|
Burn Severity Mapping Using Simulation Modeling and Satellite Imagery Robert Keane1, and Eva Karau1 (point of contact) 1USDA Forest Service, Rocky Mountain Research Station, Fire Sciences Laboratory ABSTRACT As wildland fire becomes an increasingly important issue affecting our nation’s landscapes, land managers must be able to quickly assess fire effects to efficiently allocate rehabilitation resources to areas under their custody. Satellite image-based burn maps can be quickly generated to provide a landscape view of relative fire severity, while fire effects simulation models provide biotic context to the effects of the burn. These techniques could be used synergistically to improve burn severity mapping capabilities of land managers, enabling them to quickly and effectively meet rehabilitation objectives. With this project, we seek to evaluate two tools that managers can use to assess burn severity immediately post-fire, and explore the predictive utility of a landscape scale fire effects model in prioritization of fuels reduction treatments. Considering results from a previous investigation, we propose to expand our exploration and demonstration of these tools through testing and evaluation of the FIREHARM fire effects model, and comparison of satellite burn severity images to field and modeled data. We will develop a set of tools and procedures for running the model and generating burn severity images. Additionally, we will hold two informational workshops that will teach managers how to use these tools.  PROJECT OBJECTIVES
• Evaluate ΔNBR, a satellite imagery-based burn severity mapping technique and FIREHARM, a landscape scale fire effects simulation model, as tools that managers can use to assess burn severity immediately post-fire. • Explore the predictive utility of a landscape scale fire effects model (FIREHARM) in prioritization of fuels reduction treatments • Explore the possibility that by combining satellite-based and model-based burn severity mapping technologies we can achieve biophysically relevant landscape views of fire effects. STUDY PLAN INTRODUCTION In an era where wildland fires have become an increasingly important phenomenon shaping our nation’s landscapes, the need for accurate, efficient and economical methods for assessing fire severity is imminent. New burn severity mapping innovations will aid resource managers in more effectively meeting burned area rehabilitation objectives. The USFS and other land management agencies, along with BAER (Burned Area Emergency Rehabilitation) teams, currently use maps derived from satellite imagery to help direct their post-fire rehabilitation efforts (Brennan and Hardwick, 1999; Robichaud et al., 2003). Burn severity modeling is another tool developed to provide estimates of the severity of fire effects. FIREHARM, a landscape scale fire effects model, is designed to output physically based estimates of fire effects that can be used to quantitatively describe burn severity (Keane, in prep.). Preliminary evaluation of satellite and model-derived burn severity maps suggests that these two approaches will provide a unique suite of tools that managers can use to assess burn severity (Karau and Keane, in prep.).
BACKGROUND Fire Effects Modeling FIREHARM is a landscape scale spatial model that will provide dimensioned biophysical fire effects estimates and quick views of fire severity. Model output includes data layers that describe fire behavior and effects for specific weather conditions (Keane, in prep.). The model incorporates a spatial version of FOFEM (First Order Fire Effects Model), which provides quantitative physical values for tree mortality, fuel consumption, mineral soil exposure, smoke production and soil heating (Reinhardt et al., 1997). Several input data layers are required to run FIREHARM. These include biophysical site descriptors (elevation, aspect, slope, soil depth, soil composition and leaf area index), fuel model, fuel loading class and cover type maps, and weather and fuel moisture measurements. Conveniently, most of the FIREHARM input data will be available for the continental United States upon completion of the LANDFIRE project. LANDFIRE is a multi-agency effort to provide land managers with comprehensive spatial data and planning-focused analysis tools. It will enable agencies to more efficiently and effectively manage their landscapes in accordance with the National Fire Plan (Rollins et al., 2003). In most cases, the effort required for managers to create the input data layers required to run a model like FIREHARM would be prohibitively time and resource intensive (Reinhardt et al., 2001). However the availability of LANDFIRE data layers will enable managers to easily run FIREHARM to generate fire hazard and burn severity maps. The FIREHARM model is equipped to calculate fire behavior, fire danger and fire effects variables. For this study, the pertinent output data are three fire effects estimates: fuel consumption, soil heating and tree mortality. These are quantitative and direct estimates of the physical effects of fire. For this study, we will run FIREHARM under the assumption that each fire is a single event and we will specify the weather and fuel moisture conditions for the fire in an “event” file. The resulting fire effects output will be tuned directly to the actual conditions of the fire. Image Based Burn Severity Mapping Remote sensing has been used for various burn mapping applications including: coarse spatial resolution global biomass burning and land cover/land use change research (Chuvieco and Martin, 1994; Dwyer et al., 1998; Kasischke et al., 1992), vegetation change detection and landscape ecology studies using fine spatial resolution imagery (Chuvieco, 1999; Jakubauskas et al., 1990; Turner et al., 1994; White et al., 1996), and investigations which directly support land management and post-fire rehabilitation efforts (Bobbe et al., 2004; Hardwick et al., 1998; Zarriello et al., 1994). The foundation of these studies relies upon the typical spectral reflectance characteristics of landscape features. For example, in burned areas, increased bare ground area and decreased moisture elevates reflectance in mid-infrared spectral bands (Yool, 1999), while a reduction in healthy vegetation “reduces near-infrared reflectance in direct relation to the intensity of the fire” (Jakubauskas et al., 1990). The Normalized Burn Ratio (NBR) takes advantage of these band-specific burned area reflectance characteristics as it is a linear transformation of the mid and near infrared spectral bands from a satellite image (Key and Benson, in press). NBR is calculated on a single image as: NBR = (ρMIR - ρNIR) / (ρMIR + ρNIR) (Equation 1.) Where ρMIR = mid infrared reflectance, and ρNIR = near infrared reflectance. To capture fire-caused landscape change, (Key and Benson, in press) compute a difference of pre-fire and post-fire scenes: ΔNBR = NBRprefire – NBRpostfire (Equation 2.) Where the pre-fire scene is chosen from the year previous to the post-fire scene, ideally during a phenologically similar period. Synergy of Burn Severity Mapping Approaches Remotely sensed imagery and fire effects models provide an extensive view of fire severity for large regions. Both technologies facilitate generation of quick and inexpensive maps, minimizing the need for resource-intensive and potentially dangerous field sampling. Though they share some benefits and capabilities, these methodologies differ greatly in process and product (Table 1). The ΔNBR imagery approach is based on the observed changes in linear combinations of surface reflectance values between pre and post-fire images. Thus, it is essentially the reflectance of earth features that is measured and differenced from date to date; imagery does not represent a biophysical process or effect. ΔNBR image values are unit-less indices which can be sliced into categories representing relative levels of fire effects (e.g. high, medium and low); this categorization can facilitate a quick, simple and informative summary display of relative fire severity across the landscape. ΔNBR can also be used as a continuous variable, in which case each pixel has a unique, uncategorized value. The FIREHARM modeling approach will provide dimensioned fire severity variables that are direct estimates of the damage caused by fire. Once the LANDFIRE data layers are developed and archived for the nation, managers will be able to run FIREHARM to quickly generate severity maps that have fire effects measurements in physical units, which are perhaps more meaningful, depending on the project objective, than a relative index of severity. For example, actual tree mortality, fuel consumption and soil heating estimates will allow the manager to fine tune management actions to specifically focus burned area rehabilitation efforts based on the type and extent of damage incurred. Additionally, FIREHARM has utility as a prognostic tool. Users can simulate best and worst case scenarios for possible situations that may occur in their region during the fire season, or they could use the model to guide the scheduling and location of fuels treatments.
Table 1. Comparative view of model-based and imagery-based burn mapping methods.
Advantages Disadvantages
Fire Effects Modeling (FIREHARM) The model provides biophysically based fire effects estimates. The model can be run at anytime and for any place if input data layers are available. The model will be in the public domain, so users incur no simulation cost. Upon completion of the LANDFIRE project, the data needed to run the model will be free and readily available. FIREHARM could be used as a predictive tool, and it has fire hazard mapping capabilities. Significant computing resources (memory and processor speed) are necessary to run the FIREHARM model for large landscapes. As FIREHARM is a computer model, it is a simplification of reality. FIREHARM accuracy depends on input data layer reliability. Various topographic and ecophysiological data layers must be developed in order to run the model in event mode. Pre-fire weather and fuel moisture data must be collected/calculated in order to run the model.
Remotely Sensed Imagery (ΔNBR) An image of the burned area is instantly created, providing a perimeter of the burned area.
Linear combinations of reflectance values mirror landscape patterns and can be classed into intuitive fire severity categories. In some cases, imagery is immediately available, facilitating an instantaneous ΔNBR assessment, and timely burned area rehabilitation. The image is only a picture, typically categorized into relative severity classes. It does not provide information about biophysical processes. Images may be unavailable or unacceptable due to instrument malfunction, timing of satellite overpass, or smoke/cloud obstruction. It may be difficult for some agencies or individuals to purchase images due to availability and cost. ΔNBR must be calculated after the fire. It cannot be used as a predictive tool. Whereas FIREHARM input data will be consistent and accessible to users, availability of good quality satellite images is variable, as evidenced by the recent Landsat 7 instrument malfunction which causes significant reduction in usable image data. It is also possible that smoke or clouds obscure post-fire images. In these situations, users will be able to use modeled-data as a compliment to their satellite burn mapping efforts. Models and imagery could be used in tandem to generate a suite of burn severity mapping products with more dimensionality than ever before. This study has the benefit of not just comparing imagery and models, but comparing both of these approaches to field measured fire effects. Based on study results, we will develop a protocol for efficient and accurate assessment of fire effects that will be used to quantify fire severity. This protocol may involve a combination of modeling and remote sensing methods if we find that the optimal burn mapping system involves a biophysically-focused fire modeling approach, along with the imagery-based approach which provides an integrated view of fire effects.
METHODS AND MATERIALS
We propose to implement and explore remote sensing and model-based burn severity mapping technologies for 2-3 fires that either occurred during the 2004 field season or will occur during the 2005 or 2006 seasons. This project will expand upon previous investigations of burn severity mapping as we enhance our understanding of these tools in three ways: 1. We will implement FIREHARM, a landscape scale fire effects model on several fire-affected landscapes. We will assess model performance using a comparison of model output with ground-reference data.
2. We will generate satellite burn severity images using ΔNBR methodology. Then, we will compare ΔNBR maps data with ground-reference data.
3. Finally, we will compare FIREHARM and ΔNBR maps with ground reference data. Fire effects field data will serve as ground reference. Based on comparison results, we will develop a procedure guide that will allow managers to simply and systematically generate ΔNBR images and to run FIREHARM. We will hold a two-day workshop to demonstrate these tools. The workshop and procedures guide are detailed in the Deliverables section of this proposal.
STUDY AREAS Ideally, study areas for this proposed project will fall within the boundaries of Multi-Resolution Land Cover (MRLC) zones 16 or 19, as these are the prototype areas for LANDFIRE data layer development (Figure 1 shows the boundaries of both zone 16 and 19 map zones.) Depending on fire activity in zones 16 and 19 during the 2005 and 2006 fire seasons, we will choose one or two study sites for an initial assessment of fire severity. We will use the following criteria when choosing study sites: fires should be greater than 5000 acres in size, the burned area should include mixed severity fire effects in at least some forested area, and sites should be accessible by road or trail to facilitate efficient field work. We possess field and satellite data for the Cooney Ridge and Mineral-Primm fires that burned in Montana in 2003. We will compile and analyze these data to help refine our methodology and provide preliminary project results. We have also identified a burned area that would be a suitable candidate as a subject for our proposed analysis and we could sample this fire during the summer of 2005. Located in the Manti-La Sal National Forest in southern Utah (within zone 16), the Six Mile Wildland Fire Use fire was started by lightning on July 27, 2004. The fire grew to approximately 5000 acres and burned with mixed intensity through spruce/fir, aspen, open sage and grassland cover types. As this study area burned in 2004, we would necessarily do an extended landscape assessment for this fire (see Image Processing section below). Figure 1. MRLC Zones 16 and 19 FIELD WORK
We will select field plots based on the criteria mentioned above, with the goal of attaining 30-40 burned plots per fire. Because we need matched burned/unburned pairs in order to calculate fuel consumption, plots cannot be randomly selected. A selected burned plot must have a corresponding unburned plot that is very similar in site characteristics (slope, aspect, elevation) and vegetation conditions (cover type, structural stage, fuel type). It is critical that the unburned plot match the burned plot in order to adequately characterize pre-burn fuel loads. The total number of plots per fire will vary, as there will be cases where one unburned plot may serve as a surrogate for two or more burned plots. We will aim to collect an equal number of plots in each of three burn severity classes (high, medium and low). We will follow FIREMON sampling protocol (http://fire.org/firemon) and use corresponding plot sheets for all data listed in Table 2, with the exception of soil char depth. The same data is collected on burned an unburned plots, with additional sampling of CBI and soil char depth on burned plots only. Example plot sheets are attached to this document, along with field and safety equipment lists. Table 2. Proposed data collection for burned and unburned plots.
Burned Plots Unburned Plots
Plot Description (PD)
Tree Data (TD)
Fuel Loading (FL)
Composite Burn Index (CBI)
Soil Char Depth Plot Description (PD)
Fuel Loading (FL)
The circular macroplot for Plot Description (PD) and Tree Data (TD) encompasses a 1/10 acre (37.2 foot or 11.34 meter radius). Trees that are less than 4.5 inches in diameter are counted as saplings and are measured at the macroplot level. Snags will also be measured at the macroplot level. Seedlings, trees that are less than 4.5 inches and diameter and also less than 4.5 feet tall, are measured in a 1/100 acre (11.7 foot or 3.57 meter radius) microplot nested within the macroplot.
For fuel load sampling (FL), we establish as many sampling planes as it takes to get 100 pieces of down woody debris (DWD); at minimum, we establish three planes, oriented 90˚, 300˚ and 270˚ true north. The sampling plane for 1-hour and 10-hour fuels extends 6 feet (2 meters) from the 10 foot (3.048 meter) mark of the tape, which has its origin at plot center. The sampling plane for 100-hour fuels extends 10 feet from the 10 foot mark of the tape. We estimate vegetation cover and height and take duff/litter depth measurements at the 30 foot (9.144 meter) mark and 60 foot (18.288 meter) marks of the tape. We take soil char depth measurements along the fuels transects at the 30 foot (9.144 meter) mark and 60 foot (18.288 meter) marks of the tape. The CBI methods will originate from the same center point as the PD, FL and TD macroplot, but we measure out 30 meters from plot center for the plot radius. We estimate CBI scores for all rating factors on the CBI plot sheet and an overall rating is determined for each plot. We take at least two digital photos at each plot, one facing due north and one facing due east. Infrequently we encounter unusual circumstances that force us to modify sampling methods. In these cases, we use judgment to reconcile the situation and take detailed notes about the modification. For example, if we began to set up a fuels transect and found that the next compass direction would route the transect through a river, we would choose the next listed compass direction for that transect, and take notes to document the situation. Similarly, we encountered an issue at the Sanford fire site that may not be specific to that region. Because there is a spruce beetle infestation in the area, it was sometimes difficult to determine if a tree were dead or dying due to the beetle or due to fire. If it was clear that fire had moved through the area and seemed to be the cause of death for trees of other species, we counted those spruce trees as dead due to fire.
Field data will be entered into the FIREMON database, where fuel loading, tree data and CBI values will be calculated. To calculate fuel consumption values for the field data, we simply subtract fuel loads measured on burned plots from those measured in the adjacent unburned plots. Soil heating is derived from soil char depth measurements. These numbers will be used as reference data in comparisons with model-derived and image-derived burn maps.
IMAGE PROCESSING FOR SATELLITE BURN SEVERITY MAPPING We propose to demonstrate satellite burn severity mapping methodology by generating ΔNBR imagery for our study areas. Satellite imagery will be processed following methods outlined in the Landscape Assessment section of the FIREMON protocol (http://fire.org/firemon). There are two strategies described in this protocol to assess fire severity with ΔNBR imagery. The strategies differ in pre and post fire image date selection. Initial Assessment requires a post-fire image that is from directly after the fire and the pre-fire image should be from a similar period of the previous year, so that the images match phenologically as much as possible. This image timing is designed to capture immediate fire effects, but does not reflect delayed mortality and may overestimate severity, as vegetation will have had no recovery time. For the second strategy, Extended Assessment, pre-fire imagery is selected from directly before the fire and post-fire imagery is from one growing season after the fire, matching the phenology of the pre-fire imagery as much as possible. For the Six Mile fire study area, we will perform an Extended Assessment, as it is only possible to gather our field data the year after the fire. For fires that occur during the 2005 fire season, we will perform an Initial Assessment. To avoid the problem of scan lines of missing data due to a Landsat 7 ETM+ sensor malfunction in May of 2003, we will use Landsat 5 TM data. After acquiring the imagery, we will perform any necessary image pre-processing (image calibration and atmospheric correction) (Jensen, 1996; Key and Benson, in press). For each image, in each assessment, the Normalized Burn Ratio (NBR) will be calculated as described in the equations 1 and 2. Under this formulation, positively increasing ΔNBR values theoretically correspond to increasing fire severity and ΔNBR values that increase negatively (in absolute value) correspond to areas that have enhanced vegetative regeneration after fire. Unburned areas should have ΔNBR values around zero, though there is often some “noise” in the image, where there are high or low ΔNBR patches due to landscape changes that are not related to fire (e.g. agriculture, urbanization or logging). IMAGE-BASED BURN SEVERITY MAPPING ANALYSIS We will use linear regression to statistically explore the relationship between ΔNBR imagery and field data. Regression model slope and intercepts, along with correlation coefficients will be used to assess the nature of the linear association between ΔNBR index values and field-based measurements of CBI, fuel consumption, tree mortality and soil heating. Scatter plots, charts and tables will be generated in Excel or SAS to illustrate the relationship between the variables. With these results, we will gain an understanding about the relation between satellite burn severity data and actual fire effects for our study areas. FIREHARM SIMULATION For each record in a list, the FIREHARM model will estimate fire danger indices, fire behavior measurements and fire effects. For this study, we are concerned with three fire effects estimates: fuel consumption, tree mortality and soil heating. The records in the user-defined list could represent stands, polygons or pixels and model output will be mapped back to those units after simulation.. To run FIREHARM, we need to obtain pre-fire weather and fuel moistures for the event file, create the polygon input file and enter appropriate driver file parameters. ArcInfo AMLs will be developed that prepare the following input data for the polygon file: • Site Map – from LANDFIRE input layers
• Cover Type map – from Eros Data Center
• NFDRS Fuel Model map – already available at the lab?
• Fire Behavior fuel model map – already available at the lab?
• Fuel loading class map – from Jennifer and Duncan (June 2004)
• Topographic layers: elevation, aspect, slope – from LANDFIRE input layers
• Leaf Area Index – from LANDFIRE input layers
• Soil depth – from LANDFIRE input layers
• Soil composition (percent sand, silt and clay) – from LANDFIRE input layers
The model will output ASCII files with tree mortality, fuel consumption and soil heating as columns. This tabular output will then be mapped back to the original polygon units using the Polygon ID field. MODEL DEMONSTRATION For our study areas, we propose to simulate fire effects variables by running the FIREHARM model to generate continuous layers of fuel consumption, tree mortality and soil heating. We will demonstrate the use of FIREHARM as a tool for fire effects resource allocation with two different approaches: 1. We will edit the weather and fuel moisture file (or “event” file) for three conditions: hot and dry, average, and cool and wet, to demonstrate the difference in burn severity output maps for these environmental conditions. This part of the project is designed to show managers that, using a weather forecast, they can edit the FIREHARM event file and run the model for predicted conditions or they could use this methodology to simulate best and worst case scenarios. Output maps will aid them in making resource allocation decisions, depending on expected conditions. 2. If fire progression perimeter data is available, we will run the model for each progression, adjusting weather and moisture conditions for the day. This will demonstrate another way of using the model to assess changes in severity within a burned area. MODEL ASSESSMENT We will evaluate the extent to which model output matches our ground reference data. We will use linear regression to statistically explore the relationship between the field and modeled data. Regression model slope and intercepts, along with correlation coefficients will be used to assess the nature of the linear association between field-measured and modeled burn severity estimates. Scatter plots, charts and tables will be generated in Excel or SAS to illustrate the relationship between the variables. With these results, we will have an understanding of how the model output corresponds to fire effects for our study areas.
To gain a more universal view of model performance, we would eventually like to conduct further investigation into model dynamics. Ideas for future research include model behavior and sensitivity analyses, along with implementation of the model in nationally diverse fire and vegetation conditions. COMBINATION OF MODEL AND IMAGE APPROACHES Along with the comparisons of each burn mapping technique with field data, we will explore combinations of ΔNBR and model-derived maps to determine the utility of a “hybrid-method” burn severity map. We will perform a step-wise elimination of independent variables from linear regression models. Three models will be evaluated, one for each of the following field-derived dependent variables: fuel consumption, tree mortality, and soil heating. These models will allow us to evaluate the degree to which each of the variables (CBI, ΔNBR and FIREHARM outputs) explains the variability in the dependent variables (field measured tree mortality, fuel consumption and soil heating). We will examine the main effects and interaction effects to determine the modeled and remotely sensed grid layers that combine to make an optimal burn severity map.
STATUS
We have finished all of the field work for this project and we are currently in the GIS/image processing and model simulation phase. We will soon begin analysis, followed by report writing and journal article submission. FUNDING ORGANIZATIONS Joint Fire Sciences Program (http://jfsp.nifc.gov/)
National Fire Plan (http://www.fireplan.gov/)
PRODUCT FINISH DATE: We hope to complete the project in the summer/fall of 2008.
PRODUCTS GIS files PDF Modeling Program
| 
