The methodology used for this project was driven by two fundamental requirements: 1) the need to develop consistent information across all lands within the project extent, and 2) the need to develop consistent information spanning a significant historical period. Based on these requirements, remotely sensed images were considered to be the only cost effective geospatial data source to consistently delineate and measure the response of thousands of individual fires across a continental extent and multi-decadal time frame. Many researchers have evaluated the effectiveness of various scales of remotely sensed data to characterize fire severity (Milne 1986, Chuvieco and Congalton 1988, Justice et al. 1993, Kasischke and French 1995, White et al. 1996, Fernandez et al. 1997, Patterson and Yool 1998, Pereira 1999, Sunar and Ozkan 2001, Diaz-Delgado et al. 2003, Sa et al. 2003, van Wagtendonk et al. 2004, Brewer et al. 2005, Key 2005, Roy and Landmann 2005, Smith et al. 2005). Scientific and operational precedent exists for the use of an approach based on remote sensing.
Landsat TM and ETM+ data provide the longest consistent record of relatively high spatial and spectral resolution data for mapping fire severity. Not only does this record enable the mapping of historical fire severity, it also facilitates the use of time-series approaches for characterizing post-fire effects. Landsat data have been shown to be responsive to relative changes in above-ground biomass as a result of fire (Lopez-Garcia and Caselles 1991, Kushla and Ripple 1998, Miller and Yool 2002, Epting and Verbya 2005). More specifically, multitemporal change detection approaches based on pre- and post-fire Landsat data have proven to be a cost effective and relatively accurate means of mapping fire severity (Brewer et al. 2005). The availability and low cost of Landsat data were additional factors supporting their use for a project of this geographic and temporal extent.
Multi-temporal approaches that apply image ratios and image differencing techniques to Landsat data have been developed for a variety of assessment objectives. Imagery is commonly transformed mathematically into indices by ratioing one or more spectral components or bands for each pixel. The transformation of Landsat data into vegetation indices (e.g., Normalized Difference Vegetation Index) has been widely used to strengthen the relationship between spectral response and vegetation characteristics, and a number of such indices exist (Lyon et al. 1998). Lopez-Garcia and Caselles (1991) published the first index specifically derived to enhance the relationship between Landsat spectral response and burned vegetation. This Normalized Difference index was combined with multi-temporal differencing and subsequently adapted and operationally implemented by Key and Benson (2002), who used it to develop historical fire severity data and atlases on several National Park Service lands. This approach has been named the Normalized Burn Ratio (NBR) and has been used in fire severity mapping efforts by the USGS and the Forest Service since 2002.
The Normalized Burn Ratio is used to enhance the spectral response of fire-affected vegetation. The Normalized Burn Ratio is calculated from TM bands 4 and 7 as: (TM4 − TM7)/(TM4 + TM7) where TM4 represents the near-infrared spectral range (0.76 µm to 0.90 µm) and TM7 represents the shortwave infrared spectral range (2.08 µm to 2.35 µm). Differenced NBR images (post-fire NBR subtracted from pre-fire NBR) are referred to as dNBR images. The differenced pre-fire and post-fire NBR images result in a fire-related change image that is classified into severity classes and provides an unbiased basis for analyzing additional fire effects. Figure 2 illustrates the process of deriving fire change and severity images from Landsat data.
The dNBR data have been operationally used for both rapid response and initial assessments, and for extended assessment and monitoring (Bobbe et al. 2003, Key and Benson 2002, Gmelin and Brewer 2002). For initial assessments, imagery acquired immediately after a fire is used to characterize first-order fire effects on vegetation and soils, and to facilitate the prioritization of rehabilitation resources. Extended assessments have relied on image data typically acquired during the growing season following the fire in order to capture delayed first-order effects (e.g., delayed tree mortality) and dominant second-order effects that are ecologically significant (e.g., initial site response and early secondary mortality agents). Extended assessments are intended to provide a more comprehensive ecological indication of fire severity than initial assessments. In both initial and extended assessments, there is a level of uncertainty in the characterization of fire severity. Pre-fire vegetation conditions and post-fire management activity influence the nature and magnitude of this uncertainty. The decision to use an initial or extended assessment should be based on specific management objectives.
Based on the scientific foundation in the literature and on operational precedent, the dNBR approach was selected to characterize fire severity and to delineate fire perimeters for this project. Extended assessments will be conducted on forest and shrub ecosystems and initial assessments will be conducted on grasslands and specific vegetation communities known to recover from fire within a single growing season. A simple production model was developed around this approach to ensure timely and consistent products. The following steps outline the process:
-
Fire history database compilation
-
Image data selection and pre-processing
-
○ scene selection
-
○ pre-processing
-
○ delivery and archiving
-
Fire severity interpretation and perimeter delineation
-
○ Normalized Burn Ratio calculation and differencing
-
○ interpretation and thresholding into severity classes
-
○ dNBR partitioning
-
○ dNBR fire perimeter delineation
-
Stratification and summarization of severity information
Fire History Database Compilation
Existing fire history databases were compiled into a single standardized project database that formed the basis for image scene selection. Fire history sources were generally from federal agency databases and state databases. In some cases, state and federal agencies have collaborated in developing and maintaining a single database for state and federal incidents. Federal agency data are aggregated into the Incident Command System database known as the ICS 209 (named after the form number used to report incident status), maintained by the National Interagency Fire Center (NIFC) in Boise, Idaho (http://www.nifc.gov/). ICS 209 data make up most of the records in the MTBS project database. States were solicited for fire occurrence data when it was uncertain whether the fires were included in the ICS 209.
The ICS 209 and state databases required preprocessing to ensure data accuracy and consistency. There is some level of standardization within ICS 209, but federal land management agencies have varying standards for content, geospatial accuracy, and nomenclature that are reflected in the database. Duplicate records are common because a given incident may be reported by several agencies, and there are cases of gross geospatial inaccuracies. Similar inconsistencies and errors have been observed within and across state databases. Data were standardized and corrected as part of the compilation of an MTBS project database. For the purposes of this project, standardization was accomplished by selecting data elements common to the source databases and not through record editing or manipulation of the source data, except for geospatial coordinates. Coordinates were adjusted if a record was grossly and obviously incorrect, and a correction could be made confidently. The elements that comprise the MTBS fire history database are as follows:
-
ID — Unique MTBS ID that include source ID
-
Fire Name — Incident Name from the source database
-
Agency — Reporting Agency from the source database
-
Year — Year Occurred from the source database
-
Start Date — Incident Start Day/Month/Year from the source database
-
Reported Area — Incident area from the source database
-
Long — Longitude
-
Lat — Latitude
-
Path — Landsat Path
-
Row— Landsat Row
-
Disposition — Description of issues relative to a fire’s visibility or spatial accuracy on the imagery
Records for these elements were extracted from the ICS 209 and state data sets, and source links were included to ensure data could be traced to their databases of origin. The spatial distribution and relative frequency of fire occurrences across the United States is depicted in Figure 3. Some discrepancies in the fire records are likely because of omissions in reporting and error in geographic locations within the fire records, particularly in the central and eastern United States. The fire history database compiled by MTBS will be a geospatial record of fires greater than 202 hectares (500 acres) in the east, and 404 hectares (1,000 acres) in the west.
Image Scene Selection and Data Pre-processing
Scene selection is driven by the MTBS fire history database. Scenes are selected using the USGS Global Visualization Viewer (GloVis) developed by USGS EROS (http://glovis.usgs.gov/). Enhancements were made to GloVis to accommodate the magnitude of effort required to select scenes for this project. These enhancements, available to all GloVis users, include the ability to load ArcGIS shapefiles in the viewer to aid scene selection, and to view scene-specific graphs of seasonal patterns of vegetation condition to help determine peak periods of photosynthetic activity, or ‘peak of green’ periods. A shapefile of the fire history for the specific area of interest can be loaded into the GloVis viewer and analysts use fire locations to guide scene selection for each fire. Pre- and post-fire images are selected for each incident. Scenes selected for an extended assessment are based on ‘peak of green’ condition or as close in time as cloud-free data are obtainable. Limitations in data availability because of cloud cover will naturally compromise scene selections for fires. Northern latitudes will also be subject to a shorter period of optimal scene selection because of low sun angles throughout late fall and early spring.
Selected scenes are processed according to existing USGS EROS protocols. Image data are geometrically registered, terrain-corrected, and radiometrically corrected using the NLAPS system, and then delivered to EROS and RSAC analysts to be processed into fire severity information. It is estimated that the MTBS project will acquire more than 7,000 Landsat scenes, all of which will be available for download or on media for a nominal charge. The USGS National Satellite Land Remote Sensing Data Archive will serve as the primary repository for MTBS image data. GloVis can be used to acquire the imagery.
Fire Severity and Perimeter Mapping
The NBR index is calculated for pre- and post-fire images as described in the Methods. Pre- and post-fire images are inspected for co-registration accuracy and corrected if spatial differences are systematic, excessive, and extensive (>30 meters). NBR images are differenced for each fire scene pair to generate the dNBR. A “relativized” dNBR (RdNBR) is also calculated, using a formula based on the work of Miller and Thode (2007). The RdNBR data have been shown to have stronger correlations than dNBR to Composite Burn Index plot data in some western ecosystems (Thode 2005, Miller and Thode 2007). While dNBR data and associated analysis are more extensively represented in the literature and operational use, RdNBR data have recently been used to report trends in fire severity in the Sierra Nevada (J.D. Miller, Forest Service, unpublished data) and can be expected to support future analysis in other western regions. The MTBS project intends to provide data calculated from both dNBR and RdNBR algorithms to support more localized trend analysis. The sequence of data layers generated is shown in Figure 2.
Ecological Severity Thresholding. Deriving the dNBR from Landsat imagery is a straightforward series of objective calculations requiring limited analyst interaction and relying principally on automated production sequences. After dNBR is calculated, the process of developing fire severity and perimeter maps is much more dependent on analyst interpretation. The dNBR data are calculated as signed 16-bit integers with a maximum digital number (DN) range of −32,282 to +32,282. However, the practical range of DN values representing fire-related change and no change is typically within −2,000 to +2,000. Values increasing from zero represent greater change as a result of both first- and second-order fire effects (which occur within the fire perimeter). Negative values of dNBR indicate a positive vegetation response (growth) and positive values indicate a negative vegetation response (mortality). A dNBR image for the Cerro Grande fire (2003) is shown in Figure 4a and the associated data range is shown in Figure 4b. The analyst evaluates the RdNBR and dNBR data range and determines where significant thresholds exist in the data to discriminate between severity classes. Interpretations of the dNBR and RdNBR data are aided by raw pre- and post-fire satellite imagery, plot data, and the analyst’s own experience with fire behavior and effects in a given ecological setting. Composite Burn Index (CBI) data (Key and Benson 2006) have been the most commonly collected ground-based data to estimate post-fire effects. Correlations between CBI and dNBR have been used to demonstrate the sensitivity of dNBR to post-fire effects and to establish numerical thresholds in dNBR data that discriminate severity categories (Cocke et al. 2005, Key 2005). When published dNBR relationships are available, analysts will use them to guide their interpretations. Limited interpolation of plot-based thresholds within ecologically similar conditions are examined.
Thresholding dNBR data into thematic class values results in an intuitive map depicting a representative number of ecologically significant classes. Within this project, the thematic raster data will characterize severity in five discrete classes: unburned/unchanged, low severity, moderate severity, high severity, and increased post-fire response. A single theme labeled Non-processing Area Mask is used to identify areas affected by clouds, cloud shadows, and data gaps, specifically the gaps within a Landsat 7 SLC-off product as described by the USGS Landsat Project (2007).
Determining thresholds for the burn severity classes is a significant quality control issue. It is understood that when several individuals are involved in mapping burn severity over a wide variety of landscapes that some subjectivity will be introduced. Consistency in characterizing burn severity is critical to the understanding of long term trends. In order to maintain consistency of results, a series of fires over a wide variety of landscapes have been selected for cross calibration of the burn severity thresholds. Each member of the mapping team maps the series of fires. The results of each member of the mapping team are discussed to identify what the rational was for quantifying the thresholds. When feasible, fires with associated plot data are chosen for analysis. A consensus approach is identified and the results are registered in a reference database. The mapping team uses the reference database as training and validation for mapping fires occurring in similar conditions. This approach will provide a practical, intuitive means to sum severity by area burned across broad scales and they provide a coarse look at the gradient of effects within fires. Finer-scale analysis may best be conducted on the continuous dNBR data, which provide the greatest range of data quantifying post-fire change.
Although not a direct measure of fire severity, dNBR data have been shown to correlate to field-based estimations of fire severity (Hudak 2006, Key 2005). Since these correlations will vary between fires, the grain of continuous data offers the most flexibility to evaluate severity at the individual fire scale. Analysis of multiple fires with continuous data requires the data to be normalized due to variation in reflectance data caused by inter-annual variation in phenology and site moisture. Variation due to atmospheric conditions and sensor anomalies are assumed to be corrected through the satellite data processing.
Ecological significance is issue-dependent and one set of thresholds cannot be expected to apply equally well to all analysis objectives and management issues. Other severity classifications such as those described by Stephens and Ruth (2005) may be used as the basis for thresholding, but must be considered for the appropriateness of their application to dNBR data. Fire severity classifications that are based on fire effects not readily discernible on Landsat data (e.g., subsurface biomass combustion or soil chemistry changes) should not be applied to these data.
dNBR Partitioning
In addition to setting ecological thresholds as a means of discriminating severity classes, dNBR will be arithmetically partitioned into discrete classes to facilitate objective and flexible pattern and trend analysis. Arithmetic partitioning is not intended to provide information on the ecological severity of fires at large spatial scales or for short time periods. Methods for partitioning dNBR have yet to be determined and the algorithm(s) and subsequent grain of partitioning will depend on the ability to reveal meaningful patterns in fire severity over time. Gmelin and Brewer (2002) used a simple equal interval calculation to establish objective burn severity classes between observed unburned and high severity conditions in the Northern Region of the Forest Service. Brewer et al. (2005) used the same approach in a methods comparison study that concluded dNBR was the most effective approach of those evaluated for mapping fire severity. The relative ease and quickness of arithmetically partitioning dNBR data will allow for rapid evaluation of meaningful spatial and temporal scales in the context of fire severity trends. Moreover, dNBR data can be efficiently analyzed and classified to suit the fire severity information needs of a specific management issue.
Perimeter Delineation. Fire perimeters are generated by on-screen interpretation and delineation of dNBR images. Analysts will digitize perimeters around dNBR values reflecting fire-induced change. To ensure consistency and high spatial precision, digitization will be performed at on-screen display scales between 1:24,000 and 1:50,000. Data showing incident perimeters, where available, will be used in an ancillary fashion to aid the analyst. Incident perimeters can be particularly useful in identifying unburned islands within a fire or outlining an isolated, disjunct burned area outside the main fire perimeter. Because of limited availability and inconsistent spatial precision, incident perimeters were not considered appropriate as a source for MTBS project perimeters.
Data Summarization
Tabular data will be generated from statistical summaries of the fire severity class layers. Reporting units will vary in extent depending on the needs of WFLC, but at a minimum summary data will be produced for each project mapping zone as well as at a national extent. Three sets of tabular data are currently specified in the MTBS product suite and are listed in the Introduction. Of the three, “area burned by severity class” is the statistical summary that is most directly extractable from the spatial data.
Summarizing area burned by severity class and vegetation cover type requires consistent geospatial vegetation data of similar resolution. Initial MTBS reporting efforts will use land cover classes from the 2001 National Land Cover Database (NLCD, Homer et al. 2001) for national and state summaries. Other land cover strata, such as the existing vegetation types currently being mapped by the LANDFIRE program, will offer a spatially extensive, nationally consistent, and more detailed alternative by which severity classes can be summarized.
A composite database containing additional ecological and administrative spatial units, including fourth-level hydrologic units (cataloging units) (Seaber et al. 1987) and federal ownership, will be available to enable users to summarize MTBS data for larger areas. The production and distribution of the spatial data sets described in the Methods constitute the primary geospatial data legacies available to scientific and operational interests outside this project. Summarization of area burned by severity class in relation to other geospatial information is feasible. For example, the National Fire Plan Operations Reporting System (NFPORS) database is the primary standardized federal database containing fuel treatment data in digital format, as described on the Web site (http://www.nfpors.gov/). Tabular data generated under these criteria will only be applicable to specific administrative and geographic extents.
Data Distribution
All spatial and tabular data will be distributed through Web-based interfaces. Existing data portals maintained by Forest Service and the USGS (http://www.mtbs.gov/ and http://mtbs.cr.usgs.gov/viewer.htm) will be primary access points as the data and associated reports are completed and become available. Additional distribution nodes may be developed in partnership with other federal and academic institutions.
After completion of the first historical data sets, a technology transfer phase of the project will be initiated. This effort will educate potential users about the structure and content of burn severity data, and explore applications of the data at multiple scales. Independent studies will reveal how useful MTBS data are and discover limitations that will guide operational use. The technology transfer phase will attempt to synthesize internal and external assessments of data usefulness and provide an efficient means to access these assessments. Web-based tutorials and workshops will be used to engage potential users.