The Monitoring Trends in Burn Severity (MTBS) program was established in 2006, with a mission to remotely assess the location, extent, burned area boundaries, and burn severity (see Fig. 1 for an example) of large fires using Landsat imagery on all lands across the conterminous United States (CONUS), and Alaska, Hawaii, and Puerto Rico, USA, for the period from 1984 to present (Eidenshink et al. 2007). This includes all fires ≥405 hectares in the western United States, Alaska, and Hawaii. In addition, the program maps and assesses fires ≥202 hectares in the eastern United States and Puerto Rico.
The MTBS program originally defined burn severity as visible alteration of vegetation, dead biomass, and soil that occurs within a fire perimeter (Eidenshink et al. 2007). These changes can be assessed on the ground (e.g., via the Composite Burn Index, CBI) and subsequently related to remotely sensed (e.g., via the differenced Normalized Burn Ratio, dNBR) estimates of burn severity (Eidenshink et al. 2007; see Fig. 1 for an example comparison between CBI and dNBR). CBI assesses damage to vegetated and dead biomass using a continuous index with values ranging from 0.0 to 3.0, while dNBR has been used to assess the changes in reflection in vegetated and nonvegetated surfaces resulting from fire (Key and Benson 2006). The dNBR metric is a measure of the difference between a pre- and post-fire NBR image, with typical values ranging between −2000 and 2000. Initial regression relationships between CBI and dNBR informed the MTBS program (Eidenshink et al. 2007), and subsequent investigations suggest that dNBR or a variant (e.g., Relativized dNBR [RdNBR] or Relativized Burn Ratio) can be used as a viable estimate of burn severity within some ecosystems in the United States ( Zhu et al. 2006; Picotte and Robertson 2011b; Cansler and McKenzie 2012; Parks et al. 2019).
Data produced by MTBS, including burn perimeters and severity products (https://www.mtbs.gov), have become increasingly critical in fire-related research in the United States. For example, MTBS data have been used in assessing trends in burned area extent (Finco et al. 2012; Dennison et al. 2014; Zhao et al. 2015; Picotte et al. 2016), burn severity (Zhao et al. 2015; Picotte et al. 2016), wildfire emissions (Urbanski et al. 2011; French et al. 2014), wildfire exposure and risk (Radeloff et al. 2018), and the effects of fuel reduction (Meigs et al. 2016). Although the MTBS mapping protocols outlined in Eidenshink et al. (2007) are largely still applied, some of the protocols have changed. Most importantly, the Landsat image archive becoming publicly and freely available in 2008 removed a major cost constraint, enabling the MTBS program to expand its scope and map many more fires.
Protocol evolution
Fire Occurrence Database
Current and historical fire records were initially compiled from Federal fire reporting databases, including the Incident Command System database and available state fire reporting databases, as part of the original data request for the MTBS Fire Occurrence Database (FOD). Inclusion of prescribed fires as part of the scope of MTBS resulted in a large availability of fire occurrence data that included both wildland fire and prescribed fire incident types. Because of the volume of fires, the lack of availability of prescribed fire records for all states, and the need for consistency in the MTBS data record, the decision was made in 2014 to no longer consider state-prescribed fire records for assessment. This decision impacted the mapping of fires in Florida, USA, which accounted for approximately 20% of mapped fires before 2013 and 9% afterwards (https://mtbs.gov/viewer/index.html, accessed 5 August 2019).
Other major changes that occurred in 2014 included the transition of fire occurrence tracking and mapping information into a relational database management system (MTBS internally used Event Tracking Database, ETD) and the change to a more automated system for compiling fire occurrence records. The large number of fire records assessed and mapped by MTBS resulted in the need for the development of an expanded ETD to also track mapping parameters, metadata, and all mapped fire occurrence information. Fire identifications (IDs) based on the new system were generated for current and historical fire records and were inserted into the ETD. Each fire record is uniquely established to intuitively identify the US state where the fire is located, the specific geographic location of the fire origin, and the date of fire ignition. Fire information (i.e., fire name, geographic coordinates for the location of ignition, size, ignition date, containment date, etc.) and input data, output products, processing and analysis parameters, and other metadata information collected for each mapped fire were uploaded into the ETD.
During the same timeframe as the ETD development, Short (2014) developed the Fire Program Analysis (FPA) FOD that included federal and non-federal fires from 1992 to 2011, which was subsequently expanded to include 2012 and 2013 (Short 2015). The FPA FOD was subsequently replaced by the Integrated Reporting of Wild-Fire Information (IRWIN; (https://www.forestsandrangelands.gov/WFIT/applications/IRWIN/background.shtml, accessed 5 August 2019) tool to collect and report fire event data with a unique identification for both federal and non-federal data. IRWIN was designed to ingest data from multiple, disparate fire reporting systems of record with automated capabilities to eliminate redundant records. Consequently, MTBS adopted IRWIN in 2014 as its primary source of fire records for ingest into the ETD. Once ingested into the ETD, each IRWIN record receives its own unique MTBS fire ID and is targeted for assessment and potential mapping by the MTBS program.
Landsat image considerations
The long history of the Landsat 30 m products (1984 to present), including the Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+), and Operational Land Imager, provided the MTBS program with a continuous source of data. Throughout the lifetime of the MTBS program, the Normalized Burn Ratio (NBR; García and Caselles 1991; Key and Benson 2006) index, calculated with the near-infrared and shortwave-infrared Landsat bands, has been used because of its sensitivity to identify spectral variation in burn severity. During the first year of the MTBS program, the dNBR was the only remotely sensed change detection product that was produced. To create the thematic burn severity product from dNBR, MTBS analysts visually interpreted the burn severity thresholds and compared these thresholds to outputs from techniques developed by Key and Benson (2006) that were originally developed and applied within Glacier National Park (Montana, USA) and further validated within CONUS and Alaska (Zhu et al. 2006). The thresholding process is therefore largely subjective, but see Fig. 1 for an example of an MTBS thematic image compared to ground-validated CBI plots and dNBR values. The MTBS program extended this methodology for use in biophysical settings throughout the United States.
The MTBS program continued to solely use dNBR for mapping burn severity until the development of the Relativized dNBR by Miller and Thode (2007). The two indices are fundamentally different: dNBR is an estimate of the absolute magnitude of change to vegetation and soil strata, whereas RdNBR, a variant of dNBR, estimates relative magnitude of change and potentially removes any bias of pre-fire vegetation conditions (Miller et al. 2009). Using RdNBR, for example, a stand-replacing fire in sparse shrub is rated as severely as one in dense forest (Miller and Thode 2007). Previous studies suggest that RdNBR has a stronger correlation with ground-collected metrics of burn severity (Eidenshink et al. 2007), especially forested settings and areas of high burn severity in relatively lower vegetation density settings. Beginning in 2007, the MTBS program began incorporating the standard RdNBR thresholds determined by Miller and Thode (2007) into the MTBS mapping workflow in order to provide a starting point for analysts to define burn severity thresholds (see Fig. 1 for an example of thematic burn severity classification).
At the inception of MTBS, it was anticipated that burn severity assessments would utilize imagery acquired at the peak-of-green during the growing season following a fire (i.e., an extended assessment). The intent of an extended assessment is to allow for delayed effects (mortality, survival, and recovery) to manifest on the landscape (Key 2005). Rapid regrowth of burned vegetation, however, made the exclusive use of extended assessments problematic in the southwestern and southeastern United States, and herbaceous areas throughout the CONUS. In the Southwest, fire effects that were easily observed in immediate post-fire imagery for low elevation fires (e.g., for the 26 May 2005 Duzak Fire in southern Nevada, USA) but disappeared by the next growing season because of rapidly regenerating herbaceous and shrub vegetation. This resulted in large areas that were characterized as “unburned to low” severity in an extended assessment, so an “initial assessment” became the preferred assessment strategy within areas of vegetation that rapidly regenerate.
Most fires in the Southeast occur in the late winter and early spring months and burned areas can green up within two to three months due to rapid regeneration of understory vegetation. Additionally, some areas in the Southeast exhibit low relief and poor drainage (e.g., wetlands), resulting in changing hydrologic conditions over the course of a year and between years (dry versus saturated soils). These areas frequently burn every two to four years and typically have relatively persistent cloud cover, which made the requirement for phenologically matching scene pairs difficult (Picotte and Robertson 2011a). After receiving critical feedback from Southeast resource managers and field assessments by MTBS staff, the MTBS assessment and mapping protocol for the Southeast was also modified to preferentially use an initial assessment within two months of fire start. Extended assessments are only performed if no suitable post-fire imagery is available. If no suitable pre-fire image is available, particularly in areas with inter-annual variability in hydrologic conditions, then a single-scene assessment is made using the post-fire imagery alone.
Special accommodations to the mapping protocol were also necessary to assess some fires in Alaska. The Landsat scene acquisition footprints overlap significantly in higher latitudes (Bindschadler 2003), which allows an area to be imaged two to three times in the nominal 16-day Landsat orbit cycle. However, persistent cloud cover can negate this advantage and many fires were not imaged for an entire year. Also, there were substantial periods of time (years) when no Landsat data were acquired due to ground station problems (Goward et al. 2006). The availability of quality and comprehensive Landsat data was further impacted by the lost capacity of the Landsat TM to temporarily store collected imagery on board and the preferences of ground stations in Canada and Alaska to acquire Landsat ETM+ over TM. If MTBS had strictly adhered to the one-year post-fire scene acquisition, many Alaska fires would not have been assessed by the program. In these cases, post-fire Landsat imagery acquired more than one year after the fire with no other significant land cover disturbances were used. Vegetation recovery in boreal forests can be slower than in forests at more southern latitudes (White et al. 2017; e.g., Alaska versus CONUS), which alleviates the need for rapid post-fire image acquisition.
Costs for imagery acquisition also had an important impact on the MTBS program. Before 2008, when the Landsat image archive became publicly and freely available, MTBS had to purchase Landsat imagery, which generally limited the project to two Landsat images per year for most Worldwide Reference System-2 (WRS-2) Path/Rows, unless additional scenes were acquired previously by federal partners in the Multi-Resolution Land Cover (MRLC) consortium. This cost-related restriction limited the ability of analysts to select the best available image for each fire assessment, resulting in potential burned area boundary and burn severity product errors as well as unmappable fires. In 2008, the Landsat archive was opened for free distribution of data (Wulder et al. 2012), which allowed for the reassessment of previously image-limited and unmappable fires. The MTBS recently completed revisiting 5012 fire records that were declared unmappable in the first few years of the program and was able to map 2248 of these fires. Meanwhile, the program revisited the largest fires that had been mapped spanning the period from 1984 to 2007 to determine if the best available imagery was used. Fires ranging from 2000 to 20 000 hectares were evaluated for their prescribed assessment strategy, errors in area burned, consistency in dNBR offsets (average dNBR value of unburned areas outside the burned area), applied burn severity thresholds, availability of quality pre- and post-fire imagery, and the phenological compatibility of selected imagery. When issues were identified for one or more of these evaluation criteria, fires were remapped using imagery previously not available because of costs and program budget constraints. Additionally, over 6000 occurrence records for relatively smaller fires were declared unmappable in the first few years of the MTBS program due to limits on imagery procurement. The MTBS program is currently revisiting these fires to determine if suitable imagery is now available to support their assessment, and mapping and will continue to revisit smaller fires from 1984 to 2008 as time and funding allow.
MTBS data caveats and limitations
The purpose of the MTBS burned area mapping product is to provide an estimate of the area that may have been affected by fire. This liberal approach to mapping burned areas leads to the potential for incorrect inclusion of unburned areas (i.e., errors of commission). Indeed, at the onset of the MTBS program, Kolden and Weisberg (2007) identified potential problems with commission error (mean commission error = 18%) when using Landsat data to map burned area boundaries in Nevada. Picotte and Robertson (2010) found similar commission errors in the southeastern United States (mean commission error = 15%). Incorrect exclusion of pixels (i.e., omission error) is also possible and was estimated to range between 0% and 45% by Kolden and Weisberg (2007). Errors in estimating burned area from Landsat data can result from the amount of time between when the area burned and image acquisition (Picotte and Robertson 2010; Picotte and Robertson 2011a), terrain complexity (Kolden and Weisberg 2007), and vegetation composition (Vanderhoof et al. 2017). Although MTBS attempts to obtain the best available pre- and post-fire Landsat images for burned area mapping, commission errors in the burned area extent can be problematic as a result of inherent problems with Landsat data quality and availability, and because MTBS does not remove unburned islands from within the burned area boundary (Eidenshink et al. 2007). MTBS product commission errors have been estimated to be 46.4% in the US Pacific Northwest (Meddens et al. 2016) and to range between 4.3% to 15.5% in the northern US Great Basin (Sparks et al. 2015). Burned area extent commission error could be mitigated by removing unburned areas as identified in the MTBS thematic burn severity product (Kolden et al. 2015), although this could result in the incorrect removal of low severity pixels, resulting in omission error.
Kolden et al. (2015) also identified three potential problems with the MTBS thematic burn severity production: the dNBR offset (phenology difference) is not applied in producing thematic severity data, burn thresholds are variable and subjectively determined, and burn severity thresholds are not tied to a quantifiable ecological measure (i.e., field validated). The dNBR offset, the mean value of unburned pixels near the fire boundary that occur within the same vegetation type (i.e., evergreen forests) as most-burned pixels, is currently calculated for each fire mapped by MTBS and is contained within the metadata for each MTBS fire. This offset could be applied (i.e., subtracted) from every dNBR image and from the burn severity thresholds to remove the phenological differences to potentially create more universally comparable burn severity estimates across time. Deriving dNBR offsets can be a subjective process. Consequently, applying the dNBR offset does not actually change the classified burn severity image. Offsets also do not mitigate the subjectivity of the burn severity thresholds but may correct the thresholds to make them potentially comparable between fires. Comparisons between analyst-selected thresholds and an automated Otsu (Otsu 1979) spectral thresholding procedure for 18 497 MTBS fires mapped between 1984 and 2014 suggest that the low, moderate, and high severity class breakpoints could be highly variable, although low severity breakpoints were similar (Picotte 2019). The MTBS program applies measures to ensure consistency among analysts in determining burn severity thresholds. Burn severity thresholds are consistently reviewed before release and analysts examine both classified vegetation products and high-resolution imagery to examine where low, moderate, and high burn severity breakpoints are occurring. There is also currently an effort to visually review all fires to ensure that visually determined burn severity thresholds occur with appropriate vegetation (e.g., no high severity in grasslands) and are near dNBR ranges suggested by Key and Benson (2006). None of these remote sensing efforts directly measure the ecological effects of burn severity, primarily because spatially and temporally comprehensive field data are not available for most fires throughout the MTBS data record. However, recent work does suggest that MTBS severity classifications can be associated with tree mortality (Meigs et al. 2011) or vegetation regrowth after fires (Johnston et al. 2019). Additionally, the program plans to leverage the increasing availability of CBI data (see Picotte et al. 2019 for dataset) to better calibrate remotely sensed MTBS burn severity estimates to ground conditions (see Fig. 1 for example CBI and dNBR values stratified by thematic classification). Regression relationships between CBI and dNBR and NBR could be developed to allow for automatic conversion of MTBS dNBR and NBR products to CBI estimates of burn severity. A similar approach using Random Forests (Pal 2005) to convert dNBR and NBR to CBI has already been developed by Parks et al. (2019).
