Sensitivity analysis of fire behavior modeling with LIDAR-derived surfacefuel maps

Muge Mutlu *, Sorin C. Popescu, Kaiguang Zhao

Spatial Science Laboratory, Department of Ecosystem Science and Management, Texas A&M University, 1500 Research Parkway, Suite B221,

College Station, TX 77843, United States

Forest Ecology and Management 256 (2008) 289–294

A R T I C L E I N F O

Article history:

Received 20 December 2007

Received in revised form 31 March 2008

Accepted 5 April 2008

Keywords:

LIDAR

Fuel model

FARSITE

QuickBird

A B S T R A C T

Each year, wildland fires burn millions of hectares of forest worldwide. Fire managers need to provide

effective methods for mapping fire fuels accurately. Fuel distribution is very important for predicting fire

behavior. The overall aim of this project is to model fire behavior using FARSITE (Fire Area Simulator) and

investigate differences in modeling outputs using fuel model maps, which differ in accuracy, in east

Texas. This simulator model requires as input spatial data themes such as elevation, slope, aspect, surface

fuel model, and canopy cover along with separate weather and wind data. Seven fuel models, including

grass, brush, and timber models, are identified in the study area. To perform modeling sensitivity

analysis, two different fuel model maps were used, one obtained by classifying a QuickBird image and the

other obtained by classifying a LIDAR (LIght Detection and Ranging) and QuickBird fused data set. Our

previous investigations showed that LIDAR improves the accuracy of fuel mapping by at least 13%.

According to our new results, LIDAR-derived variables also provides more detailed information about

characteristics of fire. This study will show the importance of using accurate maps of fuel models derived

using new LIDAR remote sensing techniques.

� 2008 Elsevier B.V. All rights reserved.

Contents lists available at ScienceDirect

Forest Ecology and Management

journal homepage: www.elsevier .com/locate/ foreco

1. Introduction

Forest fires destroy many houses and natural resources such asplant and animal life each year. To decrease the loss of lives andproperty because of the wildfires, fire managers need to activelyevaluate fire risks. However, managing fire risks is very difficultbecause fuel hazards are always changing. Fire behavior is verysensitive to changes in land growth, fuel risks, weather and windconditions, and topography (Keane et al., 1998). Therefore, firemanagers must provide more accurate fuel model predictions(Mutlu et al., 2008). Humans are the primary modifiers of fuelsources and ignition source vectors for the propagation of fire (Pyne,1992). To reduce the threat of the wildfires, Texas fire managers needa tool that helps them assess fire risks more accurately.

Surface fuels are the greatest concern since they are majorcontributors to the intensity and spread of fires. According toAnderson (1982), grass, brush, slash, and timber are the four majorgroups of surface fuels. Thirteen surface fuel models are identifiedfor the United Stated, each varying in amount, size, and arran-gement of the fuel model (Anderson, 1982). Table 1 illustrates thedescription of each fuel model. Seven fuel models are available in

* Corresponding author. Tel.: +1 979 458 0742; fax: +1 979 862 2607.

E-mail address: mugemutlu@tamu.edu (M. Mutlu).

0378-1127/$ – see front matter � 2008 Elsevier B.V. All rights reserved.

doi:10.1016/j.foreco.2008.04.014

our study area: grass fuel models 1 and 2, brush fuel models 4, 5,and 7, and timber litter fuel models 8 and 9.

Important assessments include the potential size, rate, andintensity of a wildland fire (Keane et al., 2000). Recent advances incomputer software technology have allowed development of severalspatially explicit fire behavior simulation models, which predict thespread and intensity of fire (Andrews and Queen, 1999). Some of thissoftware can be used to predict future fire growth and computepossible parameters of wildland fires for real-time simulations(Campbell et al., 1995; Richards, 1990). An example of such softwareis FARSITE, a spatially explicit fire growth model developed byFinney (1994). FARSITE is a two-dimensional deterministic firegrowth simulation model (Finney, 1998). This software incorporatesmodels for surface fire (Rothermel, 1972), spotting (Albini, 1979),crown fire (Wagner, 1993), and fuel moisture (Nelson, 2000).FARSITE produces maps of fire growth and behavior in vector andraster format by using Huygens’ Principle (Stratton, 2004; Finney,1998). In addition, to calculate surface fire spread, FARSITEimplements the Rothermel (1972) equations (Miller and Yool,2002). Many wildland fire managers use this powerful tool tosimulate characteristics of prescribed wildfires (Finney, 1998;Grupe, 1998). FARSITE is specially designed for forest fire modeling.

In order to run FARSITE, spatial data derived from GIS(Geographic Information Systems) and/or remote sensing isrequired and should be imported into the program. These data

Table 1Description of fuel models

Fuel model Typical fuel complex

Grass and grass-dominated

1 Short grass (1 ft)

2 Timber (grass and understory)

3 Tall grass (2.5 ft)

Chaparral and shrub fields

4 Chaparral (6 ft)

5 Brush (2 ft)

6 Dominant brush, hardwood slash

7 Southern rough

Timber litter

8 Closed timber litter

9 Hardwood litter

10 Timber (litter and understory)

Slash

11 Light logging slash

12 Medium logging slash

13 Heavy logging slash

Fig. 1. The location of our study area and false color composite of a QuickBird image

(Digital Globe, Inc.).

M. Mutlu et al. / Forest Ecology and Management 256 (2008) 289–294290

layers must be reliable for all lands and ecosystems (Keane et al.,2000). The accuracy of the input data layers is very important forrealistic predictions of fire growth (Keane et al., 1998; Finney,1998). The fuel model map is the key input for the FARSITEsimulation software and many fire managers do not have the fuelmaps needed to run the software for their area. Recently, FARSITEhas been used by many fire managers all over the world (Finney,1998; Keane et al., 1998). Fallowski et al. (2005) evaluated theaccuracy and utility of imagery from the Advanced SpaceborneThermal Emission and Reflection (ASTER) radiometer satellitesensor and gradient modeling, for mapping fuel layers for firebehavior modeling with FARSITE and FlamMap (FlammabilityMapping), fire simulation software. They created the surface fuelmodels map using a classification tree based on three gradientlayers: cover type, potential vegetation type, and structural stage.The final surface fuel model layer had an overall accuracy of 0.632.Stephens (1997) used FARSITE to spatially simulate fire growth andbehavior in mixed-conifer forest and to investigate how silvicul-tural and fuel treatments affect potential fire behavior in the NorthCrane Creek watershed of Yosemite National Park. Keane et al.(2000) combined both gradient modeling and remote sensing tomap fuels spatial data layer required by FARSITE to spatially modelfire behavior on the Gila National Forest, New Mexico. They usedsampled field data to guide the classification criteria for eachcategory and to assess the value of each category to the overallclassification. They created the three vegetation spatial data layers:potential vegetation type (PVT), structural stage, and cover type(CT). All vegetation classifications were coded into Paradoxdatabase queries which use canopy cover and plant typeinformation along with other relevant site descriptions. Then,they used these three vegetation layers to map fuels and inputlayers required to run FARSITE. Stratton (2004) used FireFamilyPlus to evaluate historical weather and calculate seasonal severityand percentile reports. Then, they used this information in FARSITEand FlamMap to model pre-treatment and post-treatment effectson fire growth, spotting, fireline intensity, surface flame length andthe occurrence of crown fire. Miller and Yool (2002) evaluated thesensitivity of FARSITE to the level of detail in the fuels data, bothspatially and quantitatively, which provided land managersknowledge about the effectiveness of detailed fuels mapping inmodeling fire spread.

Satellite technology can assist in providing data for theFARSITE software (Cheuvieco, 1997). LIDAR remote sensing is anadvance technology for forestry applications over the large areas.

The airborne LIDAR has big potential for the direct measurement ofvertical forest structure and it allows researches to measure thethree-dimensional distribution of forests. More accurate andefficient estimation of canopy fuel characteristics can be obtainedusing LIDAR techniques over large areas of forests (Andersen et al.,2005). LIDAR is an active remote sensing technique that transmitslasers to an object and measures the distance between the sensorand the object sensed. This technology is useful for high-resolutiontopographic mapping and accurate measurements of surfaceelevations. Airborne LIDAR systems can be used for fire detection,location and mapping (Justice et al., 1993), for burned areaassessment, and, important to this study, for fuel mapping (Keaneet al., 1998). Multispectral image classification is the mostimportant part of digital image analysis. Mutlu et al. (2008)specifically mapped fuels for FARSITE software use and theirresults are used in this paper. The authors applied supervisedimage classification to determine which classifier is more efficientand useful for two different fuel model maps than they created.These two fuel model maps include seven fuel models shown inTable 1. The first fuel model map was obtained by classifying only ahigh-resolution QuickBird satellite image and the second one wasobtained by classifying a LIDAR and QuickBird fused data set. Theinvestigations of Mutlu et al. (2008) show that LIDAR improves theaccuracy of fuel mapping by at least 13%.

Our research is unique and differs from other studies by usingLIDAR-derived input data layers to set initial conditions for firesimulations in FARSITE. We developed all the spatial data layersincluding the fuel model map, canopy cover, DEM, slope, andaspect, required by FARSITE using LIDAR remote sensing techni-ques and multispectral data.

The main objectives of this paper are to model fire behaviorusing FARSITE and investigate differences in modeling outputsusing fuel model maps, which differ in accuracy, in east Texas. Thisstudy will show the significance of using accurate input data layers

Fig. 2. (a) The fuel map obtained by classifying a LIDAR and QuickBird fused data set and (b) the fuel model map obtained by classifying a QuickBird image.

Fig. 3. LIDAR–QuickBird data fusion.

M. Mutlu et al. / Forest Ecology and Management 256 (2008) 289–294 291

derived from LIDAR remote sensing technique into FARSITEsoftware for realistic predictions of fire growth.

2. Study area

Fire behavior was modeled for an area in east Texasnear Huntsville, centered within the rectangle defined by9582405700W–3083903600N and 9582103300W–3084401200N, coveringabout 47.15 km2. The study area contains part of the Sam HoustonNational Forest, characterized by deciduous, coniferous, mixedstands, open ground with fuels consisting of grasses and brushes.Fig. 1 represents the high-resolution (2.5 m � 2.5 m) multispectralQuickBird image, owned and operated by DigitalGlobe, of the studyarea. Yellow marks on the image illustrate the locations of fieldplots on entire study area.

3. Materials and methods

Two different fuel model maps obtained from Mutlu et al.(2008) were used to see the differences in fire growth with fuelmodel maps of different accuracies (see Fig. 2(a) and (b)).

3.1. Data

FARSITE version 4.1 was used in this study. This softwarerequires eight data layers which include Digital Elevation Model(DEM), slope, aspect, canopy cover, fuel models map, weather,wind and fuel moisture for surface fire simulations (Finney, 1995).Two different input data sets were created in this study to generatereal-time fire simulation outputs. In dataset I, all the spatial datalayers required by FARSITE software were developed using LIDARremote sensing and GIS techniques. In dataset II, spatial data layerswere obtained from different sources.

For both datasets I and II, fuel moisture data was obtained usingRAWS (Remote Automated Weather Station) maintained by theTexas Forest Service. The weather and wind information weregathered from the Texas Forest Service’s weather station which ispart of the interagency RAWS network in Huntsville, TX (http://www.fs.fed.us/raws).

3.1.1. Dataset I: LIDAR-derived dataset

LIDAR scanning data was provided by M7 Visual IntelligenceInc. in LAS format. A total of 47 flight lines were obtained over6474.9 ha area with 28 flight lines obtained from East to West and19 flight lines obtained from North to South in leaf off condition.

Based on the study of Mutlu et al. (2008), a LIDAR–QuickBird stackimage of 10 bands was created by stacking the 4 bands of theQuickBird image with 4 LIDAR height bins, one band from thecanopy cover model, and one band from the canopy cover variance.Popescu and Zhao (2008) generated the height bin approach thatwas used to create LIDAR multiband data from scanning data. Theycreated the LIDAR bins by counting the occurrence of LIDAR pointswithin each volume unit and normalizing by the total number ofpoints. Fig. 3 shows the LIDAR–QuickBird stack image. This imagewas classified with Mahalonobis distance algorithm in ENVI (ITTVisual Systems, Inc.) with an accuracy of 90%. A total of seven fuelmodels are available in the study area: grass fuel models 1 and 2,brush fuel models 4, 5, and 7, and timber litter fuel models 8 and 9,shown in Table 1.

Canopy cover, the horizontal percentage of area covered by treecrowns at the stand level, was found using methods developed byGriffin (2006). Their study developed the use of airborne lasermethods to evaluate various canopy parameters such as canopycover and Leaf Area Index (LAI). To summarize, canopy cover isestimated by using LIDAR-derived height bins and calculating thepercentage of laser canopy hits 2 m above ground.

A digital elevation model was also derived from LIDAR. By usingENVI, slope and aspect were derived from the DEM.

3.1.2. Dataset II: multispectral data and other sources

The second map, shown in Fig. 2(b), was derived from QuickBirddata at 2.5 m resolution. Based on the report from Mutlu et al.

Fig. 4. The overall study steps.

Fig. 5. Comparison of burned areas for both fuel model maps.

M. Mutlu et al. / Forest Ecology and Management 256 (2008) 289–294292

(2008), a maximum likelihood image classification technique wasused to classify the multispectral image with an accuracy of 77%.This fuel model map includes seven fuel models as well.

Hemispherical photos were used to create canopy cover dataexpressed as a percent density at point locations used for trainingthe image classifier and for accuracy assessment. Maximumlikelihood supervised image classification technique was appliedto the QuickBird multispectral image to classify the canopy coverover the entire area. Classification is achieved by determining acanopy cover percentage range. This classification was imple-mented on a 1–4 scale including 1: 0–20%, 2: 21–50%, 3: 51–80%,and 4: 81–100%.

The DEM was downloaded from the NFAMWA (2007) (http://www.fs.fed.us/fire/planning/nist/wims_web_userguide.htm). Theresolution of DEM was at 30 m resolutions, and then converted to2.5 m resolution to match it with the resolution of multispectralimagery. The slope and aspect data were derived from the DEMusing ENVI 4.2.

4. Fire simulation

FARSITE requires data input as an ASCII file format becauseASCII text files can be viewed or created with any text editor(Finney, 1995). Since all the data were in the Band Sequential (BSQ)file format, the data were saved in Leica Geosystems’ ERDASImagine (Leica Geosystems, Inc.) image processing software imagefile format using version 9.1, then converted to ARC GRID formatfor incorporation into the development and implementation of thefire behavior model.

A total of 62 plots were measured in the study area and plotcenter locations were used as ignition points with FARSITEsimulations. FARSITE was run 124 times, 62 times on the datasetwith the LIDAR-derived fuel model map and 62 times on the

dataset with the QuickBird-derived fuel map. The duration of eachsimulation was 72 h beginning at 8:00 a.m. and ending at 8:00 a.m.3 days later.

5. Processing approach

The overall study steps are shown in Fig. 4.

6. Results and discussions

Different accuracy maps provided different results dependingon fuel model on the study area, which were expected. However,the difference is much greater than we expected. The average burnarea time for fires in Texas is between 3 and 5 days (Mark Stanford,

Fig. 6. Comparison of fire perimeter results for both fuel model maps.

Table 2Fire fighting resource characteristics for LIDAR-derived fuel model map

Resource (equipment) costs

Resource Pre Cost/h Cost/72 h

Dozer $300 175 $12,600

Tractor plow $500 150 $10,800

Type I crew $500 125 $9000

Type II crew $600 175 $12,600

Engine#1 $400 75 $5400

Engine#2 $900 100 $7200

Engine#3 $600 125 $9000

Total $3800 $66,600

Table 3Fire fighting resource characteristics for QuickBird-derived fuel model map

Resource (equipment) costs

Resource Pre Cost/h Cost/72 h

Dozer $300 175 $12,600

Tractor plow $500 150 $10,800

Type I crew $500 125 $9000

Engine#1 $400 75 $5400

Engine#2 $600 125 $9000

Total $2300 $46,800

M. Mutlu et al. / Forest Ecology and Management 256 (2008) 289–294 293

personal communication, October 2006). We have decided to runthe simulation for 72 h. The comparisons of histogram for burnedarea and perimeter results are illustrated in Figs. 5 and 6,respectively, for 72 h. Based upon the fire simulation results, fuelmodel map derived from LIDAR shows larger fire growth areas thanthe other fuel model map derived from QuickBird imagery asshown in Figs. 5 and 6.

The estimated average fire growth areas from LIDAR-derivedfuel model map and QuickBird-derived fuel model map wereapproximately 174.2 ha (430.6 ac) and 121.9 ha (301.3 ac), respec-tively. Apparently, there is a considerable difference between thetwo outputs. There are some extreme situations in our results. Forinstance, while 368.4 ha (910.3 ac) were burned on the LIDAR-derived fuel model map on the ninth run, 112.1 ha (277.2 ac) wereburned on the QuickBird-derived fuel model map. On the 38 run,the burned area is 337.1 ha (833.1 ac) on the LIDAR-derived fuelmodel map and 75 ha (185.2 ac) on the QuickBird-derived fuelmodel map. The reason for the difference is because both mapsshow different fuel models on the second ignition points, FuelModel #2, grass group, on the LIDAR-derived fuel model map andFuel Model #8, timber litter group, on the QuickBird-derived fuelmodel map. The same situation can be seen in Fig. 6 for fireperimeters for both models. For example, on the ninth run, the fireperimeter is 17.5 km on the LIDAR-derived fuel model map and4 km on the QuickBird-derived fuel model map. The fire perimetersare 8.8 and 2 km on the LIDAR-derived and the QuickBird-derivedfuel model maps, respectively, on the 38 run.

Cost of fire is one of the biggest issues in wildland firemanagement. Wildfires can have significant local economic effectsboth long-term and short-term. Donovan and Rideout (2003) usedthe Cost plus Net Value Change (C + NVC) model. The C + NVCmodel has the potential to minimize wildfire cost by reducing thetotal amount of money spent on presuppression, suppression andNVC (net wild fire damages). They assume that the cost of fire per ahectare is $100. Fire fighting resources, such as crews, dozer,engine, tractor, cost approximately $3800 based on NationalWildlife Coordinating Group fireline handbook (1998). The sum ofall costs and damages is minimized. In our case, if we take a look atthe results, average burned area is 174.2 ha on the LIDAR-derivedfuel model map. This burned area will cost around $87,826 bysumming ‘‘$3800 + $66,600 + $17,426’’ based on Donovan andRideout’s calculations. In this case, the NVC is $17,426 which wascalculated by multiplying the total burned area times 100. Table 2shows the fire fighting resources needed for a 72-h burn time basedon the LIDAR-derived fuel model map. However, the averageburned area is 121.9 ha on the QuickBird-derived fuel model mapand the cost of fire is $61,290. Since the average burned area on the

QuickBird fuel model map is smaller than that on the LIDAR-derived fuel model map, we created another fire fighting resourcestable (Table 3) for a 72-h burn based on the QuickBird-derived fuelmodel map, using fewer fire fighting resources. Apparently, there isa significant difference between these two outputs.

7. Conclusions

This study indicates the influence of a more accurate fuel mapon modeling fire behavior and assessing fire risk. FARSITE wasdeveloped mainly to be used as a tool in the fire management.FARSITE will assist fire managers with the mitigation of theharmful effects of wildfire. Also, it gives the power of sound,accurate and efficient fire behavior modeling technology to forestfire fighters (Finney, 1998).

Airborne LIDAR systems can be used for fire detection, location,fuel mapping and burned area assessment (Justice et al., 1993). Inthis study, LIDAR data processing by the height bin approach isused to generate accurate estimates of surface fuels parametersand created a fuel model map. Many fire managers do not have thefuel maps needed to run the fire simulation models for their area.We were able to develop all the spatial data layers required byFARSITE using a unique remote sensing approach that fusesinnovative LIDAR-derived products and multispectral imagery.Results from this study indicate the influence of a more accuratefuel map on modeling fire behavior and assessing fire risk.According to results, LIDAR-derived products were able to assessfuel models with high accuracy and provide different fireperimeters and fire growth area. For fire mitigation purposes,we need to know both fire perimeters and fire growth areas. Firegrowth area results are helpful to determine the cost of fire. Fireperimeter results are important because they help in determiningan optimal mix of fire fighting resources needed to fight fires suchas dozer, tractor, crews, helicopter, engines, hourly cost ofoperating the resources, arrival time, etc. Using two differentdatasets, one derived from LIDAR and the other one derived fromQuickBird imagery and different data sources, provided signifi-cantly different outputs. The differences could be attributed todifferent fuel model map, canopy cover, DEM, slope, and aspect.

M. Mutlu et al. / Forest Ecology and Management 256 (2008) 289–294294

The cost of our LIDAR data were around $32,645 and theQuickBird data were obtained for an approximate cost of $3890 forthe whole study area, 4741.83 ha. The cost of first fuel model mapwhich includes LIDAR and QuickBird data is $36,535. The cost ofsecond fuel model map which includes only QuickBird data is$3890. Based on this, LIDAR-derived fuel model map is moreexpensive than QuickBird-derived fuel models map. However, ifwe look at the results LIDAR data will save us thousands of dollars.The cost of LIDAR data for the entire study area is a lot less than thecost of average burned areas.

Accurate estimation of fire growth area and the direction of firegrowth is extremely important information for the fire manage-ment process. With the knowledge of this essential informationwill avoid any health risk for local people who live in that vicinity.This information will be more useful if it is used by firemanagement authorities. In case of fire, if the fire managers usethe QuickBird-derived fuel model map they may not be able tosend enough sources to fight with the fire and this situation maycause more serious problems. Small errors in fuel models may notbe significant for a small area; however, for large areas, small errorsmay change the result of fire behavior and fire sizes significantly.

Acknowledgements

This project was founded by the Texas Forest Service (award #02-DG-11083148-050). We want to thank all Texas Forest Servicepersonnel especially Curt Stripling for his help on collecting andanalyzing the field data and determining fuel models based onthese datasets. We thank Tom Spencer for determining fuel modelsfor our study area and thank Alicia M.R. Griffin for providing us acanopy cover data needed to run the FARSITE software and Dr.Lewis Ntaimo for helping us with the cost analysis. We also want tothank Texas A&M University Department of Ecosystem Science andManagement and the Spatial Science Laboratory for their support.

References

Albini, F.A., 1979. Spot fire distance from burning trees: a predictive model. GeneralTechnical Report INT-56. USDA Forest Service, Intermountain Forest and RangeExperiment Station, Ogden, UT, p. 73.

Andersen, E.H., McGaughey, J.R., Reutebuch, E.S., 2005. Estimating forest canopyfuel parameters using LIDAR data. Remote Sensing Environ. 94 (4), 441–449.

Anderson, H.E., 1982. Aids to determining fuel models for estimating fire behavior,United States Department of Agriculture, Forest Science, General TehcnicalReport INT-122.

Andrews, Queen, 1999. Fire modeling and information system technology. Int. J.Wildland Fire 10, 343–352.

Campbell, J., Weinstein, D., Finney, M., 1995. Forest fire behavior modeling inte-grating GIS and BEHAVE. In: Thompson, J.E. (Compiler): Analysis in Support ofEcosystem Management. USDA Forest Service Ecosystem Management CenterReport. Washington, DC, pp. 184–192.

Cheuvieco E., 1997. A review of remote sensing methods for the study of largewildland fires. Megafires Project ENV-CT96-0256. Spain, p. 192.

Donovan, G.H., Rideout, D.B., 2003. An integer programming model to optimizeresource allocation for wildfire containment. Forest Sci. 2, 49.

Fallowski, M.J., Gessler, P.E., Morgan, P., Hudak, A.T., Smith, A.M.S., 2005. Char-acterizing and mapping forest fire fuels using Aster imagery and gradientmodeling. Forest Ecol. Manage. 217 (2–3), 129–146.

Finney, M.A., 1994. Modeling the spread and behavior of prescribed natural fires. In:Cohen, J.D., Saveland, J.M., Wade, D.D. (Eds.), Proceedings of the 12th Conferenceon Fire and Forest Meteorology. Society of American Foresters, Bethesda, MD,pp. 138–143.

Finney, M.A., 1995. FARSITE Fire Area Simulator Version 1.0 User guide andtechnical documentation. Systems for Environmental Management Report.Systems for Environmental Management, Missoula, MT, p. 76.

Finney, M.A., 1998. FARSITE: Fire Area Simulator Model development and evalua-tion. Research Paper RMRS-RP-4. USDA Forest Service, Rocky MountainResearch Station, Ogden, UT, p. 47.

Griffin, A.M.R., 2006. Using LIDAR and normalized difference vegetation index toremotely determine LAI and percent canopy cover at varying scales. Master ofScience Thesis, Texas A&M University, College Station, TX.

Grupe, M.A. 1998. Assessing the applicability of the Terrestrial Ecosystem Survey forFARSITE. Master’s Thesis. University of New Mexico, Albuquerque, New Mexico,p. 95.

Justice, C.O., Malingreau, J.P., Setzer, A.W., 1993. Satellite remote sensing of fires:potentials and limitations. In: Crutzen, P.J., Goldammer, J.G. (Eds.), Fire in theEnvironment: the Ecological, Atmospheric, and Climatic Importance of Vegeta-tion Fires. Willey, pp. 77–88.

Keane, R.E., Garner, J.L., Schmidt, K.M., Long, D.G., Menakis, J.P., Finney, M.A. 1998.Development of the input data layers for the FARSITE fire growth model for theSelway–Bitterroot wilderness complex, USA. USDA Forest Service GeneralTechnical Report RMRS-GTR-3, p. 121.

Keane, R.E., Mincemoyer, S.A., Schmidt, K.M., Long, D.G., Garner, J.L. 2000. Mappingvegetation and fuels for fire management on the Gila National Forest Complex,New Mexico. Gen. Tech. Rep. RMRS-GTR-46-CD. U.S. Department of Agriculture,Forest Service, Rocky Mountain Research Station, Ogden, UT, p. 126 (CD-ROM).

Miller, J.D., Yool, S.R., 2002. Modeling fire in semi-desert grassland/oak woodland:the spatial implication. Ecol. Model. 153 (3), 229–245.

Mutlu, M., Popescu, C.S., Stripling, C., Spencer, T., 2008. Mapping surface fuel modelsusing lidar and multispectral data fusion for fire behavior. Rem. Sens. Environ.112 (1), 274–285.

Nelson Jr., R.M., 2000. Prediction of diurnal change in 10-h fuel stick moisturecontent. Can. J. Forest Res. 30, 1071–1087.

National Fire and Aviation Management Web Applications (NFAMWA), 2007.http://www.fs.fed.us/fire/planning/nist/wims_web_userguide.htm. Accessedon October 3, 2007.

Popescu, C.S., Zhao, K., 2008. A voxel-based lidar method for estimating crown baseheight for deciduous and pine trees. Rem. Sens. Environ. 112 (3), 767–781.

Pyne, S.J., 1992. Fire in America: A Cultural History of Wildland and Rural Fire.Princeton University Press, 654 pp.

Richards, G.D., 1990. An elliptical growth model of forest fire fronts and itsnumerical solution. Int. J. Numer. Meth. Eng. 30, 1163–1179.

Remote Automated Weather Station (RAWS, 2007). http://www.fs.fed.us/raws.Accessed on December 5, 2007.

Rothermel, R.C., 1972. A mathematical model for predicting fire spread in wildlandfuels. Research Paper INT-115. USDA Forest Service, Intermountain Forest andRange Experiment Station, Ogden, UT.

Stephens, S.L., 1997. Evaluation of the effects of sivilcultural and fuels treatments onpotential fire behavior in Sierra Nevada mixed-conifer forests. Forest Ecol.Manage. 105, 21–35.

Stratton, D.R., 2004. Assessing the effectiveness of landscape fuel treatments on firegrowth and behavior. Forestry (October/November), 32–40.

Wagner, W., 1993. Prediction of crown fire behavior in two stands of jack pine. Can.J. Forest Res. 23, 442–449.