Global Fire Mapping and Fire DangerEstimation Using AVHRR lmages

Emilio Chuvieco and M. Pilar Martin

AbstractA critical rcuiew of avunn images for large-scale fire appli-cations is presented. Discussion is divided into firc detectionand mapping, on one side, and fire danger estimation, onthe other. In both cases, the main lines of research are re-ported, as well as the problems faced. Finally, some futurcresearch guidelines are suggested.

The Role of Fire in Global StudiesBiomass burning is one of the key factors affecting globalprocesses (Figure 1). During the combustion itself, a largeamount of trace gasses are released, strongly affecting atmos-pheric chemistry (Crutzen et a1.,7979). Frequently, fire re-sults in a partial or complete destruction of vegetation coverwhich modifies the radiation balance by increasing the al-bedo, as well as the hydrological cycle, while raising soilerosion,

The increased rate of biomass burning in the last decadeis estimated to be caused by the increment of tropical defo-restation in the Amazon basin and from the augmented rateof burning in cultivated areas in the African savanna. Thisincrease in the amount of released gases to the atmospherehas important effects on atmospheric processes and globalclimate. The main effects include the following (Kaufman efaL, 7992):

o Some of the emitted gases (COr, CO, CH4, CH"CI) representan important contribution to the greenhouse effect that heatsthe atmosphere. Concentrations of CO as high as 3,700 ppbv(particles per billion volume), which are 50 times higher thanduring normal atmospheric conditions (Stearns ef a/., 1986),have been measured over a smoke plume.

o Smoke emissions also include reactive gases, such as NO"and CHo, which cause elevated levels of ozone and acid rain.Ozone concentrations in intense fire areas have been found tobe as much as four times greater than regular levels, reaching70 to 80 ppbv (Kaufmat et aL.,1992; Setzer and Pereira,1e91b) .

o Biomass burning is also a major source of organic hygro-scopic particles which increase the available cloud condensa-tion nuclei. Such increase generates brighter clouds thatreflect solar radiation to space, thus reducing the temperature(Robock, 1988).

o Finally, burning is also a source of graphitic carbon whichincreases absorption of solar radiation by the atmosphere andby clouds and, as a result, implies a heating effect.

On a local scale, biomass burning also has strong effects onthe landscape and ecology, especially in semiarid lands.Changes in traditional land-use patterns have recently modi-fied the incidence of fire in these territories. For instance,

I i"."" ]Caru&tties

L I ]

iI

Gaseous * ['tpn .* LandcoverEmissiom charge

Figure 1. Global implications of biomassburning.

t

iltrdt"tt."|'udget- r ' - -

IIlydrological Soil

CYcle Eroeion

rural abandonment in the European Mediterranean basin hasimplied an unusual accumulation of forest fuels which in-crease fire risk and fire severity. In tropical areas, recurrentuse of fire as a soil fertilizer, or as a means of pasture im-provement, involves soil and grass degradation in manyireas as the fire cycles are shortened (Cahoon et al',7ss2)'Fires also modify the hydrological cycle, by increasing therunoff as a consequence of less soil protection. Finally, wild-land fires also have direct effects on human casualties ()iiiaef o1. , 1989),

Remote Sensing of Forest FiresRemote sensing systems may be very valuable in different as-pects of fire management problems. First, they contribute topre-fire planning, which is directed to the optimum organi-zation of resources for fire emergencies, This phase involvesknowledge of risk conditions, forest value, land quality,weather patterns, fuel conditions, and topographic data.These variables are closely related to fire ignition and rate ofspread.'

The second phase of fire management is directly associ-ated to the fire itself: detection, suppression, and burnedland mapping. Systems are required to provide informationon fire aciivsareas. They should operate day and night, de-spite low visibility caused by smoke or dense timber cover,

Photogrammetric Engineering & Remote Sensing,VoI. 60, No. 5, May 1994, pp. 563-570,

0099-7772 I 94i600 1-5 63$o3.oo/o01994 American Society for Photogrammetry

and Remote SensingDepartment of Geography, Universidad de Alcal6 de Henares,Colegios, 2,28807 Alcald de Henares, Spain.

and Sive some Suidelines to distinguish between potentiallydangerous fires and those of no concern. Finally, post-fireevaluation makes it possible to reduce the effects of wildlandfires. Timely damage assessment should provide critical in-formation for planning the recovery of forest stands (Smithand Woodgate, 1985) and to avoid soil erosion (Isaacson etal., 1,982).

Airborne remote sensors have been used since the latesixties for fire spot detection (Hirsch et aL, L977). With thedevelopment of the Landsat program during the 70s, severalprojects were conducted to test the reliability of satellite im-agery for forest fire research. These applications, as well astlie ones derived from NoAA-A'rHm data, may be classified inthe following themes: (1) detection of fire spots, (2) cartogra-phy of burned areas, (3) estimation of gas emissions, (4) as-sessment of fire effects and plant cover evolution after thefire, (5) monitoring forest fuel conditions (vegetation mois-ture stress), and (6) development of risk models by the inte-gration of image interpretation within a GeographicInformation System.

All of these topics complement and improve traditionaltechniques of fire management (Table 1). In this paper, wewill emphasize the global, large-scale approach, which hasbeen mainly based on NOAA-AVHRR images. We have groupedour review in two primary areas: fire detection and mapping,and fire danger estimation, The first one has received moreattention by remote sensing specialists, as traditional sourcesprovide generally poor and costly estimations of burned area.Fire danger estimation requires more research, in order toaddress the complexity of factors affecting fire ignition andfire rate of spread.

Fire Mapping from AVHRR lmagesOne of the main problems affecting fire management is theIack of appropriate statistics on burned land, Even the coun-tries more severely affected by this problem do not haveproper data on fire incidence, as most of the time fires arenot mapped and only general statistics are available. The ex-ample of Brazil is especially outstanding, The Iatest statisticsavailable from the Food and Agriculture Organization (FAo)(Calabri and Ciesla, 1992) estimated that 54,280 hectareswere burned in that country during 1987. A project based onAVHRR images for the same summer season increased that es-timation to 20 million hectares (Setzer and Pereira, 1991a),almost 400 times higher than official statisticsl However, itshould be remarked that FAo statistics only include the lit-toral states of Brazil. This results from lack of access to dataon the Amazonian basin, which is the most affected area.

Although the Brazilian case can not be considered as therule, it is a good example of the weaknesses associated withtraditional sources of wildland fire data. Logistical inaccessi-bility of many tropical forest areas make global statistics forthose countries ouite uncertain, which causes severe difficul-ties in setting up globat fire management strategies. In thoseareas. the use of satellite information can be considered themost reliable method to generate global statistics on wildlandfires.

Countries with more detailed fire reports generally pro-vide more orecise statistics. However, the data are not avail-able until ieveral weeks (or even months) following the fireevent. As a result, vegetation recovery is not assessed, and alack of regrowth may constitute a severe soil erosion hazardflsaacson et a].,7s82). Moreover, these field inventories areoften very general. Usually, only the scorched perimeter isdrawn, but no information about the species affected or se-

564

verity of damage is provided. Finally, studies on vegetationsuccession after fire are seldom done.

Fire detection and mapping have been based mainly onmiddle infrared data. Considering that forest fires tempera-tures commoniy range from 500" to 1,000"K (Robinson,1991), the most suitable band for f ire detection is located be-tween 5.8 and 2.9 pm according to Wien's displacement law.Hot areas in this channel show a strong radiometric contrastwith respect to thermal infrared bands, which are betteradapted to the average Earth temperatures (300'K).

Fire detection from space is obviously very much depen-dent on observation frequency. The Earth resources satellites(such as Landsat or SPor) do not provide enough time cover-age for fire detection. On the contrary, meteorological satel-Iites have proven to be very useful for these purposes. NOAA-AVHRR images are weil suited for fire detection and mappingbecause of their adequate coverage cycle (1.2 hours) and widespectral range, which also includes the middle infrared (Kid-wel l , 1991).

The use of anuRR images for fire detection and fire map-ping has been successfully tested in southeast Asia (Malin-greau ef o1., tges; Malingreau, 1990), Canada (Chung and Le,1984; Flannigan and Vonder Haar, 1986), the U.S.A. (Matsonand Dozier, 1.981; Matson et aL, 'l.g14), Australia (Hick ef o/.,19BO), Africa (Langaas, 1992; Malingreau ef o1., 1990), China(fi j ia ef a1., tgBgl Chandga ef a1., 1991), and Brazil (Matsonand Holben, 1987; Kaufman et aI., 1990a; Setzer and Pereira,1991b).

In Canada, high accuracy for detecting large-scale forestfires was found in a pilot study conducted in Alberta (Flan-nigan and Vonder Haar, 1986). Overall, 46 percent of the to-tal number of fires were located with data obtained fromAVHRR channel 3. This accuracy raised to B0 percent whenonly cloud-free fires were considered. Precision reached 95percent, when refering only to medium or large fires (over 40hectares). Accuracy for small fires was limited, 10 to 12 per-cerrt, although better results were reported if only cloud-freeareas were taken into account (up to 87 percent).

In the Brazilian Amazonia several projects have been de-veloped to estimate the total area burned yearly, in order toderive deforestation rates for the whole country. For thesummer season of 1987, analysis of avgRn channel-3 imagesyielded a total estimation of 350,000 fires, covering as muchas 200,000 sq km (Setzer and Pereira, 1991b), These resultswere obtained from the extrapolation of observed fires in 46AVHRR images acquired from July to October. Pixels withhigh temperatures in channel 3 were labeled as fires. Thiscalculation was extended to the whole studv period usinsthe following formula:

a 1 = @ * d ) x A o x ft '

where rtt is the total area burned, n is the avefage number ofpixels labeled as fire within the summer season (n : num-ber of fire pixels / number of images), d is the length in daysof the summer season (80 in this case), t is the average timeof fire duration in days (1.5 in this case),Ap is the pixel area(sq km), and/is an adjusted factor to account for the satura-tion of channel 3 (see below). Successful results of the 1987project based the extension of this methodology to an opera-tional program. This program has been carried out since thenby the Brazilian Institute of Space Research (wrn) (Setzerand Pereira, 1991a).

The study of Setzer and Pereira also estimated the

PE&RS

Tlet-E 1. Usr or RruorE SEr.lsrr.lc Tecurureues Vensus TnaorrorunL MgrHoos roR Frne MANAGEMENT Ar LocAL euo Rectorual SclLEs

Danger,lRiskEstimation

DetectiorVMapping

SmokeEmissions Effects

TraditionalMethods

RemoteS a n c i n o

I raortlonal

Methods

Remote( o n c i n o

Weather StationsFuel maps

TM-MSS-HRV.,.+ G.I.S. (Risk)

Weather networkHistoric records

A\TTRR NDVIand Thermal

Sketches

Airborne IRTM-MSS.HRV...

Statistics fromfire reports

AVHRR Ch. 3Geostationary

AerialSamples

Airborne/Groundborne

Not available

AVHRR visibleMAPS/TOMS

n n l t , f ^ r

managed areas

Multitemporaihigh resolution

Not available

A\rHRR multi-temporal series

amount of gases released to the atmosphere from the 1987Brazilian Amazonian fires. Based on a simple method, totalamounts of 7,700 Tg of co,, 520 Tg of c, ind 94 Tg of cowere calculated (Setzer and Pereira, 1991b). More detailedprojects to measure gas emissions from tropical forest havebeen carried out using satellite, airborne, and field sensors(Kaufman et 01., 1990a and t00z).

Most of the work related to gas release is based on thedetermination of optical thickness, particle size, and singlescattering albedo using short-wave bands. Kaufman ef a/.(1990b) developed a self-consistent algorithm for simulta-neous determination of smoke characteristics. This method isbased on satellite images of the surface (land and water) inthe visible and near infrared bands. The algorithm derivesthe aerosol characteristics from the difference in upward ra-diances between two images of the same area, one acquiredon a clear day, and one during hazy conditions. This methodwas applied to the estimation of smoke concentrations overthe Chesapeake Bay coming from Canadian fires. A close re-lation was found between the aerosol estimation from AVHRRdata and ground measurements (Kaufman et o1., 1990b). Amore qualitative approach was adopted by other authors toestimate the area affected by smoke plumes in large-scale for-est f ires [Chung and Le, 1984; Jij ia et o/., 19Bg). There arealso some examples of using space photographs for smokemonitorinS. From the comparison of Skylab (1973) and SpaceShuttle photographs (1980), a tenfold increase in smoke pallsurfaces was estimated (Herfert and Lulla, 1990).

In addition to the use of avHRR imagery, other sensorshave proven to be efficient for fire detection and mapping.These include geostationary data (Prins and Menzel, t99z)and oMSp night images (Cahoon et al,, 79s2). Smoke emis-sions have also been monitored from the MAPS (Measurementof Air Pollution from Satellites) experiment on board theSpace Shuttle (Watson et o1., 1990).

Problems for Operalional Fire MappingThe main problem of avsRR data for these applications isthe low thermal sensitivity of channel 3, which is saturatedat 320'K. As a result, fire spots can be easily confused withagriculture burns or even bare soils, which frequently reachthese temperatures during the summer at the afternoon pass(Belward, 1991). Discrimination from agricultural fires couldbe partially achieved by choosing evening or night images,because this type of burning tends to be done during day-Iight periods (Malingreau, 1990). The confusion with adja-cent areas is not a severe problem in tropical forests, becausethe vegetation surrounding the fire is much cooler than satu-ration temperatures caused by evapotranspiration. In the caseof semiarid lands, fire active areas and bare soils may be dis-

criminated on night images, due to the fact that bare soils arecooler than fires at that time (Langaas, 1992)' Monitoring thetemporal dynamism of the target surfaces also provides goodclassification of fire pixels (Lee and Tang, 1990)' In any case,sensors with higher thermal sensitivity are desirable. In theBrazilian Amazonia, airborne experimental fire detectionscanners with saturation levels up to 900"K have been suc-cessfully tested (Riggan et al., 1993). The future ModerateResolution Imaging Spectroradiometer (MoDIS) will include amiddle infrared channel that saturates at 500'K, which willnotably increase the potential of satellite fire detection sys-tems.

On the other hand, low thermal sensitivity of AVHRR cre-ates problems of overestimation of the burned area. It shouldbe n6ted that a pixel may be saturated even if only a smallproportion is actually occupied by fire, In fact, fires with aiemperature above 500'K, even if they occupy less than 0.1perCent of the pixel size, will most Iikely reach pixel satura-iion temperatures (Kaufman et al., 1990aJ. For this reason, aconsistent tendency of overestimation has been found whenAVIIRR channel-3 fire mapping estimations are comparedwith higher resolution data, such as Landsat Mss or TM. Ac-cording to several experiences in the Amazon basin, thisoverestimation is on the order of 43 percent, or 1.5 times theareal extent observed with Landsat TM (Setzer and Pereira,199 1 a) ,

Cross analysis of channels 3 and 4 might partially solvethis problem as the difference in radiances between bothchannels will increase when a greater percentage of a pixel isoccupied by an active fire. This proportion, as well as the ac-tual temperatures of the fire, may be calculated following amethod proposed by Dozier (19S1) and Matson and Dozier(1981), whiCh uses the radiance contrast between channels 3and 4 of AVHRR: i.e.,

L , : p L " ( T ) + ( 1 - p ) 1 . ( T o )

L" : pLn(T ) + (7 - P)L n(T 6)

where I,(4) is the radiance in band i as a result of the firetemperature (7,), To is the background temperature (whichcan be estimated from areas nearby the fire), and p is theproportion of the pixel size occupied by the fire. Il spite-ofits interest, it should be noted that Dozier's algorithm onlyworks properly when pixels are not saturated and does nottake into account the attenuation caused by water vapor inboth bands. Pixel saturation may be solved by using a widerIFov sensor, less likely to be saturated by a small fire. Thiswas the basis of using coES-vaS thermal channel (7- by 7-kmpixel size) and middle infrared for detecting fires in theAmazonian basin (Prins and Menzel, 1992). Emphasizing

these radiance differences between channels 3 and 4 by tree-decision rules and contextual methods have also been pro-posed for an automatic discrimination of fire active areas(Lee and Tag, 1990; Illera et ql., rggs).

Another source of problems for using AVHRR channel 3in fire detection is cloud cover. Although the middle infraredchannel makes it possible to discriminate fires when cloudsare thin (Figure 2), cloud contamination is still an importantsource of errors. Small fires are frequently obscured by thickclouds which make their discrimination especially problem-atic (Flannigan and Vonder Haar, 1986; Setzer and Pereira,1991a).

Fire Danger EstimationForest fire literature usually discerns between the conceptsof fire risk and fire danger. The former is related to causality,and is commonly divided between human induced and light-ening. The latter is more associated with weather conditionswhich account for vegetation stress status and, consequently,are closely connected to flammability. Most of the currentfire danger indices rely on temperature, air humidity, andwind for estimating vegetative status (Van Wagner, 1987), Di-rect measures of vegetation and duff moisture are much morecomplex and require costly spatial sampling. Consequently,they cannot be used in global fire danger estimations.

The main problem associated with meteorological mea-sures is the generally sparse geographical distribution of ob-servation points. In spite of the increasing use of automaticweather stations, climate data are frequently available onlyfor areas that are distant from the forest land. Interpolation isperformed on a very rough basis, especially in complex ter-rain zones; this may cause fire danger values to vary signifi-cantly from the original points of measure.

Remote Sensing of Vegetation StressSeveral studies have shown strong correlations betweenA\,TIRR spectral vegetation indexes and critical physiologicalvariables. Primary examples are green biomass, Leaf Area In-dex, evapotranspiration, primary productivity, and ActivePhotosynthetically Absorbed Radiation by the plant (APAR)(Sellers, 1989). Most of these studies found significant statis-tical correlations between NDM and vegetation trends (sea-sonality: Cihlar et o1., 1991), but the relationships are muchweaker for short periods, i.e., one or two weeks (Deblondeand Cihlar, 1993).

Fire danger estimation demands frequent monitoring ofvegetation stress, Vegetation moisture is a particularly difficult parameter to estimate as it accounts for little spectralvariation with respect to other environmental factors (Cohen,1991). However, spectral characterization of vegetation stressis possible if temporal profiles are derived, and the conbastbetween living and dead components is emphasized.

The most common vegetation stress estimation has beenthe analysis of IvovI multitemporal series. As mentioned pre-viously, NDVI increase is related to LAI and apAn. Therefore,this vegetation index provides a good indication of vegeta-tion healthiness. On the other hand, decrements of NDtr'I arerelated to reduction of plant vigor and greenness, which arealso connected with vegetation moisture content, Multitem-poral accumulated Novt composites have been successfullycorrelated to accumulated evapotranspiration (Kerr et o/.,1989; Deblonde and Cihlar, 1993), rainfall amount (Kerr efa1., 1989), and crop moisture indices (Walsh, 1987) for sem-iarid environments. Strong correlations between NDVI andsoil temperature have also been measured (Dabrowska, 1991).

566

Figure 2, AVHRR images of a wildland fire in the Mediterra-nean coast of Spain. (Left) Clouds in the visible channelscreen the active flre. (Ri€ht) Channel-3 image clearly dis-plays the fire focus because the fire has a strong radio-metric contrast with clouds in this spectral channel.

An alternative to the use of NDVI series for vegetationmoisture estimation is to follow the thermal dynamism of thevegetation cover. Although not directly related to fire danger,several studies on crop moisture might be quite useful forfire danger estimation, as they deal with vegetation spectralbehavior under drought conditions.

Vegetation moisture stress can be measured as the ratioof actual and potential evapotranspiration (ET/PET). ET hasproven to be Iinearly related to net radiation and to the dif-ference between air and surface temperature. The consis-tency of this relationship is increased when applied toaccumulated periods of five or ten days (Seguin et al,, 1991):i .e . ,

) n r : > R " + n a - b 2 ( 7 " - T . )

where sT is the actual evapotranspiration, ff is net radiation,I and Q are surface and air temperatures, respectively, anda and b are adjustment constants. Over long term periodsand at regional scales, correiation coefficients ofr=0.99 be-tween accumulated values of (T" - T") and > ET - R" havebeen measured (Seguin et al,, 7s97). This relationship mightsuccessfully be applied to fuel moisture estimation, comple-menting the analysis of unvt time series.

Use of AVHRR Data in Fire Danger EstimationSeveral projects developed so far have addressed the opera-tional application of NDVI multitemporal profiles for fire dan-ger estimation. To avoid atmospheric disturbances, most ofthese studies have used weekly or biweekly maximum valuecomposites, The results have been very precise for areas ofhomogeneous and herbaceous vegetation. In Nebraska, Nnvtwas used to estimate percent of green cover and biomass.Correlations of. r:O.75 with biomass percent and r:0.82with green cover were obtained using several grassland areasas test fields (sadowski and Westover, 1986), A similar studyin four states (North and South Dakota, Nebraska and Kan-sas) obtained very good correlations between weekly Nnvtcomposites and percent of green cover (r€ ranged from 0.86to 0.6), These results were used as an input of a fire dangerindex at county level (Eidenshink ef a1,, 1989).

Significant relationships between NDVI and fuel moisturehave also been found in herbaceous lands of Australia (Pal-tridge and Barber, 1988), In this case, a variation of NDVI

PE&RS

was suggested to better identify the maximum green compo-nent: i.e.,

y : (NrR - R * 1.2) / (Nrn + n)

where NIR and R are the near infrared reflectance and red re-flectance, respectively. From the V value, fuel moisture con-tent was estimated according to the following formula:

FMC = 2s0 y(t)/ V(nov)

where FMC is the fuel moisture content, V(t) is the vegetationindex at any specific time, and V(nov) is the same measureon a November image, the time of maximum vegetation vigorin Australia.

Temporal profiles of rvc estimated by AVHRR imagesand ground measured data followed a similar trend, withclose relationship between NDVI estimation and fuel moisturemeasures on the ground (Paltridge and Barber, 1988).

These results from grasslands need to be extended toshrub and forest lands, where fire has more severe effects.With this objective, an experimental study has been con-ducted in four States of the Great Plains {Eidenshink et 41.,1990). This study assumed that the relation between NDrr't ev-olution and vegetation moisture would be very different forgrasslands and shrub-forest lands. For the former, drynessand wetness are much more uniform. As a result, NDvI calcu-lated from grasslands varies along with LAI and chlorophyllproduction. On the contrary, in shrubs and forest lands newgrowths and higher moisture contents may be spectrallyveiled by the more cured components of the plant. Conse-quently, to monitor actual vegetation conditions, it has beensuggested that the relative instead of the absolute variationsof Novl be analyzed. The authors proposed three indices:

o GRN.or = 100 (ND. - ND",I")/(ND-", - ND^,")where GRN,.1 is relative percent green, NDo is the observedNovt for an specific pixel, and Nn-'" and ND-." are the maxi-mum and minimum Novt of that pixel for the whole studyperiod.

o GRNur," : 100 (ND. - NR"lin) NRn,'xwhere cRNo6" is the absolute percent green, NRmin correspondsto the minimum NDVI for fully cured grass (0.05), and NR-", isthe maximum NDvI observed in the historical record of multi-temporal composites of United States (0.66).

o sM = 250 (cRN,.r + cRN"b)/zwhere SM is site moisture, GRN".I and cRN"n, are the fractionalrelative and absolute percent green for any specific pixel on agiven date.

Although no quantitative evaluation with ground data wasperformed in this study, seasonal trends of these indicesu/ere clearly associated so they could provide updated infor-mation concerning greenness evolution for forest stands. Theintroduction of these greenness calculations in forest firedanger indices has already been addressed fBurgan and Hart-ford, 1993).

Finally, another method to estimate fire danger is basedon the accumulative decrements of wovt values in each pixelover a dry season. A project developed on the Mediterraneancoast of Spain tested this method, Daily images were usedinstead of maximum composites. Cloud cover and view-angleeffects were minimized by selecting only the most suitableimages for the study area. A set of seven images on the sum-mer of 1987 was used to point out those areas with higherNDVI decrements over the study period. One of those areashappened to be affected by a large forest fire at the end ofthat per iod (L6pez ef 01. , 19s1) .

PE&RS

Fuel MapsForm, size, arrangement, and continuity of vegetation arecritical variables for fire behavior prediction, Usually, thevast diversity of vegetation associations require that they beorganized into generalized categories referred to as fueltypes. It is then possible to define the fuel types as com'plexes of sufficient homogeneity and size to maintain a co-herent fire behavior over time (Van Wagner et al., 7992).

Although fuel types are less time dependent than fuelmoisture, updated cartography is required for fuel manage-ment and fire behavior prediction systems. Remote sensing isan ideal tool for mapping fuel types, but it faces severalproblems. Fuel types are defined by both the canopy and bythe understorey vegetation. Strictly speaking, this discrimi-nation would be very complex in dense stands as remotesensors only obtain information from the canopy, However,most of these fuel types such as boreal spruce or coniferplantation are considered standard vegetation associations,which might be related to canopy spectral characteristics.

Several studies have successfully mapped fuel typesusing Landsat images and auxiliary variables, such as topoS-raphy or soils types (Burgan and Shasby, 1984). These stud-ies are useful for local-scale management, but are notsuitable for global fuel management where low resolutionsensors are iequired. This global approach has been adoptedby several projects which compiled fuel maps from AVHRRdata. Areas as large as the ten western states of the U.S.A.were used to test the suitability of these images (Werth ef o1.,1985). Results have shown that general discrimination wasaccurate enough, but some fuel models of the National FireDanger Risk System (NFDRS) had to be digitized from auxil-iary sources.

A more detailed study conducted in Oregon (Miller eto1., 1986) showed 92 percent accuracy from an unsupervisedclassification of four AVHRR images. Ancillary informationwas also used in this project in order to label the 45 clustersgonerated, as well as to group them into 11 fuel types of theNFDRS. Elevation, roads, and visual interpretation were ap-plied to discriminate categories with high spectral overlap.

Mosi recent developments of cls techniques wil l allowfor easy generation of fuei maps from AVHRR images and aux-iliary information, as more global data will be available forIabeling and discrimination of the most problematic cate-gories. The experiences gained from the land-cover maps ofthe conterminous U.S.A. are quite promising for this applica-tion (Brown et o/., 1993).

Conclusions: Scenarios of Future ResearchAs reviewed in this paper, the combined use of AVHRR chan-nels makes it oossible to acouire valuable information con-nels makes it poss acouire valuable information con-cerning (1) the geographical origin of the fires, (2) theircerning (1) the geogtspatial and temporalspatial and temporal development, (3) their effects, and (a)those parameters related to fire ignition and behavior.

Fire MappingIn spite of the efforts developed so far, several possibilitiesin using AVHRR data for fire detection and mapping are stillsubject to future improvements. First, there is no quantitativeapplication of a daily series of AVHRR data to monitor firegrowth. This product could be quite interesting as an inputto improve fire behavior programs such as BEHAVE (Burganand Rothermel, 1984), or as a way of testing their perform-ance. Fire growth studies might take advantage of the timecoverage of nvHRR channel-3 data to monitor daily (or eventwice a day, using night images) fire evolution in the case of

567

Jdy l? - 21 July 26 -29 Augrut4 - 13

Figure 3. NDV temporal profile for healthy andburned areas of pine and scrub vegetation,Data were obtained from weekly maximumvalue composites of evunn images on theMediterranean coast of Soain. The reductionof NDV| values after the fire is evident.

Iarge events. A preliminary study of a 20,000-hectare forestfire on the Mediterranean coast of Spain has shown a vervpromising future for this approach (^Chuvieco ef a/,, 1993i

On the other hand, most of the fire mapping research de-veloped so far is based exclusively on channel-3 data. Multi-temporal analysis of wovI values has rarely been used forthis purpose (Matson and Holben, 1987). Working with mul-titemporal series of NDVI values will solve some of the prob-Iems previously reported, such as those derived from thethermal sensitivity of AVHRR, Cloud-free and near-nadir im-ages from before and after the fire might be used for quanti-tative evaluation of burned land because the ND'rrI value isseverely reduced after a fire (Figure 3). In using ND'fl forburned land mapping, maximum value composites have beenemployed (Kasische ef o1., 1993). However, the effect of mis-registration errors and cloud contamination may make diffi-cult an accurate evaluation for small fires (Martin andChuvieco, 1993),

Finally, future improvements in sensor design mightgreatly affect fire mapping from space. The new generation ofAVHRR sensors will produce better radiometric contrast, withhigher saturation levels, although this channel will uot beoperated at daytime. On the other hand, the MoDIS instru-ment will notably augment the spectral information providedby A\TIRR, with two new channels at 2.1 and 1.65 pm andhigher thermal sensitivity.

Fire Danger EstimationIn spite of the complexity of using AVHRR data for fire dangerestimation in forest stands, the experiences with grasslandsappear quite favorable for inputing remote sensing data tofire danger indices. However, several problems should be ex-amined in order to make operational use of this technology.Among them, we highlight the following:

o Fire danger requires a short-term estimation. Maximum ValueComposites techniques remove most of the atmospheric andview-angle disturbances of daily NDvt values, but they do notprovide enough time coverage for fire danger estimation.Methods of radiometric and atmospheric conection of dailvimages should be improved, in ordlr to have more frequeniinformation on vegetation state.

o Thermal channels of avnRn should be considered in fuelmoisture estimation. As they are clearly related to vegeta-

568

tion's actual and potential evapotranspiration, the accumu-Iated difference of air and satellite derived temperaturesmight greatly improve current methods based on NDVI series,especially in dense forest stands.

r A more quantitative assessment of rqovt relations with fuelmoisture and fire danger indices should be performed.

AcknowledgmentsA great part of the material reviewed in this paper has beenprovided by the nnsons library while Emilio Chuvieco was avisiting scientist at the Canada Centre of Remote Sensing.Special thanks to Josef Cihlar, head of the Global MonitoringSection, for his suggestions in this topic. We also greatly ap-preciate the support provided by calvtr (University of Ne-braska, Lincoln), and specially to Prof, James Merchantduring the stay of Pilar Martin at this department, Sugges-tions and material provided by Marty Alexander, from For-estry Canada, have also been very valuable. Special thanks toBrian Tolk (CALMIT) for his linguistical assistance.

ReferencesBelward, A.S., 1991. Remote sensing for vegetation monitoring on re-

gional and global scales, Remote Sensing and Geographical In-formation Systems for Resource Management in DevelopingCountries (A.S. Belward and C.R. Valenzuela, editors), KluwerAcademic Publishers, Dordrecht, The Netherlands, pp. 169-182

Brown, J.F., T.R. Loveland, f. Merchant, B.C. Reed, and D.O. Ohlen,1993. Using multisource data in global land-cover characteriza-tion: concepts, requirements, and methods, Photogrammetric En-gineering & Remote Sensrng, 59:977-987.

Burgan, R.8., and R.C. Rothermel, 1984. BEIIAW: Fire Behavior pre-diction and Fuel Modeling System. Fuel Subsystem, USDA For-est Service, Ogden, Utah.

Burgan, R.E., and M,B. Shasby, 1984. Mapping broad-area fire poten-tial from digital fuel, terrain, and weather data, lournal of For-estry, 82i228-237.

Burgan, R.E., and R,A. Hartford, 7993. MonitoringVegetation Green-ness with Satellite Data, USDA Forest Service, INT-292, Ogden.

Cahoon, D.R., B.I. Stocks, J.S. Levine, W.R. Coffer, and K.P. O'Neill,1992, Seasonal distribution of African savanna fires, Nature,3 5 9 :812-815.

Calabri, G., and W.M. Ciesla, 7992. Global Wildland Fire Statistics,Food and Agriculture Organization, Rome.

Chandga, D., H. Deyong, S. Beien, W. Xinmin and P. Xizhe, 1991.Satellite monitoring of Da Xinan Lin forest fire and post-confla-gration change,Applications ofRemote Sensing inAsia andOceania, Environmental Change Monitoring (B. Murai, editor),Geocarto International, Hong-Kong, pp. 258-262.

Chung, Y.S., and H.V. Le, 1984. Detection of forest-fires smokeplumes by satellite imagery, Atmospheric Environment,7822143-2757.

Chuvieco, E., M.P. Martin, and L. Dominguez, 1993. Using AVHRRimages for fire mapping and fire growth monitoring in the medi-terranean forest, Proc. of the 12th Pecora Synposium on LandInformation from Space-Based Systems, Sioux Falls (in press).

Cihlar, J., L. St. Laurent, and f.A. Dyer, 1991. Relation between theNormalized Difference Vegetation Index and ecological varia-bles, .Remote Sensing of Environment, 351279-298,

Cohen, W.8., 1991. Response of vegetation indices to changes inthree measures of leaf water stress, Photogrammebic Engineering & Remote Sensrng, 57:195-202.

Crutzen, P.J., L.E. Heidt, J.P. Krasnec, W.H. Pollock, and W. Seiler,1979. Biomass burning as a source of atmospheric gases, Nafure,282:253-256.

Dabrowska, K., 1991. Application of AVHRR data for hydrological

PE&RS

studies of grassland in Poland, sth. AVHRR data user's meeting,Tromso, Norway, pP. 383-391.

Deblonde, G., and J. Cihlar, 1993. A multiyear analysis of the rela-tionship between surface environmental variables and NDVIover tlle Canadian landmass, Remote Sensing neviews, Ti!S!-1 7 7 .

Dozier, J., 1981. A method for satellite identification of surface tem-perature fields of subpixel resolution, Remote Sensing of Envi'ronment, 71t227-229.

Eidenshink, J.C., R,H. Haas, R.M. Zokaites, D.O. Ohlen, and K.P'Gallo, 1989. Integration of Remote Sensrng and GIS Technologtto Monitor Fire Danger in the Northern Gteat Plains, U.S. Geo'Iogical Survey, Final Report.

Eidenshink, J.C., R.E. Burgan, and R.H. Haas, 1990. Monitoring firefuels condition by using time series composites of AdvancedVery High Resolution Radiometer (AVHRR) data, Proc, ResourceTechnologlt 90, ASPRS, Washington, D.C., pp. 68-82.

Flannisan, M.D., and T.H. Vonder Haar, 19B6' Forest fire monitoringusing NOAA satellite AVHRR, Canadian lournal of Forest Re'seorc i ,16 :975-982.

Herfert, M.R., and K.P. Lulla, 1990. Mapping continental-scale bio-mass burning and smoke palls over the Amazon basin as ob-served from the Space Shuttle, Photogrammetric Engineering &Remote Sensrng, 56:136 7-13 73.

Hick, P,T., A.|. Prata, G. Spencer, and N. Campbell, 1986. NOAA'A\TIRR satellite data evaluations for areal extent of bushfiredamage in Western Australia, Proc. First Austtalian AVHRRConference, Perth, pp. 794-200.

Hirsch, S.N., R.F. Kruckeberg, and F.H. Madden, 1971. The bi-spec-tral forest detection system, Proc. 7th Inter. Symp. on RemoteSensing of Environment, Ann Arbor, pp. 2253-2259.

Illera, P., A. Fernandez, and J.L. Casanova, 7993. Proceedings of theInternational Workshop on Satellite Technologlt and Geographiclnformation Systems for Mediterranean Forest Mapping and FireManagement, Thessaloniki, (in press)'

Isaacson, D.L., H.G. Smith, and C.J. Alexander, 1982. Erosion hazardreduction in a wildfire damaged area, Remote Sensing for Re'source Management (Johannsen and Sanders, editors), Soil Con-servation Society of America, Ankeny, pp. 179-190'

Jijla,2., Z. Quingshan, D. Chaggong, L, Cheng, and D. Chaohua,1989. Detection of forest fire in Da Hinggan Ling region by mete-orological satellite, Acta Meteorologica Sinica, 3:562-568'

Kasischke, E.S., N.H. French, P. Harrel, N.L. Christensen, S.L. Ustin'and D. Barry, 1993. Monitoring of wildfires in boreal forestusing large area AVHRR-NDVI composite image data, RemoteSensing of Environment, 45:61-71.

Kaufman, Y.1., C.l. Tucker, and L Fung, 1990a. Remote sensing of bio-mass burning in the Tropics, Journal of Geophysical Research,9 5:992 7-993 9.

Kaufman, Y.1., R.S. Fraser, and R.A. Ferrare, 1990b. Satellite mea-surements of large-scale air pollution: methods, Joutnal of Geo'physical Research, 95 :9895-9909.

Kaufman, Y.J., A. Setzer, D. Ward, D. Tanre, B.N. Holben, P. Menzel,M.C. Pereira, and R. Rasmussen, 1992. Biomass burning airborneand spaceborne experiment in the Amazonas (Base-A), loutnalof Geophysical Re search, 9 7:14581-14599.

Kerr, Y., J. Imbernon, G. De Dieu, O. Hautecoeur, f'P. Lagouarde, andB. Seguin, 1989. NOAA-A\TIRR and its uses for rainfall andevapotranspiration monitorin g, International J ournal of RemoteSensing, 70:847-854.

Kidwell, K.B. (editor), 1991. NO,4,4 Polar Orbiter Data Users Guide,NOAAAIESDIS, Washington, D.C.

Langaas, S., 1992. Temporal and spatial distribution of savanna firesin Senegal and the Gambia, West Africa, 1989-90, derived frommulti-temporal AVHRR night images; International lournal ofWildland Fire, 2:2!-36.

Lee, T.F., and P.M. Tag, 1990. Improved detection of hotspots using

PE&RS

the AVHRR 3,7 pm channel, Bulletin American MeteorologicalSociety, 7 1 :77 22-77 3O'

L6nez, S., F. Gonzi l Iez, R. Llop, and J.M. Cuevas, 1991. An evalua-' tion of the utilitv of NOAA AVHRR images for monitoring forestfire risk in Spain, International lournal of Remote Sensing,72:784\-7857.

Malinereau, I.P., 1990. The contribution of remote sensing to thegl-obal monitoring of fires in tropical and subtropical ecosys-iems, Fjre in Tropicol Biota (J.G. Goldammer, editor), SpringerVerlag, Berl in, PP. 337-370.

Malingreau, J.P., G. Stevens, and L. Fel lows, 1985. The 1982-83 for-.rt fit"t of Kalirnantan and North Borneo. Satellite observationsfor detection and monitoring, Ambio' 74:374-327'

Malingreau, J.P., N. Laporte, and J.M. Gregoire, 1990. Exceptionalfiie events in the tiopics. Southern Guinee January 1987, Intet-national lournal of Remote Sensrng, 17:2727-2123'

Martin, M.P., and E. Chuvieco, 1993. Mapping and evaluation ofburned land from multitemporal analysis of AVHRR NDVI images,Proceedings of the Intemational Woil<shop -on S-atellite Technol-ogt and deographic Information Systems fol Mediterranean ForestMapping and Fire Management, Thessaloniki, (in press).

Matson, M., and J. Dozier,1981' Identi f icat ion of subresolut ion hightemperature sources using a thermal IR sensor, PhotogrammetricEngineering & Remote Sensing, 47:7377-1378.

Matson, M., and B. Holben, 1987. Satel l i te detection of tropicalburning in Brazil, International /ournal of Remote Sensing,8 :509-5 16 .

Matson, M., S.R. Schneider, B. Aldridge, and B. Satchwell, 1984'Fire Detection IJsing the NOAA Series Sotellites, NOAA Techni-cal Report NESDIS, Washington, D.C.

Mil ler, W.A., S.M. Howard, and O.G' Moore, 1986. Use of AVHRRdata in an Information System for fire management in the West-ern United Stales, 20th Intetn. Symp' on Remote Sensing of En'vironment, Nairobi, PP. 67-79.

Paltridge, G,W., and J, Barber, 1988. Monitoring gr-assland drylressan-d fite potential in Australia with NOAA'/AVHRR data, RemoteSensing of Environmenf, 25:381-394.

Prins. E.M., and W,P. Menzel, 1992. Geostationary satel l i te detectionof biomass burning in South America, International /ournal ofRemote Sensing, 73:27 83-27 99'

Riggan, P.]. , J.A. Brass, and R.N. Lockwood, 1993. Assessing f ire""emissions from tropical savanna and forest of central Brazil,

Photo grammetric Engineering & Re mote Se nsing, 5 I : 1 009-1 0 1 5'

Robinson, J.M., 1991, Fire from space: global f i re evaluation usinginfrared remote sensinS, lnternational lourna) of Remote Sens-ing, 72:3-24.

Robock, A., 1988. Surface temperature effects of forest fire smokeplumes, Aerosols and Climate (P'V. Hoobs and M.P' McCormick,Ldtors;, Deepak Publ', Hampton, Virginia' pp.435-442'

Sadowski, F.G., and D.E. Westover, 1986. Monitoring the fire-dangerhazard of Nebraska rangelands with AVHRR data, Proc. 10th Ca-nadian Symp. on Remote Sensing, Edmonton, pp. 355-363'

Seguin, 8., J.P. Lagouard, and M. Savanne, 1991. The assessment of"

regional watei conditions from meteorological satellite thermalinfrared data, Remote Sensing of Environment' 35:141-148'

Sellers, P.J, 19S9. Vegetation-Canopy Spectral reflectance and Bio-physilal processes, Theory and Applications of Optical RemoteSensing (G. Asrar, editor), Wiley, New York, pp. 297-335'

Setzer, A.W., and M.C. Pereira, 1991a. Operational detection of firesin Brazil with NOAA-AVHRR, Proc. 24th. International Symp'on Remote Sensrng of Environmenf, Rio de Janeiro, pp. 469-482'

1991b. Amazonia biomass burnings in 1987 and an estimateof their tropospheric emissions, Ambio, 2O:f9-23'

Smith, R,8., and P.W. Woodgate, 1985. Appraisal of f i re damage andinventory for timber salvage by remote sensing in mountain ashforest in Victoria, AustroJjon Forestry, 481252-263.

Stearns, ] .R., M. Zahniser, C.E. Kolb, and B.P. Sandford, 1986' Air-

569

borne infrared observations and analysis of a large forest fire,Applied Optics, 25:2554-2562.

Van Wagner, C.E., 1987, Development and Structure of the CanadianForest Fire Weather Index System, Canada Forest Service, Ottawa.

Van Wagner, C.E., B.J. Stocks, B.D. Lawson, M.E. Alexander, T.f.Lynham, and R.S. McAlpine, 1992. Development and Structureof the Canadian Forcst Fire Behavior Prediction Sysfem, For-estry Canada, Ottawa.

Walsh, S.J., 1987. Comparison of NOAA-AVHRR data to meteorologi-

cal drought indices, Photogrammetric Engineering & nemoteSensing, 53:1069-1074.

Watson, C.E., f. Fishman, and H.G. Reitchle, 1990. The significanceof biomass burning as a source of carbon monoxide and ozonein the Southern tropics: a satellite analysis, lournal of Geophysi-cal Research, 95:16443-16450.

Werth, L.F., R.A. McKinley, and E.P. Chine, 1985. The use of wi ld-land fire fuel maps produced with NOAA-AVHRR scanner data,Pecora X Symposium, Fort Collins, pp. 326-331.

1993 ACSM/ASPRSTECHNTCAL PAPERS

"Looking to the Future with an Elre on the Past"

Proceedings of the 1993 ACSM/ASPRS Annual Meetingheld in New Orleans, Louisiana, February 1993.

Even if you couldn't attend the 1993 ASPRS/ACSMAnnual Convention in New Orleans. this 3 volume setcan help you discover what's new in the industry.Learn the latest research and theory in GIS, RemoteSensing, Photogrammetry, and Cartography.

yoL. r - AcsMStock # 4933-1. pp. 422.

r The 1990 Census and GIS: A Case Study of Day-care Accessibility in East Baton Rouge Parish

r Real-Time Kinematic GPS for Surveying: Centi-meters in Seconds

r Examination of the Validity of State Plane Projec-tion Computations in PLSS Applications

r Surveying and Mapping Education: An Interdisci-plinary Approach

r Disputes Between Adjoining Landowners

VOL. 2 - ASPRS: Bemote SensingStock # 4933-2. pp. 474.

r Automated Classification of Landsat Data Using anInduction-Based Model

r AVHRR for Forest Mapping: National Applica-tions and Global Implications

r Development of Modular Courses for RemoteSensing, Photogrammetry, and LIS/GIS Instruction

Forest and Wetland Monitoring Using an Integrat-ed GISChange Detection Using Multi-Temporal FeatureSpace With Digital TM Data

VOL.3 - ASPRS: GIS'Photogramrnetw. and Modeling

Stock # 4933-3. pp. 458.

Spatial Autocorrelation in Remotely-Sensed andGIS DataFeature Mapping in a Photogrammetric GISApplication of Neural Networks in Forest Eco-logical ClassificationPractical Guidelines for the use of GPS Photogram-metryAn Approach to Automated Terrain ClassificationFrom Digital Elevation Model DataIntegrating a Resource Assessment Model intoARC/Info GIS: A Spatial Decision Support Sys-tem Development

1993. Three volumes. 1,354 pp. $25 each(soficover); l.tPRS Members $15 each. Stock #4933-I, 4933-2, 4933-3.

To Order, See The ASPRS Store.

I

I