Remote Sensing of Environment 112 (2008) 3773–3783

Contents lists available at ScienceDirect

Remote Sensing of Environment

j ourna l homepage: www.e lsev ie r.com/ locate / rse

Using multiple endmember spectral mixture analysis to retrieve subpixel fireproperties from MODIS

Ted C. Eckmann ⁎, Dar A. Roberts, Christopher J. StillGeography Department and Institute for Computational Earth System Science, University of California at Santa Barbara, Santa Barbara, CA 93106-4060, United States of America

⁎ Corresponding author.E-mail addresses: ted@geog.ucsb.edu, ted.eckmann@

0034-4257/$ – see front matter © 2008 Elsevier Inc. Aldoi:10.1016/j.rse.2008.05.008

A B S T R A C T

A R T I C L E I N F O

Article history:

The Moderate-Resolution Im Received 28 February 2008Received in revised form 20 May 2008Accepted 24 May 2008

Keywords:FireSubpixelSizeMESMAFRPMODISASTER

aging Spectroradiometer (MODIS) sensors on NASA's Terra and Aqua satellitesimage most of the Earth multiple times each day, providing useful data on fires that cannot be practicallyacquired using other means. Unfortunately, current fire products from MODIS and other sensors leave largeuncertainties in measurements of fire sizes and temperatures, which strongly influence how fires spread, theamount and chemistry of their gas and aerosol emissions, and their impacts on ecosystems. In this study, weuse multiple endmember spectral mixture analysis (MESMA) to retrieve subpixel fire sizes and temperaturesfromMODIS, possibly overcoming some limitations of existing methods for characterizing fire intensities suchas estimating the fire radiative power (FRP). MESMA is evaluated using data from the Advanced SpaceborneThermal Emission and Reflection Radiometer (ASTER) to assess the performance of FRP and MESMA retrievalsof fire properties from a simultaneously acquired MODIS image, for a complex of fires in Ukraine from August21, 2002. TheMESMA retrievals of fire size described in this paper showa slightly stronger correlation than FRPdoes to fire pixel counts from the coincident ASTER image. Prior to this work, few studies, if any, had usedMESMA for retrieving fire properties from a broad-band sensor like MODIS, or compared MESMA to higher-resolutionfire data or othermeasures offire properties like FRP. In the future,MESMA retrievals could be usefulfor fire spread modeling and forecasting, reducing hazards that fires pose to property and health, andenhancing scientific understanding of fires and their effects on ecosystems and atmospheric composition.

© 2008 Elsevier Inc. All rights reserved.

1. Introduction

Fires are major sources of trace gas and aerosol emissions (e.g.Andreae &Merlet 2001), ecosystemdisturbance (e.g. Pyne et al., 1996),and land-cover change (e.g. Cochrane et al., 1999) at local, regional,and global scales. Fires can also pose hazards to property and health,and forecasts of fire spreading are often very poor (Bianchini et al.,2005). Furthermore, the roles that fires play in global climate have notbeen fully quantified (e.g. Hoelzemann et al., 2004), and climatechange is likely altering these roles. This study therefore seeks toimprove monitoring and understanding of fires and their impacts.

The Moderate-Resolution Imaging Spectroradiometer (MODIS)sensors on NASA's Terra and Aqua satellites can provide data to helpaddress these issues by imaging much of the Earth's surface multipletimes each day, providing fire data that cannot be practically acquiredusing in situ measurements or other sensors. MODIS data currentlyenable the creation of active fire products that support wildfiremanagement, and frequently offer multiple snapshots of a fire'sprogression over time, which can provide more detailed measurementsof fire behavior than are available from post-fire burned area maps. Forexample, Peterson et al. (2005) used pixel-level fire products from 11

gmail.com (T.C. Eckmann).

l rights reserved.

separate MODIS overpasses, collected over 58 h of a fire's lifespan, toinitialize and validate the widely-used Fire Area Simulator (FARSITE)model, which is described in Finney (1998). MODIS fire products have anominal spatial resolution of 1 km, which is very coarse relative to thesizes of typical wildfire features (e.g. Morisette et al., 2005), but theseMODIS products are often the best data available for monitoring andmodeling many fires. This is because in many areas of the world, firesthat are detected by MODIS are reported inconsistently or not reportedat all by suborbital or on-the-ground surveys, and other satellite sensorscurrently or previously used for monitoring fires are in somewaysmorelimited than MODIS in spatial or temporal coverage, revisit frequency,geolocational accuracy, false-alarm rates, or sensitivity to smallfires (e.g.Csiszar et al., 2005, Giglio et al., 2006).

MODIS fire detection algorithms have been validated at the pixellevel (e.g.Morisette et al., 2005) butMODISfire pixels are actuallymixedpixels that usually contain unburned areas, along with smaller flaming,smoldering, and burned components. Pixel-level MODIS fire productsthus leave substantial uncertainties about any given fire's overall size,and the sizes and temperatures of its flaming, smoldering, and burnedcomponents. Kaufman et al. (1998), Wooster et al. (2003), and Robertset al. (2005) describedmethods for estimating the total radiant power ofa firewithin aMODIS pixel, but this quantity cannot separate a fire's sizeand temperature: a small hot fire can have the same radiant power as alarger, cooler fire. Because a fire's size and its temperature have different

3774 T.C. Eckmann et al. / Remote Sensing of Environment 112 (2008) 3773–3783

influences on the amount and chemistry of its trace gas and aerosolemissions (Andreae & Merlet, 2001), ecosystem impact (e.g. Hanley &Fenner, 1998), and fire spreading behavior (Pyne et al., 1996), estimatesof a fire's radiant power (FRP) may not provide enough information forprecisely quantifying some aspects of fires and their effects.

While not designed specifically for fires, a technique for subpixelunmixing developed by Roberts et al. (1998), called multiple end-member spectral mixture analysis (MESMA), may be able to overcomethese and other limitations of existing fire products. MESMA usesspectral shape to unmix each pixel as a combination of subpixelfeatures, called endmembers. In this paper, we apply MESMA to aMODIS image of fires in Ukraine by unmixing each fire pixel using alibrary of pre-generatedfire endmembers of various temperatures, andbackground endmembers (spectra selected fromnon-burning pixels inthe same image), thus estimating the subpixel sizes and temperaturesof fires in each pixel. We also calculate FRP and compare MESMA firearea and FRP to measures of fire size from a coincident higher-resolution image from the Advanced Spaceborne Thermal Emissionand Reflection Radiometer (ASTER) to assess the relative performanceof FRP and MESMA for this Ukraine image. The Ukraine image waschosen because it contains several fires of varying sizes, and because itrepresents a geographical area that has not been investigatedpreviously in similar studies. Despite the finer spatial resolution ofASTER,measures of subpixel fire properties fromMODIS are still usefulbecauseASTERhas a narrower swath and amuch longer return intervalthan MODIS, so MODIS can image a larger number of fires and providemore frequent images of an area's fires. MODIS is thus generally muchmore useful than ASTER for monitoring fires in near real-time and forcomprehensive global fire climatologies.

2. Background

Planck's equation describes the spectral emitted radiance from ablackbody where T is the object's kinetic temperature in Kelvin, h isPlanck's constant (6.63×10−34 J s), c is the speedof light (3.00×108ms−1),λ is wavelength in meters, Lλ is emitted radiance at wavelength λ, and kis Boltzmann's constant (1.38×10−23 J K−1):

Lλ ¼ 2hc2

λ5 ehckλT−1

� � ð1Þ

In the case of remote sensing, measured radiance Lλ for a pixelusually comprises the radiance emitted by several different objects(a “mixed” pixel) that may all have different temperatures and spec-tral emissivities, instead of a single pure pixel containing only oneobject at a single temperature. Methods that can address this andother issues related to retrieving subpixel fire properties include theDozier (1981) approach, FRP (Kaufman et al., 1998), and MESMA(Roberts et al., 1998). Each approach is described in the followingsections.

2.1. The dozier approach

Dozier (1981) developed a method for retrieving the sizes andtemperatures of subpixel objects, which has beenmodified and appliedwidely to a variety of sensors bymany other investigators. Dozier (1981)showed that as the equation below exists for each band, and given atleast two bands (usually near 3.8 µm and 10.8 µm), the resulting systemof equations canbe solved to retrieve temperatures and subpixel areas of“hot” and “background” surfaces within a single pixel:

Lλ ¼ fhotβ λ; Thotð Þ þ fbackgroundβ λ; Tbackground� � ð2Þ

where Lλ is measured radiance at wavelength λ, fhot is the fraction ofthe pixel covered by the hot surface of temperature Thot, fbackground isthe fraction of the pixel covered by the background surface of

temperature Tbackground, β(λ,T) is the Planck equation (see Eq. (1)),and fhot and fbackground sum to 1. Although others have modified thisapproach, the version described in Eq. (2) assumes that the objectsemit as blackbodies, and does not account for atmospheric influenceson at-sensor radiance. This approach generally also requires assump-tions that the pixel comprises only two objects with differenttemperatures, and that some of the unknowns can be constrained,such as estimating the background object's temperature from adjacentpixels, or assuming that two adjacent pixels have the same tempera-tures for the hot and background objects.

Many have applied modifications of the Dozier approach for esti-mating the subpixel sizes and temperaturesoffires, volcanoes, andotherhot objects, using a variety of sensors. Examples of applications to firesinclude Flannigan and VonderHaar (1986) usingNOAA's Advanced VeryHigh Resolution Radiometer (AVHRR), Prins and Menzel (1992) usingNOAA's Geostationary Operational Environmental Satellite (GOES)Visible Infrared Spin Scan Radiometer Atmospheric Sounder (VAS),and Wooster et al. (2003) and Oertel et al. (2004) with the Bi-spectralInfraRed Detection (BIRD) satellite. Matson and Dozier (1981) alsoapplied the technique to estimate the sizes and temperatures of gasflares fromoil fields and steelmills using theNOAA-6AVHRR. Lombardoet al. (2006) applied a similar method to study subpixel features of avolcano using the Digital Airborne Imaging Spectrometer (DAIS) 7915spectrometer and Landsat TM.

Applications of the Dozier approach generally require the follow-ing assumptions: 1) the hot object has a single, uniform temperature;2) the background object radiates as a blackbody; 3) atmosphericeffects are minimal. According to Giglio and Kendall (2001), theseassumptions are usually unrealistic and generally produce large errorsin retrievals of fire sizes and temperatures. The choice of wavelengthsused in Eq. (2) can also produce substantial errors in retrieved fire sizeand temperature, as Giglio and Justice (2003) demonstrated.

2.2. The fire radiative power (FRP) approach

Kaufman et al. (1998) and Justice et al. (2002) described methodsfor estimating a fire's radiative power (FRP) from MODIS, and FRPvalues calculated with these methods are included in the MOD14 “Fireand Thermal Anomalies” product, which also provides binary fire/no-fire detection data for virtually all MODIS images (Giglio et al., 2003).Wooster et al. (2003) described a similar algorithm, referring to it as anestimate of the fire's radiative energy (FRE), and compared FRE fromMODIS to FRE from BIRD for a fire imaged simultaneously by bothsensors. Likewise, Roberts et al. (2005) compared FRP from MODISwith near-simultaneous FRP from the Spinning Enhanced Visible andInfrared Imager (SEVIRI) aboard Meteosat-8. All of these approachesaim to estimate a fire's true radiative power (FRPtrue) in J s−1 (or W),which, according to Roberts et al. (2005), is:

FRPtrue ¼ eσ ∑n

i¼1AnT4

n ð3Þ

where σ is the Stefan–Boltzmann constant (5.67×10−8 J s−1 m−2 K−4), Anis the area of the nth thermal component of the fire (in m2), Tn is thetemperature of the nth thermal component (in K), and ε is the effectivemean emissivity over all emitting wavelengths (unitless). Most sensorsdo not measure radiation over all the wavelengths at which fires emit,and because most pixels that contain fires also contain non-burningcomponents, this becomes a mixed-pixel problem, so for most sensorsand situations, direct measurement of FRPtrue is not practical. However,Wooster et al. (2003) showed that over a realistic range of fire andbackground sizes, temperatures, and sensor types, there is an ap-proximately linear relationship between FRPtrue and the sensor'smeasured radiance above the adjacent non-burning background pixelsin the mid-infrared (around 4 μm). Algorithms that exploit this rela-tionship to estimate FRPtrue are generally derived using a semi-empirical

3775T.C. Eckmann et al. / Remote Sensing of Environment 112 (2008) 3773–3783

approach (seeRoberts et al., 2005), as shown in the algorithmdevelopedfor MODIS by Kaufman et al. (1998):

FRPestimated ¼ 4:34� 10�19 Wm−2 K−8 T4ð Þ8− T4bð Þ8h i

ð4Þ

where FRPestimated is the estimated rate of emitted energy from thefire pixel (in W m−2), T4 is the brightness temperature in the ~4 μmband of the pixel that contains fire, and T4b is the mean brightnesstemperature in the ~4 μmband of adjacent background pixels that donot contain fire.

2.3. The MESMA approach

Dennison et al. (2006) applied MESMA to retrieve subpixel sizesand temperatures of active fire components using Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data acquired over the 2003California Simi Fire. AVIRIS is a 224 channel hyperspectral sensor thatmeasures radiance from around 400 to 2500 nm at a spatial resolutionof approximately 4 to 20 m, depending on aircraft altitude (Greenet al., 1998). MESMA unmixes each pixel as a linear combination ofendmembers, allowing a different combination of endmembers foreach pixel (Roberts et al., 1998):

Liλ ¼ ∑N

k¼1fki � Lkλ þ eiλ and ∑

N

k¼1fki ¼ 1 ð5Þ

where a mixed pixel Liλ from location i is modeled as the sum of Nendmembers, Lkλ, each covering a fraction fki of the pixel. The residualterm εiλ describes the unmodeled portion of the radiance, and thechosen model for each pixel is the one that minimizes the root meansquared error (RMSE) over the included number of bands used inunmixing, B:

RMSE¼ ∑B

k¼1eiλð Þ2=B

� �1=2ð6Þ

This MESMA approach does not require its endmembers to beblackbodies: endmembers can come from field- or lab-measuredspectra (which would be called reference endmembers), or relativelypure pixels in an image (which are known as image endmembers).Radiative transfer models can also account for atmospheric effectseither by simulating or modifying the spectra used as endmembers.Other studies have used MESMA across a diverse range of remote-sensing applications, such as mapping snow-covered area and grainsize (Painter et al., 2003), mapping landforms at the continental scale(Ballantine et al., 2005), mapping urban land cover (Powell et al.,2007), and mapping crop stress in cotton fields from spider miteinfestations (Fitzgerald et al., 2004).

The Dennison et al. (2006) approach accounted for atmosphericeffects on measured radiance by generating an endmember libraryusing MODTRAN 4.3 (Berk et al., 1989), which simulates expected at-sensor radiance for a given surface temperature, view and solargeometry, and a modeled atmosphere. In this implementation,MODTRAN assumes the fire components are emitting as blackbodiesfrom the surface, but in theory, spectra measured from actual fires (orother hot objects) over a range of known temperatures could be usedinstead of the blackbody curves generated by the Planck equation. Asdiscussed by Giglio and Kendall (2001), errors may be introduced byassuming that fires emit as blackbodies, as such an assumption mayonly be valid when the path length through the flames from thesensor's view is approximately 6 m or longer, with shorter pathlengths generally producing lower fire emissivities, and thus greatererrors in retrieved fire sizes and temperatures. Although the effectsmay be different for MESMA, the sensitivity analysis of the Dozierapproach in Giglio and Kendall (2001) found the blackbody assump-tion produced retrieved temperature errors of under ~20 K, but fire

size errors of up to 60% of the pixel's size, after testing a range of firetemperatures with a path length through the flames as low as ~1 m.

The Dennison et al. (2006) study that applied MESMA to an AVIRISfire image appeared successful, but only a few active fires have beenimaged by AVIRIS, and sampling large numbers of fires using AVIRISwould be impractical because it is an airborne sensor with verylimited spatial coverage. In contrast, an enormous number of MODISimages of active fires have been collected from around the globe, andMODIS data can provide what are in many ways the most com-prehensive global climatologies of seasonal fire patterns currentlyavailable (e.g. Giglio et al., 2006). As a result, developing methods forapplying MESMA to measure fire properties from MODIS couldprovide new contributions to many studies worldwide.

2.4. Limitations of the Dozier, FRP, and MESMA approaches

The Dozier, FRP, and MESMA approaches are all subject to thelimitations of any sensor towhich they are applied. For example, MODISis most sensitive to radiance coming from the center of each pixel'sfootprint, and this sensitivity decreases gradually towards the edges ofeach pixel in the across-track direction (e.g. Kaufman et al., 1998). This iscommonly called the “triangular response” of MODIS, as the spatialresponse of each pixel is approximately triangle-shaped in the across-track direction, although this phenomenon does not affect the along-track sensitivity of MODIS. Each MODIS thermal pixel has a nominalfootprint at nadir of 1 km by 1 km, but due to the sensor's triangularresponse, the area to which each MODIS pixel is sensitive at nadir isactually around 2 km in the across-track direction, by 1 km in the along-track direction (e.g. Morisette et al., 2005). Thus, MODIS measures ofradiance from any given fire, and thus estimates of that fire's propertiesusing the Dozier, FRP, or MESMA approaches, will vary based uponwhere that fire is located within the MODIS pixel's footprint.

Other limitations also exist for the Dozier, FRP, and MESMAapproaches. The primary weakness of FRP is that it is not truly aretrieval of subpixel fire sizes or temperatures because it cannotseparate a fire's size and temperature: a small hot fire can have thesame radiative power as a larger, cooler fire. In fact, for a given value ofFRP, an infinite number of possible combinations of the fire's size andtemperature exist that could produce that FRP value. Thus, FRPmay beof limited utility for many applications because the sizes andtemperatures of fires have such strong and separate influences onthe amount and chemistry of their trace gas and aerosol emissions,and their effects on ecosystems. For example, Hanley and Fenner(1998) demonstrated how various soil temperatures, representingdifferent fire intensities, can influence whether and the extent towhich the seeds of some species are better suited than others tosurvive, germinate, and grow rapidly following the fire. They alsoshowed how the amount of time for which seeds are exposed to thesetemperatures can influence post-fire growth rates, which can alsohave different influences across species. Thus, the sizes, temperatures,and spreading rates of fires can all have different and complexinfluences on the rates of post-fire plant succession and post-firespecies composition, many of which may not be discernible using theFRP approach.

FRP also may not be adequate for fully understanding the trace gasand aerosol emissions of fires. Many studies (e.g. Schultz 2002, van derWerf et al., 2004, Randerson et al., 2005) have used emission factors toestimate the amounts of various chemical species emitted by firesglobally, with considerable focus on gases like CO2 and CH4 that canexert significant influences on climate. These emissions factors provideestimates of the amount of individual chemical species released permass of biomass burned, which contain significant uncertaintiesbecause fire emission factors vary considerably as a function of fueltype, moisture, and structure, wind, terrain, movement of the fire'sflaming front, the fire's size, temperature, and other factors (Andreae &Merlet 2001). Unfortunately, the data products currently used in these

3776 T.C. Eckmann et al. / Remote Sensing of Environment 112 (2008) 3773–3783

emission-factor-based studies do not fully account for variations inmany of these variables, especially fire sizes and temperatures, whereagain, FRP may be of limited utility. Also, FRP may not be portable tomany sensors. Several sensors thatmayotherwise beuseful inprovidingFRP estimates at fine spatial resolutions, such as AVIRIS and ASTER,simply lack bands near 4 µm.

The Dozier, FRP, and MESMA approaches assume linear mixing,and a linear response from the sensor, but these assumptions may notalways hold (e.g. Collins et al., 2001). However, this is not an issue fortheMESMA approaches presented in Dennison et al. (2006) and in thispaper, which address this by using a separate endmember for eachtemperature. The MESMA and Dozier approaches also assume thateach band in a sensor is measuring the same area, but this is notalways the case. Shephard and Kennelly (2003) used a two-channelalgorithm, based on the Dozier approach, to show that band-to-bandcoregistration errors of only 10% with 1-km pixels and a triangularspatial response function (properties representative of MODIS) canproduce substantial errors in estimations of subpixel fire sizes andtemperatures. The effects of these errors vary with factors such as thefire's size relative to the sensor's resolution, and may be different forthe MESMA approach because it uses more channels, but FRP is notaffected by band-to-band coregistration errors because it uses onlyone channel. Uncertainties from band-to-band coregistration errors,and from the triangular response of MODIS, should produce randomerrors, not biases, in the fire sizes and temperatures retrieved byMESMA. Thus, these errors may be significant for a single pixel, butcould essentially be cancelled out in calculating means for samples ofmany pixels, such as would likely be used in studies of fire emissionsover large areas or long periods. Pixel saturation is another severelimitation for most sensors and most bands, although sensors such asMODIS and BIRD have bands designed specifically for remainingunsaturated even for high-temperature targets such as fires (Giglio &Kendall 2001, Wooster et al., 2003). The effectiveness of theseapproaches can also be limited by other factors such as sensor noise,imperfect atmospheric correction, or obscuration by clouds or ground-cover: if a sensor's view of a surface fire is partially obstructed by trees,for example, its size or FRP could be underestimated. Many of theseerror sources should be small relative to the scales of fires detectableby MODIS, especially for large, hot fires, but future studies shouldquantify the errors associated with each of these factors.

Despite these limitations, the Dozier, FRP, and MESMA approacheshave significant, and yet largely unrealized potential to improve ourunderstanding of fires. For example, Peterson et al. (2005) foundFARSITE performance was very sensitive to the size of the ignitionpolygon created from the existing MODIS pixel-level fire product. Thissuggests that fire-spread modeling may benefit greatly from theimproved initialization and validation data that could be madeavailable from subpixel retrievals of fire sizes from MODIS images.Trigg and Roy (2007) described results of focus group interviews aboutthe usefulness of current MODIS fire products with resourcemanagers, many of whom emphasized their need for better spatialresolution and information on fires smaller than the limits of what thecurrent MODIS active fire product provides, along with information onfire intensities, temperatures, and rate of spread. Thus, significantdemand exists for the data products that could be generated using theDozier, FRP, and MESMA approaches, but more work is needed inquantifying, comparing, and overcoming their uncertainties, and inapplying these approaches to expand the types of available active-firedata. In an attempt to contribute towards these goals, we describe theuse of MESMA to retrieve fire properties from MODIS and compareMESMA results to the FRP. We evaluate the performance of bothMESMA and FRP using coincident ASTER data and a threshold appliedto ASTER's band 9 to flag burning pixels. Our study does not calculateresults for the Dozier approach, primarily because MESMA theoreti-cally incorporates all the aspects of the Dozier approach but withsignificant advantages such as not needing to assume that the sub-

pixel objects radiate as blackbodies, and allowing the use of morebands at a variety of wavelengths. However, a follow-on study willinclude calculations using the Dozier approach in its analysis.

3. Methods

3.1. Validation data

Ideally, on-the-groundmeasurementswould be used for validatingretrievals of subpixel fire sizes and temperatures, but the authors arenot aware of any suitable on-the-ground data for the fires presented inthis study. This lack of suitable on-the-ground data is unfortunatelycommon, as the danger and difficulty of obtaining in situ active firemeasurements, and the rapid rates of fire spreading, explain whyremote sensing studies of active fires must typically rely on other,simultaneously acquired, higher-resolution remotely sensed data forvalidation. For example, Morisette et al. (2005) used pixels identifiedas containing fire by band 9 from Terra's ASTER, which is centered at2.43 µm and has a nominal spatial resolution of 30m, to validate pixel-level fire detection products fromMODIS-Terra that covered the sameareas simultaneously. Our study followed a similar approach, usingASTER to assess the performance of FRP and MESMA retrievals from acoincident MODIS image (Fig. 1). Morisette et al. (2005) found band 9to be the most useful of ASTER's bands for discriminating betweenburning and non-burning pixels. ASTER's band 9 was used in thispaper for the same reasons.

Morisette et al. (2005) counted ASTER pixels with a radiance of6.33 W m−2 μm−1 sr−1 or higher in ASTER's band 9 as containing fire,using only images from Africa. Manual inspection of the UkraineASTER image used in this study, along with ASTER images of fires inother areas, revealed that this threshold misses many obvious firepixels. For example, Fig. 2 shows a portion of a nighttime ASTER imagethat contains a fire in California, USA, and demonstrates that manyASTER pixels that contain fire have radiances well below theMorisetteet al. (2005) threshold of 6.33Wm−2 μm−1 sr−1. A nighttime fire scenewas used for this demonstration because all the pixels within it thathave radiances significantly above zero in ASTER's SWIR bands likelycontain fire, whereas reflected solar radiation can also produce sig-nificant radiances in ASTER's SWIR bands in daytime scenes, whichwould thus make the demonstration more complicated for a daytimescene such as the one in Fig. 1. The ideal threshold, which wouldmaximize sensitivity to fire while minimizing confusion with non-burning background land-cover types, will vary from scene to scenebased upon background brightness levels and conditions. Manualinspection determined that an appropriate threshold for the UkraineASTER image used in this study would be to count pixels withradiances of 2.00 W m−2 μm−1 sr−1 or higher in ASTER's band 9 ascontaining fire. Pixels with band 9 radiances above this threshold, butwithout radiances significantly above those of surrounding areas inASTER's thermal infrared bands or other visual indications of burningsuch as smoke plumes in ASTER's visible bands, were manuallyidentified as being highly reflective but non-burning land-cover typesand not counted as fires, as in Morisette et al. (2005). However, thishad little influence on the results presented in this paper because allareas in our analysis that were above the band 9 threshold also hadother indicators to support their classification as burning and werethus counted as ASTER fire pixels. Artifacts produced by the ASTERsensor, such as “blooms” and “spikes” of increased radiances aroundbright objects, and pixels with zero or near-zero values near saturatedpixels (which are described in Morisette et al., 2005) appear to bepresent in the scene used in this study (Fig. 1), but these artifacts aresmall enough that their influence on the results presented in thispaper should be minor.

Due to the triangular response of MODIS, Morisette et al. (2005)treated the actual ground area sampled by each MODIS fire pixel atnadir as being 2 km in the along-scan direction by 1 km in the along-

Fig. 2. Radiances fromASTER's band 9, showing a portion of a nighttime fire in Californiafrom September 29, 2006 at 06:05:40 UTC. Pixels above the Morisette et al. (2005)threshold of 6.33 W m−2 μm−1 sr−1 are indicated, but all pixels with radiances sig-nificantly above zero likely contain fires because this is a nighttime SWIR image.

Fig.1. The entire ASTER scene used in this study, which was acquired on August 21, 2002 at 09:26:23 UTC, and its positionwithin the correspondingMODIS image (orange box in insetmap), which covered the ASTER scene's area at almost exactly the same time. Red boxes outline the footprints of all the MODIS pixels that fell completely within this ASTER scene andwere flagged by the corresponding MOD14 collection 5 product as containing fire. White arrows point to areas that contain fire but were missed by the MOD14 collection 5 product.

3777T.C. Eckmann et al. / Remote Sensing of Environment 112 (2008) 3773–3783

track direction. The area corresponding to this assumed footprint isapproximately 66 by 33 ASTER band 9 pixels. Because both sensors areon the same platform, any point on the ground within the areacovered by both ASTER and MODIS will have approximately the samesensor zenith angle for both sensors. Thus, this ratio of 1 MODIS pixelto a 66-by-33-pixel polygon of ASTER band 9 pixels will remainapproximately the same throughout the area covered by both sensors.The triangular response of MODIS also causes its sensitivity todecrease towards the along-scan edge of each pixel, but in order tovalidate the MODIS fire product, which is ultimately affected by thetriangular response, Morisette et al. (2005) did not attempt to “factorout” this effect. Therefore, in this paper comparing measures of theUkraine fires from the coincident MODIS and ASTER images, all ASTERband 9 pixels within this 66-by-33-pixel area were counted equally.Likewise, geolocation error for MODIS pixels could account for somediscrepancies between MODIS and ASTER data in this kind ofcomparison, but those geolocation errors are also an inevitable part

Table 1Settings used for generating fire endmembers with MODTRAN 4.3.1, taken primarilyfrom the datasets listed in Table 2

Parameter Name Value

Date and time August 21, 2002 at 09:26:23 UTCTarget latitude & longitude 51.56° N, 26.60° EAtmosphere Mid-latitude summerPrecipitable water vapor 2.526 cm (total-column)Extinction model Rural extinction, 5 km visibilityColumn carbon dioxide 385 ppm (from climatology)Spectral albedo 0Ground altitude 0.139 kmSensor altitude 705 kmSensor zenith angle 0° off nadirSolar zenith angle 40.78°

Fig. 3. Examples of the fire endmembers generated using MODTRAN to simulate at-sensor radiance for Terra's view of the ASTER scene's area in Fig. 1, for various surfacetemperatures (in K), before and after resampling to the spectral responses of eachMODIS band. Endmembers were created using the settings in Table 1, for every 10 K stepfrom 500 to 1500 K, but only some of these are graphed here for clarity.

3778 T.C. Eckmann et al. / Remote Sensing of Environment 112 (2008) 3773–3783

of any MODIS fire product, or any MODIS-to-ASTER comparison, so noattemptwasmade to compensate for these effects. According toWolfeet al. (2002), geolocation error for the 1 km MODIS bands is ~73 m(1σ). Teshima and Iwasaki (2008) showed that geolocation error forASTER's band 9 is often less than 1 pixel (b30 m).

3.2. MESMA retrievals

For the MODIS image used in this study, MODTRAN 4.3.1 (Berket al., 1989) was used to create a fire endmember at every 10 K stepfrom 500 to 1500 K, a temperature range chosen to match the oneused in Dennison et al. (2006) primarily to facilitate comparisons withthat study, and because this range spans the temperatures typical ofbiomass-burning fires (Pyne et al., 1996). In this approach, MODTRANassumes the fires are emitting as blackbodies from the surface, andadjusts this to the expected at-sensor radiance for each of thesesimulated fire temperatures given the atmospheric conditions andsolar/sensor angles, also following Dennison et al. (2006).

Table 1 lists the settings used in MODTRAN for generating the fireendmembers used in this study, which are median values for theportion of the MODIS granule covered by the ASTER scene in Fig. 1. Allof these inputs could be available operationally in near real-time,primarily from the MODIS products listed in Table 2, which areautomatically generated for virtually every MODIS granule. Forexample, the MOD03 MODIS geolocation product supplies sensorzenith angle, solar zenith angle, and ground altitude as “height abovegeoid,” and the MOD07 product supplies total-column precipitablewater vapor, using the thermal-infrared method which works in day-time and nighttime (King et al., 2003). These at-sensor fire end-

Table 2Filenames and all the bands/data used from each file in this study, which were obtained froGeological Survey (USGS) Center for Earth Resources Observation and Science (EROS), and f

Filename: Bands/data used:

MOD021KM.A2002233.0925.005.⁎.hdf MODIS swath-perspective radianc5 (resampled to 1 km from 500 m6 (resampled to 1 km from 500 m7 (resampled to 1 km from 500 m21 (collected at 1 km)29 (collected at 1 km)31 (collected at 1 km)32 (collected at 1 km)

MOD03.A2002233.0925.005.⁎.hdf MODIS geolocation data, sensor zenMOD07_L2.A2002233.0925.005.⁎.hdf Total-column precipitable water vMOD14.A2002233.0925.005.⁎.hdf Locations of MODIS pixels flaggedAST_L1A_00308212002092623_⁎.hdf Swath-perspective radiance at senso

All of these files were from the most recent collection versions available to the public at the tcould vary because some of these products may be regenerated when ordered.

members were then resampled separately to the spectral responses ofeach MODIS band (Fig. 3).

For the entire ground area covered by the ASTER scene in Fig. 1, andstarting at the ASTER scene's center, every 5th MODIS pixel in thealong-scan and along-track direction was selected as a potentialbackground endmember, but only pixels from this pool that did not

m the Land Processes Distributed Active Archive Center (LP DAAC), located at the U.S.rom NASA's Level 1 and Atmosphere Archive and Distribution System (LAADS)

e at sensor, using these bands at 1 km: Center wavelength) 1.24 µm) 1.64 µm) 2.13 µm

3.96 µm8.55 µm11.03 µm12.02 µm

ith angle, solar zenith angle, ground altitude as “height above geoid”apor from the MODIS thermal-infrared methodas containing fire, FRPr from ASTER's band 9 (centered at 2.43 µm), ASTER geolocation data

ime of this study. The ⁎ in each filename represents the processing date and time, which

Fig. 5. The “shade” endmember used in this study, which was generated by MODTRANusing the settings in Table 1 and a surface temperature of 10 K.

Fig. 4. Mean and standard deviation for the 168 background endmembers used forunmixing in this study.

3779T.C. Eckmann et al. / Remote Sensing of Environment 112 (2008) 3773–3783

contain fire according to the MOD14 collection 5 product were used asbackground endmembers. This produced a library of 168 backgroundendmembers (after 1 pixel had been removed because it had beenflagged as containing fire by the MOD14 collection 5 algorithm). Fig. 4displays the mean values and standard deviations for the 168 back-ground endmembers used for unmixing in this study.

MODIS bands near major water vapor, ozone, and carbon dioxideabsorption features were excluded from unmixing, along with bandsshorter than 1200 nm because at those wavelengths, smoke mayinfluence unmixing results (Dennison et al., 2006). All bands withsaturated pixels within the ground area covered by the ASTER scenewere also excluded from unmixing, using the on-orbit measuredsaturation values for MODIS from Salomonson et al. (2000). Thus, onlyMODIS bands 5, 6, 7, 21, 29, 31, and 32were used for unmixing, and theproperties of these bands are listed in Table 2.

Spectral mixture analysis studies working in parts of the spectrumdominated by reflected solar radiation commonly include a “shade”endmember to represent the portion of a pixel receiving no directinsolation (e.g. Roberts et al., 1993). An endmember generated usingthe MODTRAN settings in Table 1 for a 10 K surface temperature wasused as the “shade” endmember in this study (Fig. 5), because itrepresents themodeled atmospheric contribution to a pixel's radiance,with basically no radiance from the surface, which is analogous to therole of a “shade” endmember in many other spectral mixture analysisstudies. Each MODIS pixel flagged as containing fire by the MOD14collection 5 product within the area covered by the ASTER image inFig. 1 was thus unmixed, allowing one fire endmember, one back-ground endmember, and one “shade” endmember per pixel. TheseMESMA results were then compared with ASTER fire pixel counts, andwith FRP values from the MOD14 collection 5 product.

4. Results and discussion

Within the area covered by the ASTER image from Fig. 1, 19 MODISpixels were flagged as containing fire by the MOD14 collection 5product. Each of these 19 MODIS pixels contained at least one ASTERpixel above the fire threshold of 2.00 W m−2 μm−1 sr−1 within itsassumed footprint of 33 by 66 ASTER band 9 pixels. However, severalareas covered by theASTER scene that containedfireweremissed by theMOD14 collection 5 product (Fig. 1), which is consistent with thefindings of Morisette et al. (2005) and other similar studies: theMOD14fire detection algorithm produces occasional errors of omission, anderrors of commission as well. Table 3 lists results for the MESMAretrievals of fire size described in this paper, and other details for each ofthese 19 pixels. Only two of these 19MODIS pixels containedmore than1% fire, according to MESMA retrievals, which is consistent with ourexpectations from visual inspection of the ASTER scene in Fig. 1: MODIS

pixel footprints are very large relative to the subpixel sizes ofmany fires.This illustrates the importance of subpixel measures of fire propertiesfrom MODIS, because many MODIS “fire pixels” contain predominatelynon-burning components. MESMA modeled each pixel as containingbetween 88% to over 99% of its selected background endmember, byarea,while all 19 of theseMODIS pixelsweremodeled as containing lessthan 10% of the shade endmember.

The small sample size of 19 points limits the significance of theconclusions that can be drawn from these data, and spatial auto-correlation among these points may, if not accounted for, causestatisticalmeasures of relationshipswithin these data to be overstated.Unfortunately, the small sample size may also invalidate any analysesof the spatial autocorrelation within the data, and any methods tocompensate for it (e.g. Fortin et al., 1989), but simple regressionanalyses using these data show some noteworthy results. FRP waspoorly correlated with theMESMA fire-size retrievals described in thispaper (R2 of 0.17; p=0.078), which demonstrates that MESMA and FRPmay provide significantly different information about the properties ofthe fires in these pixels (Fig. 6). For example, a MODIS pixel thatcontained 2.3%fire, according toMESMA, had an FRP of 60MW,while apixel that contained 2.6% fire, according to MESMA, had an FRP of20MW. One interpretation of this result would be that the first pixel hadhotter fire elements covering smaller areas than the second pixel.However, MESMAmodeled both of theseMODIS pixels as containing fireof the same temperature (500K), so this kind of discrepancy could also beproduced by radiance from the background portion of the pixel that wasattributed to the fire, or vice-versa, for either the MESMA or FRP ap-proaches. These two MODIS pixels contained 250 ASTER pixels, and 251ASTERpixels, respectively, thatwereflaggedas containingfire. Thiswouldseem to agree more closely with MESMA than FRP in this case: MESMAmodeled these two MODIS pixels as having similar sizes (2.3% fire and2.6% fire, respectively), while the FRP values for these two pixels (60MWand 20MW, respectively) would suggest that they had very different firesizes or temperatures, which is less likely based on the similar number ofASTER pixels flagged as containing fire within their footprints.

FRP shows moderate correlation (R2 of 0.32; p=0.012) to the numberof ASTER pixels above the fire-detection threshold within the 33-by-66pixel area corresponding to each MODIS pixel's assumed footprint(Fig. 7). The MESMA retrievals of fire size described in this paper show aslightly stronger correlation (R2 of 0.49; p=0.00080) to these ASTER pixelcounts (Fig. 8). While the small sample size limits the significance ofthese findings, and the quantitative measures of fit may be overstateddue to spatial autocorrelation, these results suggest that in some respects,MESMA may provide better information than FRP about fire sizes.

Although FRP andMESMAboth correspond to some of the variationin fire size detected in ASTER's band 9, both FRP and MESMA exhibit

Table 3Details for all the MODIS pixels flagged as containing fire by the MOD14 collection 5 product within the area covered by the ASTER scene in Fig. 1

From MOD14: From the MESMA retrievals described in this paper:

Location of MODIS firepixel's center

Fire radiativepower, in MW

% of pixelwith fire

Modeled firetemperature, in K

% of backgroundendmember

% of shadeendmember

RMSE, in W m−2

μm−1 sr−1

51.5372° N, 26.2483° E 19.654 2.551% 500 88.049% 9.400% 0.273

51.5352° N, 26.2625° E 60.267 2.255% 500 96.482% 1.263% 0.215

51.4953° N, 26.3453° E 25.042 0.411% 500 97.408% 2.181% 0.177

51.4690° N, 26.7738° E 15.839 0.389% 500 99.546% 0.065% 0.124

51.5481° N, 26.2374° E 37.517 0.335% 500 98.856% 0.810% 0.069

51.7607° N, 26.3734° E 29.718 0.259% 500 98.326% 1.415% 0.095

51.7587° N, 26.3876° E 26.879 0.231% 500 94.990% 4.779% 0.119

51.5502° N, 26.2233° E 22.892 0.178% 500 98.267% 1.555% 0.082

51.5461° N, 26.2516° E 31.001 0.143% 500 95.273% 4.585% 0.119

51.4973° N, 26.3312° E 34.155 0.067% 500 96.671% 3.262% 0.063

51.6047° N, 27.0473° E 6.514 0.063% 500 99.754% 0.183% 0.135

51.5549° N, 26.2548° E 27.428 0.062% 600 97.510% 2.429% 0.199

51.7540° N, 26.3559° E 40.416 0.033% 500 98.305% 1.662% 0.096

51.7566° N, 26.4018° E 13.810 0.004% 1070 99.604% 0.392% 0.095

51.4799° N, 26.1313° E 16.768 0.001% 1500 90.980% 9.019% 0.080

51.5335° N, 26.2104° E 6.586 0.001% 1500 96.836% 3.163% 0.077

51.4994° N, 26.3170° E 28.376 0.001% 1500 97.433% 2.566% 0.088

51.5356° N, 26.1963° E 9.397 0.000% 1500 97.693% 2.306% 0.047

51.4711° N, 26.7597° E 10.578 0.000% 1500 97.565% 2.435% 0.044

This table includes more significant figures than may be appropriate given the levels of uncertainty in order to distinguish between pixels with nearly identical values.

3780 T.C. Eckmann et al. / Remote Sensing of Environment 112 (2008) 3773–3783

very poor correlationwith ASTER pixel counts for someof theseMODISfire pixels. For example, aMODISpixelwith an FRPof 20MWcontained251 ASTER pixels flagged as containing fire, while a MODIS pixel withan FRP of 38MWcontained only 24 ASTER pixels flagged as containingfire (Fig. 7). Likewise, this latter MODIS pixel contained 0.33% fire,according to MESMA, while a MODIS pixel that contained only 0.03%fire, according to MESMA, had 239 ASTER pixels flagged as containingfirewithin its footprint (Fig. 8). SomeMODIS pixels are outliers for bothFRP and MESMA, which may indicate limitations of the MODIS sensorfor these applications, or for using ASTER pixel counts based on aradiance threshold for these kinds of comparisons.

It is not surprising that Figs. 7 and 8 do not show better correlationswith ASTER pixel counts, because ASTER band 9 pixels are also mixed

Fig. 6. Comparison between FRP from the MOD14 collection 5 product, and the MESMAfire-size retrievals using the methods described in this paper. Each point in thescatterplot represents one of the 19 MODIS pixels from Table 3.

pixels, and thus are not exact measurements of fire size. Many possiblecombinationsoffire sizes and temperatures couldproduce the thresholdused in this paper for counting an ASTER band 9 pixel as containing fire(Fig. 9). This is a limitation inherent to assessing MESMA and FRPperformance using pixel counts from an ASTER radiance threshold, andthis limitation exists for other radiance thresholds as well. For example,many possible combinations of fire sizes and temperatures could alsoproduce the Morisette et al. (2005) ASTER radiance threshold of6.33 W m−2 μm−1 sr−1 (Fig. 9), or the maximum input radiance forASTER's band 9, which is 8.04Wm−2 μm−1 sr−1 according to Yamaguchiet al. (1998). In theory, MESMA could also be used to find subpixel firesizes and temperatures from ASTER, but MESMA would not work

Fig. 7. Fire radiative power from the MOD14 collection 5 product, compared to thenumber of ASTER pixels above the fire-detection threshold within the 33-by-66 pixelarea corresponding to each MODIS pixel's assumed footprint. Each point in thescatterplot represents one of the 19 MODIS pixels from Table 3.

Fig. 8. Retrieved subpixel fire sizes from MESMA using the methods described in thispaper, compared to the number of ASTER pixels above the fire-detection thresholdwithin the corresponding area for each MODIS pixel, as in Fig. 7.

3781T.C. Eckmann et al. / Remote Sensing of Environment 112 (2008) 3773–3783

effectively on bands that are saturated, and unfortunately, the ASTERscene displayed in Fig. 1 does not contain enough unsaturated bands forMESMA to effectively retrieve subpixel fire sizes from it. A follow-upstudy, to be published later, uses an ASTER fire scene that containsenough unsaturated bands for MESMA to retrieve subpixel fire sizesfrom the ASTER image, thus potentially providing a better validationsource for MESMA and FRP values from the coincident MODIS image.

In this study, MESMA selected a variety of different backgroundendmembers to model the 19MODIS fire pixels in Table 3, but MESMAselected only a few unique fire temperatures, predominately 500 K,with five at 1500 K. While the authors are not aware of any suitabledata that exist for assessing the accuracy of the fire temperatures fromthis scene, 500 K does not seem like an unreasonable value given thatmost of these areas appear to be smoldering, and 500 K is a reasonabletemperature for a smoldering fire (Kaufman et al., 1998). However, thebimodal temperature distribution retrieved by MESMA, consistingalmost exclusively of 500 K and 1500 K fires, may not reasonable givenexpected fire behavior and temperature distributions. We hypothesizethat MESMA selected 1500 K because 1500 K is the most similar to a

Fig. 10. Distribution of ASTER band 9 radiance levels within the footprints of the threeMODIS pixels (a, b, and c) with the highest ASTER fire pixel counts, out of the 19 MODISpixels from Table 3.

Fig. 9. Combinations of fire temperatures, and sizes (expressed as the percent of anASTER band 9 pixel), that produce a band 9 radiance of 2.00 W m−2 μm−1 sr−1 [thethreshold used in this paper], 6.33 W m−2 μm−1 sr−1 [the threshold used in Morisetteet al. (2005)], and 8.04Wm−2 μm−1 sr−1 [themaximum input radiance for ASTER's band9 according to Yamaguchi et al. (1998)]. This graph uses fire endmembers that weregenerated by MODTRAN using the settings in Table 1 and resampled to the spectralresponse of ASTER's band 9, but it does not account for radiance from the non-burningportion of the pixel. Thus, the actual fire sizes and temperatures needed to producepixels with these radiance values can be lower than indicated in this graph.

solar spectrum out of the included fire endmembers, and thus wouldprobably have the best fit for reducing discrepancies between thereflected solar radiation from a pixel in the image and the backgroundendmember selected to model that pixel. To test this hypothesis,MESMA was run again on these pixels that were modeled ascontaining 1500 K fire, using the same settings and endmembers asbefore but excluding all bands shorter than 3.5 μm (thus excludingMODIS bands 5, 6 and 7), to minimize the influence of reflected solarradiation on unmixing results. After excluding those bands, MESMAmodeled two of these pixels as having significantly lower firetemperatures, which could support the hypothesis that reflected

3782 T.C. Eckmann et al. / Remote Sensing of Environment 112 (2008) 3773–3783

solar radiation may explain why MESMA modeled these pixels ascontaining 1500 K fire.

We also evaluated the distribution of radiances from ASTER's band9 within the footprints of the three MODIS pixels with the highestASTER fire counts (Fig. 10). The first of these MODIS pixels contained251 ASTER pixels above the fire threshold (Fig. 10a), while the secondcontained 250 ASTER pixels above the fire threshold (Fig. 10b). Out ofthe 19 MODIS pixels detected as containing fire within the ASTERscene in Fig. 1, these two MODIS pixels had the highest ASTER firecounts, and also the two largest subpixel fire sizes retrieved byMESMA, at 2.55% and 2.26%, respectively, so ASTER fire counts areconsistent with MESMA retrievals for these two MODIS pixels. Only afew of the ASTER band 9 pixels within the footprints of these twoMODIS pixels had radiances above 3 W m−2 μm−1 sr−1, but a largenumber of these ASTER band 9 pixels had radiances between 2 and3 W m−2 μm−1 sr−1 (Fig. 10a and b), which is also consistent with therelatively low temperature (500 K) and large areas selected byMESMAfor these pixels. In contrast, FRP produced very different values forthese two similar pixels: the first of these two MODIS pixels had amoderate FRP of 20 MW (Fig. 10a), while the second of these MODISpixels had a high FRP of 60 MW (Fig. 10b). Because the distribution ofASTER band 9 radiances for these two MODIS pixels was nearlyidentical, MESMA results are significantlymore consistent with ASTERfire counts for these pixels than are the results from FRP. One possibleexplanation for the inconsistency between FRP and the ASTER firecounts for these twoMODIS pixels is that the band used for calculatingMODIS FRP is centered at a different wavelength (3.96 μm) thanASTER's band 9 (2.43 μm), and the amounts of energy present at thesetwowavelengths could be very different. It is also possible that FRP dida poorer job of discriminating between radiance from the fire versusthe non-burning background portion of these pixels, possibly becauseMODIS FRP is based primarily on a single band (Eq. (4)), whereasMESMA uses multiple bands for discriminating between fire andbackground radiance (Eq. (5)).

However, the third MODIS pixel for which we analyzed thedistribution of ASTER band 9 radiances, which had 239 ASTER pixelsabove the fire threshold, appears to show FRP outperforming MESMA(Fig. 10c). For this third MODIS pixel, MESMA selected a very small firearea (0.03%), while FRP calculated a more intermediate value(40MW). This result for FRP is consistent with the general relationshipbetween FRP and ASTER (Fig. 7) but stands out as an outlier forMESMA (Fig. 8). MESMA selected the same background endmemberfor the first and second pixels (Fig. 10a and b), but MESMA chose adifferent background endmember for this third pixel (Fig. 10c). Thissuggests that MESMA fire size retrievals may be overly sensitive to thebackground endmember selected. Unfortunately, because ASTER firepixels were mapped using a radiance threshold, and ASTER pixelsthemselves are mixed pixels, it is difficult to fully reconcile theobserved FRP and MESMA values with ASTER fire counts for thesepixels. For example, very small, hot fires in an ASTER pixel couldproduce the exact same value of band 9 radiance as cooler, moreextensive smoldering fires (Fig. 9), yet would also likely produce verydifferent values for MESMA and FRP. Reflected solar radiation in theseMODIS pixels may have also played a role in the discrepancies thatboth MESMA and FRP show with ASTER fire counts.

The follow-up study, to be published later, explores this hypothesisfurther using nighttime images that contain fire, to assess MESMAperformance on retrieving fire sizes and temperatures from sceneswith no reflected solar radiation. The lack of variety among themodeled fire temperatures in Table 3 also does not seem realistic, or ifthese temperatures are accurate, it shows that the scene in Fig. 1 doesnot contain enough variability in fire temperature to fully evaluateMESMA's potential. Follow-on studies will assess this by analyzingmultiple images across diverse fire conditions, and attempt to useimages with available coincident measurements of fire temperaturesfor validation.

The MESMA approach presented in this paper could be applied toimages of active fires acquired by other sensors as well, if enoughunsaturated bands are available for MESMA to use spectral shapeeffectively— ideally four bands or more should be available, althoughretrievals may be possible with fewer bands. These bands should alsobe located away from the absorption features of major atmosphericgases such as water vapor, and at wavelengths longer than ~1200 nmbecause smoke can influence unmixing at shorter wavelengths (e.g.Dennison et al., 2006). Thus, the MESMA approach presented in thispaper may not be suitable for retrieving fire properties from imagesacquired by the existing Landsat, Indian Remote Sensing (IRS), orSysteme Probatoire d'Observation de la Terre (SPOT) sensors, but itmay work for Hyperion, HyMap, and other sensors.

5. Conclusions

In this paper, we evaluated a new approach, MESMA, for retrievingfire temperature and area from a daytimeMODIS image.We comparedMESMA retrievals to FRP and evaluated the performance of bothmeasures using ASTERfire countswith an empirically derived 2Wm−2

μm−1 sr−1 threshold. Overall,fire size retrieved fromMESMAcorrelatedbetter with ASTER fire counts than did FRP. However, both measuresproduced anomalous results in some instances, with pixels containingsimilar ASTER fire counts producing very different FRP in several cases,and MESMA estimates of fire area showing severe discrepancies inother instances. MESMA fire temperature estimates seemed morerealistic after restricting the analysis to MODIS bands longer than3.5 μm, hypothesized to remove potential solar contamination, whichmay have led to anomalously high fire temperatures retrieved byMESMA. This suggests that daytime and nighttime imagery mayrequire slightly different procedures for MESMA retrievals of fire sizes.A significant limitation of our analysis is that the ASTER pixelsthemselves are also mixed pixels, and thus not a direct measure of firesize because there aremany combinations of fire size and temperaturethat could produce the ASTER radiance threshold used in this paper, orany other radiance threshold. A follow-up study, to be published later,addresses this limitation byusingMESMAto retrieve subpixelfire sizesfrom ASTER.

This paper describes what may be the first use of MESMA toretrieve fire properties from a non-hyperspectral sensor, the firstcomparison of FRP to MESMA fire property retrievals, and possiblyalso the first comparison of MESMA fire property retrievals to acoincident source of higher-resolution fire data. This is importantbecause a fire's size and its temperature have different influences onthe amount and chemistry of its trace gas and aerosol emissions,ecosystem impact, and fire spreading behavior, yet existingMODIS fireproducts cannot distinguish between a small hot fire and a largercooler fire within a pixel. Based on our results, MESMAmay be able toovercome this limitation, making it a promising tool for retrievingsubpixel fire properties from MODIS and for contributing informationnot available from existing measures. However, future studies thatassess MESMA's performance across a variety of geographical areasand fire conditions using multiple validation datasets, including on-the-ground measurements, will be necessary to gain a full under-standing of its accuracy and utility.

Acknowledgements

This work was supported by NASA Headquarters under the NASAEarth and Space Science Fellowship Program — Grant NNX06AF77H.The MODIS and ASTER data used in this study were distributed by theLand Processes Distributed Active Archive Center (LP DAAC), located atthe U.S. Geological Survey (USGS) Center for Earth ResourcesObservation and Science (EROS) http://LPDAAC.usgs.gov, and byNASA's Level 1 and Atmosphere Archive and Distribution System(LAADS) at http://ladsweb.nascom.nasa.gov. We would like to thank

3783T.C. Eckmann et al. / Remote Sensing of Environment 112 (2008) 3773–3783

Joel Michaelsen, Philip Dennison, Keith Clarke, Dylan Parenti, KerryHalligan, Greg Husak, and Seth Peterson for their help and sugges-tions. We would also like to thank the undergraduate students whohave assisted with this research, including Susannah Pitman,Jacquelynn Ybarra, Dani Luna, Julie Rueckheim, Ryan Del Rosario,Andrea Romano, Cara Moore, Alex Samarin, Lindsey Everett, MonicaAltmaier, Brian Boyce, Mike Isaacs, Marjorie Tamoro, Julia Sweet,Clarissa Esteves, Brenda Salguero, and Jenny Douthett. Finally, wewould like to thank the anonymous reviewers for their very helpfulcomments.

References

Andreae, M. O., & Merlet, P. (2001). Emission of trace gases and aerosols from biomassburning. Global Biogeochemical Cycles, 15, 955−966.

Ballantine, J. -A. C., Okin, G. S., Prentiss, D. E., & Roberts, D. A. (2005). Mapping NorthAfrican landforms using continental scale unmixing of MODIS imagery. RemoteSensing of Environment, 97, 470−483.

Berk, A., Bernstein, L. S., & Robertson, D. C. (1989). MODTRAN: A Moderate ResolutionModel for LOWTRAN7, AFGL-TR-89-0122 (pp. 1−38). Bedford, MA: Air Force Geophys.Lab. 01731.

Bianchini, G., Cortés, A., Margalef, T., & Luque, E. (2005). S2 F2 M— Statistical System forForest Fire Management. Computational Science — ICCS 2005 (pp. 427−434).

Cochrane,M. A., Alencar, A., Schulze, M. D., Souza Jr., C. M., Nepstad, D. C., Lefebvre, P., et al.(1999). Positive Feedbacks in the Fire Dynamic of Closed Canopy Tropical Forests.Science, 284, 1832−1835.

Collins, E. F., Roberts, D. A., & Borel, C. C. (2001). Spectral mixture analysis of simulatedthermal infrared spectrometry data: an initial temperature estimate boundedTESSMA search approach. IEEE Transactions on Geoscience and Remote Sensing, 39,1435−1446.

Csiszar, I., Denis, L., Giglio, L., Justice, C. O., & Hewson, J. (2005). Global fire activity fromtwo years of MODIS data. International Journal of Wildland Fire, 14, 117−130.

Dennison, P. E., Charoensiri, K., Roberts, D. A., Peterson, S. H., & Green, R. O. (2006).Wildfire temperature and land cover modeling using hyperspectral data. RemoteSensing of Environment, 100, 212−222.

Dozier, J. (1981). A method for satellite identification of surface temperature fields ofsubpixel resolution. Remote Sensing of Environment, 11, 221−229.

Finney,M.A. (1998).FARSITE: Fire Area Simulator—ModelDevelopment andEvaluation (pp. 47).Ogden, UT: USDA Forest Service, Rocky Mountain Research Station RMRS-RP-4.

Fitzgerald, G., Maas, S., & Detar, W. (2004). Spider mite detection and canopycomponent mapping in cotton using hyperspectral imagery and spectral mixtureanalysis. Precision Agriculture, 5, 275−289.

Flannigan, M. D., & Vonder Haar, T. H. (1986). Forest firemonitoring using NOAA satelliteAVHRR. Canadian Journal of Forest Research/Review, 16, 975−982.

Fortin, M. -J., Drapeau, P., & Legendre, P. (1989). Spatial autocorrelation and samplingdesign in plant ecology. Plant Ecology, 83, 209−222.

Giglio, L., Csiszar, I., & Justice, C. O. (2006). Global distribution and seasonality of activefires as observed with the Terra and Aqua Moderate Resolution Imaging Spectro-radiometer (MODIS) sensors. Journal of Geophysical Research, 111, G02016.

Giglio, L., Descloitres, J., Justice, C. O., & Kaufman, Y. J. (2003). An enhanced contextualfire detection algorithm for MODIS. Remote Sensing of Environment, 87, 273−282.

Giglio, L., & Justice, C. O. (2003). Effect of wavelength selection on characterization of firesize and temperature. International Journal of Remote Sensing, 24, 3515−3520.

Giglio, L., & Kendall, J. D. (2001). Application of the Dozier retrieval to wildfirecharacterization: a sensitivity analysis. Remote Sensing of Environment, 77, 34−49.

Green, R. O., Eastwood, M. L., Sarture, C. M., Chrien, T. G., Aronsson, M., Chippendale, B. J.,et al. (1998). Imaging Spectroscopy and the Airborne Visible/Infrared ImagingSpectrometer (AVIRIS). Remote Sensing of Environment, 65, 227−248.

Hanley, M. E., & Fenner, M. (1998). Pre-germination temperature and the survivorshipand onward growth of Mediterranean fire-following plant species. Acta Oecologica,19, 181−187.

Hoelzemann, J. J., Schultz, M. G., Brasseur, G. P., Granier, C., & Simon, M. (2004). GlobalWildland Fire Emission Model (GWEM): evaluating the use of global area burntsatellite data. Journal of Geophysical Research, 109.

Justice, C. O., Giglio, L., Korontzi, S., Owens, J., Morisette, J. T., Roy, D., et al. (2002). TheMODIS fire products. Remote Sensing of Environment, 83, 244−262.

Kaufman, Y. J., Justice, C. O., Flynn, L. P., Kendall, J. D., Prins, E. M., Giglio, L., et al. (1998).Potential global fire monitoring from EOS-MODIS. Journal of Geophysical Research,103, 32215−32238.

King,M.D.,Menzel,W. P., Kaufman,Y. J., Tanre,D., Bo-Cai, G., Platnick, S., et al. (2003). Cloudand aerosol properties, precipitable water, and profiles of temperature and watervapor fromMODIS. IEEE Transactions on Geoscience and Remote Sensing, 41, 442−458.

Lombardo, V., Buongiorno, M., & Amici, S. (2006). Characterization of volcanic thermalanomalies by means of sub-pixel temperature distribution analysis. Bulletin ofVolcanology, 68, 641−651.

Matson, M., & Dozier, J. (1981). Identification of subresolution high temperature sourcesusing a thermal IR sensor. Photogrammetric Engineering and Remote Sensing, 47,1311−1318.

Morisette, J., Giglio, L., Csiszar, I., & Justice, C. O. (2005). Validation of the MODIS activefire product over Southern Africa with ASTER data. International Journal of RemoteSensing, 26, 4239−4264.

Oertel, D., Zhukov, B., Thamm, H. -P., Roehrig, J., & Orthman, B. (2004). Space-borne highresolution fire remote sensing in Benin, West Africa. International Journal of RemoteSensing, 25, 2209−2216.

Painter, T. H., Dozier, J., Roberts, D. A., Davis, R. E., & Green, R. O. (2003). Retrieval ofsubpixel snow-covered area and grain size from imaging spectrometer data. Re-mote Sensing of Environment, 85, 64−77.

Peterson, S. H., Goldstein, N. C., Clark, M. L., Halligan, K. Q., Schneider, P., Dennison, P. E.,et al. (2005). Sensitivity Analysis of the 2003 Simi Wildfire Event. Proceedings,Geocomputation, August 1–3, 2005. Ann Arbor, Michigan.

Powell, R. L., Roberts, D. A., Dennison, P. E., & Hess, L. L. (2007). Sub-pixel mapping ofurban land cover using multiple endmember spectral mixture analysis: Manaus,Brazil. Remote Sensing of Environment, 106, 253−267.

Prins, E. M., &Menzel, W. P. (1992). Geostationary satellite detection of biomass burningin South America. International Journal of Remote Sensing, 13, 2783−2799.

Pyne, S. J., Andrews, P. L., & Laven, R. D. (1996). Introduction to Wildland Fire. New York:Wiley.

Randerson, J. T., van der Werf, G. R., & Collatz, G. J. (2005). Fire emissions from C3 and C4

vegetation and their influence on interannual variability of atmospheric CO2 andδ13CO2. Global Biogeochemical Cycles, 19, GB2019.

Roberts, D. A., Gardner, M., Church, R., Ustin, S., Scheer, G., & Green, R. O. (1998).Mapping Chaparral in the Santa Monica Mountains using multiple endmemberspectral mixture models. Remote Sensing of Environment, 65, 267−279.

Roberts, D. A., Smith, M. O., & Adams, J. B. (1993). Green vegetation, nonphotosyntheticvegetation, and soils in AVIRIS data. Remote Sensing of Environment, 44, 255−269.

Roberts, G., Wooster, M. J., Perry, G. L. W., Drake, N., Rebelo, L. M., & Dipotso, F. (2005).Retrieval of biomass combustion rates and totals from fire radiative powerobservations: application to southern Africa using geostationary SEVIRI imagery.Journal of Geophysical Research, 110.

Salomonson, V. V., Guenther, B., Barnes, W. L., Therrien, N. J., & Murphy, R. E. (2000).Early instrument performance results from the Terra/Moderate Resolution ImagingSpectroradiometer (MODIS). Geoscience and Remote Sensing Symposium, 2000.Proceedings. IGARSS 2000IEEE 2000 International, vol. 943 (pp. 943-946).

Schultz, M. G. (2002). On the use of ATSR fire count data to estimate the seasonal andinterannual variability of vegetation fire emissions. Atmospheric Chemistry andPhysics, 2, 387−395.

Shephard, M. W., & Kennelly, E. J. (2003). Effect of band-to-band coregistration on fireproperty retrievals. IEEE Transactions onGeoscience and Remote Sensing, 41, 2648−2661.

Teshima, Y., & Iwasaki, A. (2008). Correction of attitude fluctuation of Terra spacecraftusing ASTER/SWIR imagery with parallax observation. IEEE Transactions on Geoscienceand Remote Sensing, 46, 222−227.

Trigg, S. N., & Roy, D. P. (2007). A focus group study of factors that promote and constrainthe use of satellite-derived fire products by resource managers in southern Africa.Journal of Environmental Management, 82, 95−110.

van derWerf, G. R., Randerson, J. T., Collatz, G. J., Giglio, L., Kasibhatla, P. S., Arellano, A. F.,Jr., et al. (2004). Continental-Scale Partitioning of Fire Emissions During the 1997 to2001 El Nino/La Nina Period. Science, 303, 73−76.

Wolfe, R. E., Nishihama, M., Fleig, A. J., Kuyper, J. A., Roy, D. P., Storey, J. C., et al. (2002).Achieving sub-pixel geolocation accuracy in support of MODIS land science. RemoteSensing of Environment, 83, 31−49.

Wooster, M. J., Zhukov, B., & Oertel, D. (2003). Fire radiative energy for quantitativestudy of biomass burning: derivation from the BIRD experimental satellite andcomparison to MODIS fire products. Remote Sensing of Environment, 86, 83−107.

Yamaguchi, Y., Kahle, A. B., Tsu, H., Kawakami, T., & Pniel, M. (1998). Overview ofAdvanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER). IEEETransactions on Geoscience and Remote Sensing, 36, 1062−1071.