surpassed at the site level in a diffuse manner across the landscape, due to stochastic site-level disturbances such as tree mortality, herbivory, flooding, or small-scale dieback, it is much less common for these small-scale disturbances to influence the majority of a 500 m pixel. We also minimized mapping non-fire disturbances by following the approach of Veraverbeke et al. (2015) and applying our dNBR approach to (1) mapped fire polygons from the ALFD and CNFD (93 % of total burned pixels; hereafter collectively referred to as the National Large Fire Databases, NLFD) and (2) MODIS active-fire acquisitions (MOD14A1 Collection 6 and MYD14A1 Collection 6; Giglio et al., 2018) outside these polygons (7 % of total burned pixels). In each case we applied a 1 km buffer (Veraverbeke et al., 2015) to capture burned pixels immediately outside these areas. Finally, our approach is motivated by a desire to balance commission and omission errors at both the 30 and 500 m scales, thereby providing an unbiased estimate of total burned area.

To map 30 m burned pixels, we first extracted dNBR at both burned and unburned control sites in our aggregated field database using available cloud-free Landsat 5, 7, and 8 Tier 1 surface reflectance images in Google Earth Engine (Gorelick et al., 2017). Landsat 5 and 7 were atmospherically corrected using the Landsat Ecosystem Disturbance Adaptive Processing System (LEDPAS; Schmidt et al., 2013), while Landsat 8 was atmospherically corrected using Land Surface Reflectance Code (LaSRC; Vermote et al., 2016). Pre- and post-fire normalized burn ratio (NBR) was calculated as the mean of all available Landsat observations between July and August. Pre-fire values were extracted 1 year before a given fire, and post-fire values were extracted 1 year after a fire. We then selected a 30 m Landsat dNBR threshold that most effectively separated burned and unburned control sites. Because there are many fewer unburned control sites in the Walker et al. (2020a) combustion database, we derived additional control sites by extracting dNBR at burned sites 2 years before a given fire, which had the advantage of controlling for any site-level spectral differences between burned and control sites represented in the database. This process generated a dNBR threshold of 0.084, which minimized 30 m site-level commission and omission errors to 6.6 % (Fig. S2).

We then created a mask at 30 m to account for unburnable land cover (i.e., non-vegetated pixels). This was created using two sources: the Joint Research Center's yearly water history product (Pekel et al., 2016) and the 2010 land cover product of the North American Land Change Monitoring System (NALCM) at 30 m resolution (Latifovic et al., 2012). The first product allowed us to capture transient water pixels in our time series, while the NALCM land cover product classified each pixel into 19 different land cover classes, from which we masked out non-vegetated pixels, including ice, water, barren land, and cropland. These two sources were combined into separate masks for each year between 2001–2019. Because areas that burned in 2010 were often classified as barren lands in the 2010 NALCM product, we considered barren lands to be vegetated in our mask for the year 2010.

Using the vegetation mask and the dNBR threshold, we created a binary burned/unburned 30 m Landsat product and upscaled this to the native MODIS 500 m resolution and projection. To determine whether or not a given 500 m pixel was classified as burned or unburned, we calculated the percentage of 30 m vegetated pixels that burned within its footprint. If more than 50 % of the 30 m vegetated pixels within the larger 500 m pixel burned (i.e., were tripped by the dNBR threshold), the entire pixel was assigned as burned, and the burned fraction was calculated as the percent of the burnable land cover (vegetation) in the 500 m pixel. Note we did not use the percent of burned 30 m pixels to determine burn fraction within a given 500 m pixel, primarily because of limitations imposed by frequently missing Landsat imagery (detailed below).

We used this approach whenever 500 m pixels contained 100 % coverage by Landsat imagery at 30 m. When, however, there was less than 100 % Landsat coverage, we needed to determine if it was more accurate to classify 500 m pixels using Landsat (with partial coverage) or MODIS Collection 6 imagery (Vermote and Wolfe, 2015a, b). To do so, we analyzed all MODIS pixels with complete Landsat coverage and masked out increasing numbers of Landsat pixel strips within the larger MODIS footprint (using increments of 5 %). After each removal of Landsat pixels, we compared the accuracy of the resulting burned/unburned classification using (i) Landsat imagery with partial coverage and (ii) MODIS imagery. This procedure suggested that using MODIS dNBR was more accurate than Landsat when less than 85 % of a 500 m MODIS pixel was covered by Landsat imagery. We therefore used Landsat to classify burned pixels when at least 85 % of a 500 m pixel was covered by Landsat imagery and otherwise used MODIS. Burned pixels were assigned a quality flag of 0 when there was complete Landsat coverage; a quality flag of 1 when Landsat coverage was less than 100 % but greater than 85 %; and a quality flag of 2 when Landsat coverage was less than 85 %, and therefore MODIS imagery was used to classify burn status. Overall, 81 % of total burned pixels were derived using Landsat (66 % from full coverage and 15 % from partial coverage), although particular regions (notably Alaska and Newfoundland and Labrador) tend to rely more on MODIS due to more limited availability of Landsat imagery (Fig. S3).

We developed a correction factor for MODIS-based dNBR to account for differences between Landsat and MODIS NIR and SWIR spectra, as well as the influence of vegetation fraction on 500 m dNBR signals. To do so, we calculated pre- and post-fire NIR and SWIR bands from MODIS and Landsat (resampled to 500 m) for a 50 % random sample of burned pixels. We then differenced the Landsat 500 m resampled bands from the 500 m MODIS bands and regressed them onto vegetation fraction to obtain a correction factor. The regression yielded an R2 of 0.74 and an equation of y=0.94x+0.01, which was applied to all pixels where burn status was classified by MODIS. We then calculated a new dNBR threshold to classify pixels at 500 m in an unbiased manner. To do so, we determined the MODIS dNBR threshold that evenly split omissions and commissions based on pixels mapped with complete Landsat coverage. This threshold was determined to be 0.0725, resulting in an omission/commission error of 14.2 % at 500 m when using MODIS.

One issue with a burned-area mapping approach such as ours that utilizes post-fire imagery 1 year after a fire is that it is difficult to determine the year(s) of burn where overlapping burns occurred in successive years. To address these cases, we created a seasonal MODIS-based product following the methodology of Giglio et al. (2018). The dNBR for each day between 15 January and 15 December was calculated using the 30 preceding days as pre-fire NBR and the 30 d after as post-fire NBR. Any pixels with fewer than 10 valid observations in either window were masked out. We used a similar thresholding approach to that described above for mapping burned pixels with MODIS, resulting in a seasonal dNBR threshold of 0.23. Any pixel mapped using the MODIS seasonal approach was assigned a quality flag of 3.

In addition to determining fire locations, fire year, and the burned fraction, we also determined the day of burning for each pixel. When possible, day of burn was taken directly from the thermal-anomaly active-fire detections from MOD14A1 Collection 6 and MYD14A1 Collection 6 (Giglio et al., 2018) active-fire products. Where an active fire was registered, day of burn was assigned by taking the earliest active-fire acquisition during the year. When an active fire was not registered for a given burned pixel, we utilized a multi-tiered approach to assign day of burn. When possible, we used a kriging technique to interpolate day of burn using the active-fire detections within each fire polygon in the NLFD following Veraverbeke et al. (2015). To implement this, we required fire polygons to contain at least five active-fire acquisitions within their boundaries and have some level of temporal variation (i.e., not all active-fire acquisitions on the same day). When this was not the case, day of burn was assigned using the closest active-fire pixel. Finally, when no active-fire acquisitions were associated with a given fire polygon, we used our MODIS-based seasonal mapping approach to determine day of burn by locating the day of maximal dNBR within a given year. For fires that were detected by MODIS thermal anomalies but were not contained in the NLFD (7 % of all burned area), we created our own polygons around the burned pixels (by converting pixels to vectors and buffering them) and used the same method to assign day of burn. Quality flags for our burn day product represent this tiered approach, with a flag of 0 for pixels with direct active-fire hits, a flag of 1 for pixels whose day of burn was determined by interpolation, and a flag of 2 for pixels whose day of burn was determined using the MODIS seasonal burned-area product. A simplified flowchart of burned-area processing methods is shown in Fig. S4.

We compared ABoVE-FED burned area to several other products including the NLFD, NBAC, MCD64A1 Collection 5, MCD64A1 Collection 6, the Alaska Fire Emissions Database version 2 (AKFED; Veraverbeke et al., 2017), GFED4s (van der Werf et al., 2017), a 500 m model by van Wees et al. (2022), and the Fire Model Intercomparison Project (FireMIP; Hantson et al., 2016; Rabin et al., 2017; Table S1 in the Supplement). NBAC is a Canada-only product and is related to the CNFD but improves upon it by incorporating multi-sensor remote sensing imagery (including Landsat) to account for waterbodies and unburned vegetation patches. FireMIP includes simulations performed by coupled fire–vegetation models forced with a standardized set of input data. We also visually compared our product and others to high-resolution imagery of fires from the WorldView-2 (1.84 m) satellite, available through DigitalGlobe, Inc., a Maxar company under the NextView license agreement through the National Geospatial Intelligence Agency (Neigh et al., 2013).

We built and applied statistical models of aboveground combustion, belowground combustion, and burn depth to every mapped burned pixel in the ABoVE domain based on field observations across Alaska and western Canada (Walker et al., 2020a). Because not all field sites included estimates of both aboveground and belowground combustion, we created two separate combustion models, one utilizing all available aboveground combustion measurements (n=515) and one utilizing all available belowground combustion measurements (n=769). Our burn depth model utilized the same field sites as belowground combustion. Further discussion of models implemented can be found in the Supplement.

Turetsky et al. (2011) discovered a positive relationship between burn depth, annual burned area, and timing of burn (day of year) in black spruce forests and peatlands of interior Alaska and also noted the influence of burn timing was more important in small-fire years. To test if these relationships held true with a larger field database in Alaska (n=286 for ABoVE-FED compared to n=178 in Turetsky et al., 2011), we performed a multiple regression of burn depth and belowground combustion using annual burned area and day of year as predictor variables. We also tested how burn depth and belowground combustion varied as a function of day of year within both small- and large-fire years. To do so, we split the field sites in Alaska into four quantiles based on annual burned area and then regressed burn depth and belowground combustion against day of year within each quantile. We also conducted this analysis using a sample of 500 ABoVE-FED pixels in Alaska instead of field observations and then repeated both of these analyses using all available field observations and 500 random pixels within the broader ABoVE domain. In each case, we sampled 500 pixels instead of using all available pixels to minimize the effect of large sample sizes on p values.

Temporally there was high variability in burned area year to year (Fig. 2a). Across the domain, ABoVE-FED reported similar burned-area totals compared to the NLFD (average of 2.87 Mha yr-1 for ABoVE-FED compared to 2.90 Mha yr-1 for NLFD; Fig. 3), although there was variability in this relationship (NLFD estimated larger annual burned area in 11 years and smaller burned area in 8 years between 2001–2019). This was the net result of two contrasting patterns: ABoVE-FED tended to report less burned area within mapped polygons, due to unmapped unburned patches and unburnable land cover (e.g., small waterbodies) in the government fire databases, but detected additional burned areas associated with MODIS active-fire acquisitions well outside mapped fire polygons (7 % of total burned area in ABoVE-FED, 6 % of total emissions; Fig. S9). The state/territory with the most burned area detected outside the mapped polygons was British Columbia (31 % of the 7 % total burned area mapped outside NLFD polygons; Fig. S9). Exploratory analysis revealed this was likely a result of commission errors due to logging (i.e., logged areas tripping dNBR thresholds in conjunction with small fires registered by MODIS active-fire hits). Across the domain, the mean fire size coincident with NLFD polygons was much larger (4954 ha) than the mean fire size outside the polygons (166 ha). Because the NBAC product accounts for more of these unburned patches within polygons (Hall et al., 2020), it tended to report lower total burned area compared to ABoVE-FED (Fig. S10). ABoVE-FED burned area was higher than MCD64A1 (Collection 5 and 6; Fig. 3) in all years, which is consistent with known omissions in these global products for boreal North America (Giglio et al., 2018; Randerson et al., 2012; Fig. S11). These large-scale patterns were corroborated by high-resolution imagery of particular fire events (Figs. 4, S12–S19). ABoVE-FED identified more burned pixels than MCD64A1 Collection 6 by being more sensitive to fire-induced spectral changes but also accounted for unburnable portions of the landscape (Fig. S20). GFED4s burned area was slightly higher (Fig. S21; average of 2.38 Mha yr-1 during 2001–2016) than both MODIS products (Fig. 3; average of 2.93 Mha yr-1 during 2001–2016), but lower than the NLFD and ABoVE-FED (Fig. 3). The MCD64A1 Collection 5 and Collection 6 and GFED4s databases underestimated burned area by 32 %, 23 %, and 18 % compared to ABoVE-FED, respectively.

ABoVE-FED burned area was similar to AKFED where it was available (Alaska, the Northwest Territories, and the Yukon Territory; Fig. S22; average of 1.27 Mha yr-1 for ABoVE-FED during 2001–2015 compared to 1.22 Mha yr-1 for AKFED). All models participating in FireMIP simulated smaller burned area than ABoVE-FED, and with a very high level of variability between models (1.34 ± 0.83 Mha yr-1 across Alaska and Canada during 2001–2012; Fig. S23a).

Burned area was highly variable interannually (Figs. S21, S24), with the largest-fire years occurring in 2004 in Alaska and the Yukon Territory; 2015 in Alaska, Saskatchewan, and Alberta; 2014 in the Northwest Territories; and 2013 in Manitoba and Quebec. Across states, provinces, and territories, total burned area was highest in Alaska, the Northwest Territories, and Saskatchewan. A total of 54 Mha burned across Alaska and Canada during all years, and 45 Mha burned in the ABoVE domain, with an annual mean of 2.87 Mha yr-1 across Alaska and Canada and 2.37 Mha yr-1 in the ABoVE domain.

Spatially ABoVE-FED estimated the most burned area in Alaska, the Northwest Territories, and Saskatchewan (Figs. 5a, S21, S24).

Our aboveground and belowground combustion models performed well, although the aboveground model performed significantly better across the suite of models examined (Fig. 6a, b). A ranger random forest model (Wright and Ziegler, 2017) performed best for aboveground and belowground combustion, with a median R2 of 0.46 and 0.25, respectively, across the 10-fold cross-validation repeated 100 times. Our secondary models that did not include information on FWIs (Sect. 2.4.3) performed similarly to our primary models, with R2 values for aboveground and belowground combustion of 0.45 and 0.24, respectively. Although both the aboveground and belowground models performed reasonably well at predicting lower and moderate combustion values, which includes the majority of field observations, they both struggled to predict larger combustion values (Fig. S25a, b). The burn depth model performed better than both combustion models, with a median R2 of 0.53 using a ranger random forest model (Fig. 6c).

There were notable differences in the feature importance of the aboveground and belowground models (Fig. S26a, b). The aboveground model was heavily influenced by its top predictor, pre-fire tree cover, followed by metrics of relative humidity, with other variables including remotely sensed fire severity and vegetation moisture content having significant but relatively low importance. In contrast, the belowground model was influenced strongly by a number of soil, terrain, climate, and tree cover variables. The most important features for the burn depth model were similar to the belowground model, with soil properties, tree cover, and climate being the most influential (Fig. S26c). Overall, the distribution of variables used in the training dataset and predicting dataset were similar (Fig. S27), with the exception of slope. Most field sites were located in relatively flat terrain, whereas the combustion predictions included locations with steeper terrain.

Spatial patterns of mean burn depth and combustion tended to follow a gradient of higher burn depth and mean combustion in the western part of the ABoVE domain (Alaska, Yukon Territory, and Alberta) to lower mean combustion in central–western Canada (Saskatchewan, Northwest Territories, and Manitoba) (Figs. 5c, d, S28). There was, however, considerable fine-scale variability at 500 m within these regions (Fig. 7), and spatial patterns were relatively consistent with previous combustion mapping efforts.

Across the ABoVE domain, 1.51 ± 0.53 Pg C was emitted over the 2001–2019 period, with a mean of 79.3 ± 27.96 Tg C yr-1. Mean combustion across all years and regions was 3.13 ± 1.17 kg C m-2. Pixel-level uncertainty (Fig. S29) tended to follow spatial patterns of mean combustion (Fig. 5c) and was relatively consistent across years (Fig. S30), with a mean value of 2.86 kg C m-2. Seasonally, the majority of burned area occurred during June, July, and August (Fig. 8), although there were substantial regional differences, with some regions recording a large fraction of burned area outside this window (e.g., May fires in Alberta). In general, monthly patterns in emissions (Fig. S31) followed patterns in burned area. Overall, combustion tended to be highest in summer compared to spring and fall fires, although this pattern was most pronounced in the Yukon Territory, Northwest Territories, Saskatchewan, and Alaska (Fig. S32).

Estimates of total carbon emissions in ABoVE-FED were similar to AKFED (Fig. S22, Table S4), with the notable exception of 2014 in the Northwest Territories: AKFED estimated 164 Tg C, and ABoVE-FED estimated 89.7 Tg C. This was primarily a result of differences in mean modeled combustion in the Northwest Territories 2014 fires, with AKFED exhibiting its highest mean combustion in 2014 (Fig. S21; 4.81 kg C m-2 in AKFED compared to 2.89 kg C m-2 in ABoVE-FED). In general, ABoVE-FED estimated slightly higher mean combustion levels than AKFED in Alaska (3.34 kg C m-2 in ABoVE-FED and 3.03 kg C m-2 in AKFED), lower combustion in the Northwest Territories (3.29 kg C m-2 in ABoVE-FED and 3.44 kg C -2 in AKFED), and substantially higher combustion in the Yukon Territory (3.71 kg C m-2 in ABoVE-FED and 2.26 kg C m-2 in AKFED) (Fig. S22, Table S4). ABoVE-FED carbon emissions were relatively similar to Walker et al. (2018) for the 2014 Northwest Territories fires and to Dieleman et al. (2020) for the 2015 Saskatchewan fires (Fig. 7, Table S4). Total carbon emissions from ABoVE-FED were substantially higher than GFED4s (Fig. S33), with the largest differences occurring in Alaska. This was primarily a function of higher mean combustion values in ABoVE-FED compared to GFED4s (Fig. S34). Between 2001–2016, ABoVE-FED estimated 80 Tg C yr-1 total emissions with a mean combustion value of 3.39 kg C m-2, and GFED4s estimated 51 Tg C yr-1 total emissions with a mean combustion value of 2.30 kg C m-2 (Table S4). However, more recently a 500 m model by van Wees et al. (2022) was completed, and both emissions and combustion match more closely to ABoVE-FED. Between 2002–2019 this 500 m product estimates 73 Tg C yr-1 total emissions (Fig. S33) with a mean combustion value of 3.38 kg C m-2 (Fig. S34). Meanwhile, between 2002–2019 ABoVE-FED estimates 83 Tg C yr-1 total emissions with a mean combustion value of 3.16 kg C m-2. Compared to GFED4s these larger emission and combustion estimates in the 500 m product are largely due to increased estimates of belowground combustion, as the van Wees et al. (2022) model is informed by the same field measurements used in ABoVE-FED.

Differences in combustion and carbon emissions were very large between ABoVE-FED and fire–vegetation models participating in FireMIP (Fig. S23b). ABoVE-FED estimated much higher emissions than FireMIP (70.1 Tg C yr-1 for ABoVE-FED during 2001–2012 compared to 4.0 Tg C yr-1 for FireMIP). This is likely because models in FireMIP mostly combust aboveground vegetation, whereas combustion from belowground sources (primarily soil organic matter) comprises 90 % of total carbon emissions in ABoVE-FED (Fig. S35) and 88 % in the field plots from Walker et al. (2020a). ABoVE-FED mean aboveground combustion (7.84 Tg C yr-1 during 2001–2012) was much more similar to FireMIP's 4.0 Tg C yr-1.

We found multiple lines of evidence that belowground fire severity (burn depth and belowground combustion) is positively related to annual burned area and seasonal day of burn (Tables S5, S6). In general, mean annual burned area had a stronger relationship with fire severity than did burn day of year using multiple linear regression. However, within quantiles of annual burned area (i.e., small- vs. large-fire years), day of year was strongly related to fire severity (particularly belowground combustion), and the slope of this relationship was generally larger in small-fire years (Table S6). When assessed using domain-wide mean severity from mapped ABoVE-FED pixels, we found no significant relationship of burn depth with burned area or combustion (Fig. S36).

There were also no significant (p value ≤ 0.10) trends in burned area, combustion, or emissions across the 2001–2019 time series (Figs. 2a, b, S37).

Our approach to mapping burned area across boreal North America has several advantages compared to past approaches. Although our burned-area product is at 500 m resolution, the majority of pixels (81 %) were mapped using 30 m Landsat imagery. Using finer-scale 30 m imagery allowed us to directly calibrate dNBR thresholds to site-level information and account for unburnable fractions of 500 m pixels. We also calibrated these dNBR thresholds for both 30 m Landsat and 500 m MODIS imagery to most effectively balance omissions and commissions. This allowed us to provide an unbiased estimate of burned area, which is a critical variable for understanding the impacts of fire on arctic–boreal ecosystems and climate.

In theory, ABoVE-FED burned area would be expected to be higher than other available products because of its increased sensitivity to fire-induced spectral changes (compared to, for example, global MODIS burned-area products, via our focus on splitting omissions and commissions) and our accounting for active-fire acquisitions outside mapped fire polygons by the Alaskan and Canadian government agencies. Alternatively, ABoVE-FED accounts for sub-pixel heterogeneity of burnable land surfaces, which would otherwise result in lower burned-area estimates compared to existing products. The net result is that ABoVE-FED burned area tends to be higher than other products, but not exclusively.

We suggest future research efforts focused on burned-area mapping in arctic–boreal environments could be conducted at resolutions finer than 500 m. Doing so will allow for improved understanding of fire spread and behavior patterns and interactions between fire behavior and vegetation/land cover type. Finer-scale mapping should also allow for more accurate assessments of burned area by accounting for the presence of unburned patches of vegetation and waterbodies, thereby facilitating increased understanding of the drivers of fire spread and effects on ecosystem processes (Hall et al., 2020). Fires have typically been mapped at landscape scales using 500 m MODIS imagery because of the frequent revisit times (multiple acquisitions per day). With a resolution of 30 m, Landsat imagery has been less commonly used for mapping burned area at landscape scales because the revisit time (16 d) is much longer and because data coverage can be highly variable regionally and spatially depending on available downlink stations and cloud cover (Hilker et al., 2009; Ju and Masek, 2016; Fig. S3), but this revisit frequency is improving with two Landsat satellites (Landsat 8 and 9) and two Sentinel satellites (2a and 2b) in orbit, which provide much more frequent overpasses (2–3 d when combined).

Similar to ABoVE-FED, approaches for mapping burned area using satellite imagery have typically relied on image differencing of vegetation indices, particularly dNBR (French et al., 2015). This requires pre- and post-fire image pairs and thus compounds issues related to image availability at fine scales (30 m; Chen et al., 2021). Future burned-area mapping at landscape scales could potentially be improved by using machine learning. More specifically, deep learning approaches have been shown to be highly effective at mapping wildfires across different landscapes and vegetation types (Jain et al., 2020; Knopp et al., 2020). Convolutional neural networks, which use a spatial moving window and therefore account for the spatial characteristics of fire scars (Jain et al., 2020), are particularly promising. Finally, developing burned-area products in near real time, as opposed to active-fire-based assessments of hot pixel counts, would help scientists, fire managers, and society contextualize and potentially mitigate rapidly progressing fire seasons as they evolve.

Similar to previous studies (e.g., Veraverbeke et al., 2015), our aboveground combustion model performed substantially better than our belowground model. This is due primarily to the challenge of estimating belowground carbon consumption using remote-sensing-based observations, which are more sensitive to aboveground properties. For example, the ABoVE-FED aboveground combustion model was heavily influenced by remotely sensed properties such as pre-fire tree cover, fire severity (represented by dNBR), and vegetation wetness (represented by NDII), whereas the belowground model was strongly influenced by soil metrics, topography, and solar radiation (Fig. S26). This occurred despite our model utilizing considerably more field observations (n=515 for aboveground combustion and 769 for belowground) than past efforts in boreal North America (e.g., Dieleman et al., 2020: n=47; Veraverbeke et al., 2015: n=126; Walker et al., 2018: n=211), suggesting an inherently limited capacity to model belowground combustion using these techniques. Previous analysis of the field observations we used showed site-level drainage is the dominant driver of combustion in the ABoVE domain, due in part to the large contribution towards total combustion from belowground carbon stocks (Walker et al., 2018, 2020b). We therefore suggest prioritizing the use of geospatial products that adequately capture drainage, and thereby its impact on belowground carbon stocks and vulnerability to combustion, for improving future estimates of carbon emissions from fire disturbance across boreal North America.

Despite these limitations, our model performance is similar to past efforts. For example, Veraverbeke et al. (2015) reported an aboveground combustion model fit of R2=0.53 and a belowground fit of R2=0.29 for Alaska. Walker et al. (2018) implemented a 10-fold cross-validation approach and reported a model fit of R2=0.26 for total (aboveground and belowground) combustion in the Northwest Territories, Canada. Comparatively, we report a median R2 of 0.46 and 0.25 for ABoVE-FED aboveground and belowground combustion models, respectively. However, model performance was substantially higher in Dieleman et al. (2020), who reported a cross-validated R2 of 0.73 for total combustion in Saskatchewan. This is likely due to the higher relative contribution from aboveground combustion in the younger and more productive boreal forests of southern Canada, combined with high-quality provincial spatial datasets such as logging history (Dieleman et al., 2020). In all these cases, spatial patterns from ABoVE-FED are generally consistent with previous efforts (Fig. 7), lending confidence to assessments of drivers and spatiotemporal patterns of combustion.

Somewhat surprisingly, our models of burn depth performed better than both aboveground and belowground combustion models (cross-validated R2=0.53), which is considerably better than the R2 model fit of 0.40 reported for the burn depth model in Veraverbeke et al. (2015). This suggests substantial uncertainty in translating burn depth to carbon emissions in these boreal forests, which underscores the need for improved spatial layers of soil properties such as bulk density (Houle et al., 2017) and carbon fraction. The field and laboratory techniques used to calculate carbon emissions from burn depth also contain uncertainty, which is not always quantified. These errors are likely compounded when aggregating data across field campaigns, ecozones, and research groups, such as we did here. Nevertheless, burn depth is a critical fire severity property in its own right, with applications ranging from understanding the changing boreal carbon cycle (Walker et al., 2019) to post-fire succession and vegetation patterns (Baltzer et al., 2021; Johnstone et al., 2010). Our results suggest geospatial statistical modeling is well suited for capturing and extrapolating depth of burn in organic soils, at least within the ABoVE domain.

Finally, we assessed the influence of spatial and sensor differences when building the combustion and burn depth models at 30 m but predicting them at 500 m. Overall, biases introduced by model nonlinearities, sub-grid heterogeneity, and vegetation fractions were found to be negligible (slope = 0.98 for aboveground and 0.97 for belowground combustion when regressing 500 m against aggregated 30 m predictions). This suggests that approaches to map fire carbon emissions at large scales using 500 m MODIS imagery are not fundamentally biased because of spatial scale.

The machine learning models we employed allow insights into the drivers of both aboveground and belowground combustion. Partial dependence plots indicated that aboveground combustion tended to increase when tree cover and dNBR increased and when relative humidity and vegetation water content (NDII) decreased (Fig. S38). These patterns are consistent with understanding of fire behavior and aboveground consumption dynamics, which are generally driven by aboveground fuels and climate conditions that facilitate fuel drying and fire spread (Beck et al., 2011; Rogers et al., 2014; Walker et al., 2020b). Alternatively, belowground combustion increased with higher silt (and lower sand) content, higher tree cover, and lower relative humidity (Fig. S39). At moderate slopes (less than 20 %), at which the majority of field observations were located, belowground combustion was higher in flatter landscapes. These relationships are consistent with current understanding about the drivers of soil organic matter accumulation and vulnerability to combustion (Walker et al., 2018, 2020b; Scholten et al., 2021). Drivers of burn depth were similar to those for belowground combustion, with the exception of higher burn depth occurring in areas with lower extreme maximum temperatures and tasseled cap greenness (Fig. S40). The former is likely related to deeper burn depths occurring in the northern portions of the ABoVE domain (Fig. 5d), where long-term maximum temperatures are generally lower. Tasseled cap greenness was assessed after a given fire and can therefore be considered to be a metric of fire severity (low greenness = high severity).

Total emissions from ABoVE-FED are relatively consistent with past efforts, including AKFED and GFED4s, but with some important differences. Total emissions and mean combustion (Fig. S22) in Alaska were similar between ABoVE-FED and AKFED, which is expected given the similar field observations from Alaska used to develop these models. However, although AKFED was extended to the Yukon and Northwest Territories (Veraverbeke et al., 2017), it did not incorporate field observations from these regions. By utilizing 797 field plots across these provinces (albeit heavily dominated by the Northwest Territories), our results suggest AKFED tended to underestimate combustion in the Yukon and overestimate combustion in the Northwest Territories, especially during the large-fire year of 2014. ABoVE-FED also includes many more predictor variables than AKFED and is based on a different statistical model. We did not find large variations in mean combustion from year to year (Fig. 2), which is likely related to both the tendency of the random forest models to regress to the mean (Fig. S25) and relatively consistent observed mean combustion across large regions of the ABoVE domain (Walker et al., 2020a, c).

GFED4s is a widely used data source for global and regional burned area and fire emissions. Our results suggest GFED4s underestimates combustion across the ABoVE domain by roughly 1/3 (32 %; Fig. S34; mean of 3.39 kg C m-2 in ABoVE-FED compared to 2.30 kg C m-2 in GFED4s), leading to 36 % lower total emissions compared to ABoVE-FED (Fig. S33). This is consistent with previous regional studies noting a consistent underestimation for GFED4s emissions in Alaska (Veraverbeke et al., 2015) and the Northwest Territories (Walker et al., 2018). This result has important implications for quantifying and understanding the role of arctic–boreal fires in the global carbon cycle and climate. Regional- to continental-scale upscaling efforts such as ABoVE-FED, including the underlying field observation database (Walker et al., 2020a), can help inform further versions of global fire models and thereby improve our quantification and understanding of the role of wildfire in the global carbon cycle.

In contrast to AKFED and GFED4s, fire carbon emissions in FireMIP were an order of magnitude lower (94 %) than ABoVE-FED (Fig. S23b). This is likely due to the fact that most models in FireMIP only combust aboveground vegetation, whereas combustion of belowground soil organic matter constitutes the majority of emissions in boreal Alaska and Canada. This underscores the importance of developing algorithms that accumulate and burn soil organic matter within global fire models, which is important for both direct fire emissions and post-fire permafrost thaw and degradation (Genet et al., 2013; Jafarov et al., 2013; Natali et al., 2021; Treharne et al., 2022).

ABoVE-FED confirms the high interannual variability in fire carbon emissions in the ABoVE domain, including the large-fire years of 2004 in Alaska and the Yukon Territory, 2005 in Alaska, 2010 in Saskatchewan, 2014 in the Northwest Territories, and 2015 in Alaska and Saskatchewan. We also found general agreement with previous work (Turetsky et al., 2011) that large-fire years and later-season fires facilitate deeper burning and higher belowground carbon emissions, including the phenomenon that burn timing has a stronger influence on severity in small-fire years (i.e., extreme fire years result in high severity regardless of timing). However, these relationships varied depending on region and analysis technique and were often confounded by site-level factors and fire weather at the time of burn. Overall, however, this underscores the influence that climate change (warming, drying, and longer fire seasons) has on boreal fire severity.

Consistent with previous studies (Rogers et al., 2014; Veraverbeke et al., 2015; Walker et al., 2018; Dieleman et al., 2020), ABoVE-FED includes high uncertainty in combustion at the pixel level (2.86 kg C m-2). Much of this uncertainty likely arises from difficulty in predicting large combustion values, particularly from belowground sources (Fig. S25b). This suggests ABoVE-FED is underpredicting emissions coming from the most severe fire events between 2001–2019. We attempted to correct for this bias in a number of ways, including testing a variety of models (Fig. 6), tuning model parameters, assigning higher weights to the highest combustion values, and applying the synthetic minority oversampling technique (SMOTE; Chawla et al., 2002) to synthetically create more samples with higher combustion values. Ultimately, none of these approaches were able to correct for the low bias at high combustion levels without sacrificing performance for low combustion values. More field observations of high combustion combined with improved predictor variables (particularly drainage) may improve future model performance. Also consistent with previous studies, these pixel-level uncertainties were dampened through spatial averaging, such that domain-wide mean combustion had comparatively lower uncertainty (3.13 ± 1.17 kg C m-2).