h b c

c Department of Geography and Geology, University of Alcd Instituto de Economa, Geografa y Demografa (IEGD), C28037 Madrid, Spaine Center for Spatial Technologies and Remote Sensing (CST

ushres,ratory, BSchool of

LFMC to re behavior at intermediate scales. Changes in LFMC have both direct (liquid water absorption)

. . . . . . . . . 456

. . . . . . . . . 456

. . . . . . . . . 457

. . . . . . . . . 458

Remote Sensing of Environment 136 (2013) 455468

Contents lists available at SciVerse ScienceDirect

Remote Sensing of Environment5. Estimation of LFMC from satellite data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4585.1. Physical basis for a RS-based estimation of LFMC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458

5.2. Statistical versus physical model-based approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4601. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2. The importance of LFMC for re risk assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3. Measures of vegetation water content and relationships with LFMC . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. Field measurements of LFMC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .Spectral indicesFire occurrenceWildreFire risk

and indirect (pigment and structural changes) impacts on spectral reectance. The literature is dominatedby studies that have used statistical (empirical) and physical model-based methods applied to coarse resolu-tion data covering the visible, near infrared, and/or shortwave infrared regions of the spectrum. Empiricalrelationships often have the drawback of being site-specic, while the selection and parameterizationof physically-based algorithms are far more complex. Challenges remain in quantifying error of remotesensing-based LFMC estimations and linking LFMC to re behavior and risk. The review concludes with alist of priority areas where advancement is needed to transition remote sensing of LFMC to operational use.

2013 Elsevier Inc. All rights reserved.

Contents5.3. Fine versus coarse spatial and spec5.4. Input data used to estimate LFMC

Corresponding author. Tel.: +61 2 6246 5742.E-mail address: marta.yebra@csiro.au (M. Yebra).

0034-4257/$ see front matter 2013 Elsevier Inc. Allhttp://dx.doi.org/10.1016/j.rse.2013.05.029for estimating LFMC, with particular concern towards the operational use of LFMC products for re riskassessment. Relationships between LFMC and re behavior have been found in fuel ignition experimentsand at landscape scales, but the complexity of re interactions with fuel structure has prevented linkingWater thicknessDry matter contentRadiative transfer modelsf Centre for Environmental Risk Management of Bg USDA-ARS Hydrology and Remote Sensing Laboh Ecosystems and Environment Research Centre,

a r t i c l e i n f o

Article history:Received 10 March 2013Received in revised form 29 May 2013Accepted 29 May 2013Available online 22 June 2013

Keywords:Vegetation water contentLive fuel moisture contentchnological Hazards, University of Utah, 270 S Central Campus Dr, Rm 270, Salt Lake City, UT 84123, USAal, Calle Colegios 2, 28801 Alcal de Henares, Spainentro de Ciencias Humanas y Sociales (CCHS), Consejo Superior de Investigaciones Cientcas (CSIC), Albasanz 26-28,

ARS), University of California, 250-N, The Barn, One Shields Avenue, Davis, CA 95616-8617, United StatesUniversity of Wollongong, Australialdg 007 BARC-West, 10300 Baltimore Avenue, Beltsville, MD 20705, USAEnvironment and Life Sciences, University of Salford, Salford M5 4WT, UK

a b s t r a c t

One of the primary variables affecting ignition and spread of wildre is fuel moisture content (FMC). LiveFMC (LFMC) is responsive to long term climate and plant adaptations to drought, requiring remote sensingfor monitoring of spatial and temporal variations in LFMC. Liquid water has strong absorption features inthe near- and shortwave-infrared spectral regions, which provide a physical basis for direct estimation ofLFMC. Complexity introduced by biophysical and biochemical properties at leaf and canopy scales presentstheoretical and methodological problems that must be addressed before remote sensing can be used foroperational monitoring of LFMC. The objective of this paper is to review the use of remotely sensed dataa CSIRO Water for a Healthy Country Flagship, CSIRO Land ab Department of Geography and Center for Natural and Tend Water, Canberra ACT 2601, AustraliaF. Mark Danson , Yi Qi , Sara Jurdao

Marta Yebra a,, Philip E. Dennison b, Emilio ChuviecoReview

A global review of remote sensing of live fuel moisture content forre danger assessment: Moving towards operational products

c, David Riao d,e, Philip Zylstra f, E. Raymond Hunt Jr. g,

j ourna l homepage: www.e lsev ie r .com/ locate / rsetral resolution input data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461

rights reserved.

..ce

.

.roaver.......

456 M. Yebra et al. / Remote Sensing of Environment 136 (2013) 455468(1000 ha) in the Sydney Basin Bioregion, New SouthWales, Australia,through drying of fuels over extended areas (Bradstock et al., 2009). The2003 Sant Lloren Fire in Catalonia, Spain, also occurred following aperiodwhere very hot, dry Saharan air produced criticalmoisture stressin plants (Oliveras et al., 2009). Very low LFMC and dry, warm, SantaAna winds contributed to large bushre events in southern Californiain 2003 and 2007 (Keeley, 2004; Keeley et al., 2009).

In grasses, live and dead fuel loads are variable through time as se-nescence converts live fuel to dead fuel. The proportion of herbaceousfuel that is dead is important in determining the probability of ignitionand rate of spread (ROS) of wildres (Cheney et al., 1998), which hasled to the use of vegetation indices for estimation of dead versus livefuels proportion to compute re danger potential (Burgan et al., 1998;Newnham et al., 2011).

Obtaining spatially comprehensive and temporally frequent estimatesspecically for LFMC ismore problematic. Field sampling and gravimetric

2. The importance of LFMC for re risk assessment

The effects of LFMC on re behavior are complex and not alwayseasy to identify empirically. Elevation of fuel temperature to the com-bustion point requires loss of water through evaporation; thus, higherLFMC should increase the time to ignition and decrease the probabil-ity of ignition. LFMC has therefore been demonstrated to be a primarydeterminant of time to ignition across multiple species at low to mod-erate temperatures (Xanthopoulos & Wakimoto, 1993), exhibiting ageometrically decreasing effect as temperatures are raised (Zylstra,2011a) until the effect is negligible at temperatures correspondingwith the hottest parts of a ame (Fletcher et al., 2007). LFMC hasalso been shown to correlate negatively with ame length from burn-ing leaves (Zylstra, 2011a).

The way in which these factors affect re behavior and risk is com-plex and the subject of debate. Plucinski et al. (2010) found that in6. Challenges in estimating FMC from satellite data . . . . . . . . .6.1. Theoretical challenges . . . . . . . . . . . . . . . . . .

6.1.1. Decoupling the impact of EWT from DMC on reectan6.1.2. Confounding effects of LAI and other canopy variables

6.2. Methodological challenges . . . . . . . . . . . . . . . .6.2.1. Site-specicity of statistical methods . . . . . . .6.2.2. Constraining parameters of physical model-based app6.2.3. Dependence of physical model-based accuracies on in

7. Main problems for the operational use of RS derived LFMC models . .7.1. LFMC estimation error . . . . . . . . . . . . . . . . . .7.2. LFMC and re behavior . . . . . . . . . . . . . . . . . .7.3. LFMC and re risk . . . . . . . . . . . . . . . . . . . .

8. Concluding remarks . . . . . . . . . . . . . . . . . . . . . . .Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . .References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1. Introduction

Fuel moisture content (FMC), the mass of water contained withinvegetation in relation to the dry mass, is a critical variable affecting reinteractions with fuel. FMC is one of the primary variables in many rebehavior prediction models and re danger indices, as it affects ignition,combustion, the amount of available fuel, re severity and spread,and smoke generation and composition (Anderson & Anderson, 2010;Deeming et al., 1978; Finney, 1998; McArthur, 1967; Nelson, 2001;Plucinski et al., 2010; Viegas et al., 1992). FMC is usually separated intodead (DFMC) and live (LFMC) components. In many re risk models,DFMC is empirically determined from weather variables, diameter ofthe material and biochemical compositions (Viney, 1991).

LFMC ismuchmore difcult to estimate frommeteorological indicesthan DFMC, because living plants have a variety of drought adaptationstrategies (Viegas et al., 2001) and can draw upon moisture stored inthe soil. The Keetch-Byram Drought Index (KBDI, (Keetch & Byram,1968)) has been indirectly correlated with LFMC for some species(Dimitrakopoulos & Bemmerzouk, 2003; Xanthopoulos et al., 2006),while other species appear to be driven more by medium-term meteo-rological conditions or phenology (Castro et al., 2006; Pellizzaro et al.,2007; Zylstra, 2011a). Even though DFMC across a landscape is a deter-minant of landscape connectivity and therefore the potential area burnt(Caccamo et al., 2012a), the correlation of intense re behavior withmore persistent indicators such as deep soil dryness and/or extendedheatwave conditions suggests that the moisture conditions of livefuels are also important determinants of re behavior. For instance,the Black Saturday bushres of February 2009 inVictoria, Australia, dur-ingwhich 173 people died andmore than 3500 houses were destroyed,occurred after weeks of extreme record-breaking high temperatures,which dried many plants to critically low levels (Gellie et al., 2010).Drought had a major inuence on the incidence of large bushresmethods (Lawson & Hawkes, 1989) are locally accurate but very costly.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463ches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463sion procedure . . . . . . . . . . . . . . . . . . . . . . . . . . 464. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466

Furthermore, generalization of thesemeasurements to landscape, region-al, or global scales is not feasible from eld sampling. Remotely sensed(RS) data provide the opportunity to estimate LFMC over large areas atne spatial and temporal resolutions, but as illustrated in this review,these data require calibration and validation. The initial hypothesis be-hind satellite-based estimation of LFMC is that the impact of LFMC varia-tion on the RS signal is strong enough to be discriminated from otherfactors affecting spectral variation such as the atmosphere, soil back-ground, solar and sensor geometry, and other plant characteristics. Sever-al studies have been published in recent years to test this hypothesis(Bowyer & Danson, 2004; Ceccato et al., 2001; Gillon et al., 2004;Riao et al., 2005) and multiple methods have been developed to es-timate LFMC from both coarse and ne spatial resolution remotesensors (e.g. Chuvieco et al., 2002; Peterson et al., 2008; Wang et al.,2013; Yebra et al., 2008b). However, despite the success of thesemethodsfew of the resulting products have been operationally integrated intowildre danger systems to date.

The objective of this paper is to review the use of RS data for esti-mating LFMC to assess its potential, with the anticipation that theseRS-based products will soon become useful for operational use. To ad-dress this objective we will cover the following aspects: (i) importanceof estimating LFMC in the context of re risk assessment (Section 2);(ii) methods of measuring vegetation water content and their relation-ships with LFMC (Section 3); (iii) eld data collection challenges andrecommendations (Section 4); (iv) models that have been developedto derive LFMC from RS data as well as a brief review of the physicalbases for a RS based estimation of LFMC (Section 5); (v) challenges anddevelopments in satellite-based estimation of this variable (Section 6);(vi) obstacles for the operational use of LFMC models and products(Section 7); and (vii) priorities for research and applications withinthis eld (Section 8).laboratory recreations of shrub res, the most important factors for

457M. Yebra et al. / Remote Sensing of Environment 136 (2013) 455468spread initiation were LFMC and the presence of litter, and that shrubdensity has a greater inuence at lower moisture contents (see gure3 Plucinski et al. (2010)). Similar effects on re propagation in shrubfuels have also been identied in eld conditions (Davies et al., 2009;Weise et al., 2005).

While these studies demonstrate the importance of LFMC in deter-mining the incidence and scale of res, specic inuences on someareas of re behavior are less obvious. In a continuous fuel bed, ROScan be expected to decrease as FMC increases, since re velocityis equal to the distance to ignition divided by time to ignition. Thisis widely accepted for DFMC in a range of re behavior models(e.g. Burrows, 1999; Gould et al., 2007; McArthur, 1962; Rothermel,1972), and Viegas et al. (in press) demonstrated that ROS in amixed bed of live and dead fuels was explained by the compositeFMC derived from both live and dead values. The relationship is farless obvious in a discontinuous fuel bed however, and in their reviewof historical heathland and crown re experiments Alexander andCruz (2012) found an inuence in laboratory experiments but notin the eld, proposing that the discrepancy was potentially due tothe larger heat uxes present in eld conditions. In this context,Zylstra (2011a) noted that re behaviors in discontinuous fuels arenot directly determined by causal factors, but are rather emergentproperties of the interactions between the ammability of plant parts,forest geometry and exogenous factors, and are therefore subject tonumerous complex behaviors and feedbacks that unless specicallyaccounted for in experimental process will confound any clear results.This model also suggests that greater heat uxes do not override the in-uence of LFMC on leaf ignitability, but rather have the effect of ignitingleaves at a greater distance ahead of the burning fuels. This implies thatthe laboratory experiments identied an inuence of LFMC becausethey were able to control for a wider range of variables relevant to thescale of the fuel complexes examined.

Although the immediate inuence of LFMC on re behavior maybe difcult to capture empirically, the overall effect becomes appar-ent as an emergent property when the relationship is studied on alandscape scale. Multiple studies using re histories fromMediterra-nean ecosystems have demonstrated relationships between LFMCand actual re occurrence and size. Schoenberg et al. (2003) exam-ined monthly trends in chaparral LFMC in Los Angeles County,California, USA, and found that burned area increased when LFMCdropped below 90%. Dennison et al. (2008) investigated relation-ships between chamise (Adenostoma fasciculatum) LFMC and resize and occurrence in the Santa Monica Mountains of California.They found that while small res occurred across a wide range ofLFMC values (59139%), large res (>10 km2) only occurred whenchamise LFMC was below 77%. Dennison and Moritz (2009) expandedthis analysis to LosAngeles County, California, and found similar relation-ships between the occurrence of large res and chamise LFMC, with aLFMC threshold near 79%.

Chuvieco et al. (2009) also showed that lower LFMC increases reoccurrence for shrublands and grasslands in the Cabaeros NationalPark located in central Spain. Moisture variations in grasslands werefound to be good predictors of the number of res and total burnedsurface, while both the total burned area and the occurrence oflarge res were more sensitive to moisture variations of two shrubs(Cistus ladanifer L. and Rosmarinus ofcinalis L.). Jurdao et al. (2012)related seven years of MODIS thermal anomalies (MOD14, (Giglioet al., 2003)) with LFMC calculated from Advanced Very High Resolu-tion Radiometer (AVHRR) data using an empirical algorithm developedby Garcia et al. (2008). They found signicant differences between reand non-re pixels and several spatial and temporal variables derivedfrom LFMC.

Outside Mediterranean ecosystems, there have been few compar-isons between LFMC and re occurrence. Maki et al. (2004) usedburning areas obtained from AVHRR to conrm the relationships be-

tween a vegetation dryness index (VDI) and ignited pixels and rebehavior. A hypothetical trapezoidal shape was rst dened by thenormalized difference vegetation index and normalized differencewater index of the research region bounded by four vertexes oftheir maximum and minimum values using SPOT/VEGETATION data.To estimate water status per fractional vegetation cover, the VDIwas calculated as 1 subtracted by the ratio between a point's distanceto the trapezoidal left side and distance to the right side. The VDIvalues of re-spread pixels were higher than those of the non-re-spread pixels and the re front length was strongly related to VDI.Drought conditions have been found to be closely related to the areaburned by unplanned re in the Sydney sandstone area (Bradstocket al., 2009), and Caccamo et al. (2011) demonstrated that this wasclosely tied to greater landscape connectivity resulting from reductionsin LFMC.

Overall then, there is strong empirical evidence demonstrating theeffect of LFMC on re occurrence and the extent of its impact, and labo-ratory evidence demonstrating the direct relationship to re behavior.Field evidence for this connection is inconclusive; however there arephysical arguments that explain why the connection is not alwaysclear, and the observed effect of LFMC on re size demonstrates thatsuch relationships must exist as emergent properties even if not yetcaptured experimentally.

3. Measures of vegetation water content and relationships with LFMC

LFMC is a ratio of the mass of water contained in a live plant to thetotal dry mass of the plant. Changes in LFMC are driven by bothchanges in the moisture status of the leaves and seasonal changesin dry matter, which represents the amount of available fuel to beburned. Field sampled LFMC is calculated as:

LFMC mfmdmd

1

wheremf is the mass of the fresh collected sample andmd is the drymass of the same sample.

When measuring moisture content in RS-based studies, the mostcommonly used metric is the amount of leaf water divided by its area(A), which is called the equivalent water thickness (EWT, g cm2):

EWT mfmd

A: 2

EWT is more relevant to RS studies as the water absorption of in-coming radiation is directly related to the depth of the water layer,particularly in the near infrared (NIR: 700 to 1400 nm) and short-wave infrared (SWIR: 1400 to 2500 nm) spectral bands.

EWT and FMC are linked through the dry matter content (DMC,g cm2), which is the inverse of the specic leaf area (SLA, cm2 g1),a more common term in plant physiology and functional ecology(Garnier & Navas, 2012). DMC describes the areal expression of cellu-lose, lignin and other components of plants that remain after drying.DMC is dened as:

DMC mdA

: 3

DMC can be used to normalize thewatermass per area and calculateFMC following:

LFMC mfmd

AmdA

0BBBB@

1CCCCA

EWTDMC

: 4Eqs. (1) to (4) also apply to the calculation of DFMC.

458 M. Yebra et al. / Remote Sensing of Environment 136 (2013) 455468Since only EWT is directly related to radiation absorption by water,the estimation of LFMCpartially relies on the accuracy of DMC estimates(Riao et al., 2005). DMC does not have strong specic absorption fea-tures (Fourty & Baret, 1997), and therefore the estimation of LFMCfrom RS data is more challenging than estimation of EWT.

LFMC calculated in Eqs. (1) and (4) is independent of the leaf areaand the amount of leaves that constitute the canopy or leaf area index(LAI). However, this last parameter is extremely helpful when calcu-lating the total canopy water content in the leaves per unit groundarea (CWC, g m2) which is commonly computed by multiplyingthe mean leaf EWT of a sample area by the LAI:

CWC EWT LAI: 5

Consequently, CWC allows scaling leaf water content to the canopylevel and inuences the amount of reected radiation measured by re-mote sensing. CWC has been related to LFMC (Zarco-Tejada et al., 2003)andmay be correlatedwith LFMC if species happen to have similar DMC(Chuvieco et al., 2003; Yebra et al., 2008b).

Senescence in grasses results in conversion of live fuel to dead fuelover time. LFMC may not be separable from DFMC in partially curedgrasses, and estimated LFMC may represent an aggregate of both mea-sures. In the sections that follow, grassland LFMC is considered as theaverage moisture content of both live and dead fuel components, andthe degree of curing is not explicitly accounted for.

4. Field measurements of LFMC

Field measurements of LFMC are essential for calibration and val-idation of satellite-based methods for estimating LFMC. Field mea-surement of LFMC is straightforward requiring destructive samplingof a representative sample of leaf material which is then weighedfresh, oven dried, and reweighed to determine dry weight. Howeverthe sampling design, in terms of the number of samples and their lo-cation, must take into account the variability in LFMC that may occurwithin individual plants in both horizontal and vertical directions,and the variability between plants within the sampled area. Samplingleaves from broadleaved trees may be simple, but sampling leaves insome shrub species, where there is a mix of woody and leafy material,may be more complex.

The key difculty in making LFMC eld measurements is ensuringthat they are spatially comparable to the RS data and this problem hasoften limited scientists' ability to develop methods for monitoringLFMC from space. The primary limitation is the spatial scale of sam-pling. RS data suitable for monitoring LFMC are typically collected atspatial scales ranging from 0.1 to 100 ha, while LFMC is typically sam-pled at smaller scales ranging from 0.01 to 0.1 ha. Field measurementof LFMC for comparison with satellite-based measurements requires:

(i) Site metadata, including sampling locations and dates, shouldalways be recorded. The coordinates of the center of the sam-pling site is often sufcient, but a polygon around the perimeterof the sampling site is much more useful. As much as possible,the same site should be sampled to create a long-term recordof LFMC for that site. A change of a few meters that results in adifferent slope or aspect being sampled could make time seriescomparison much more difcult.

(ii) The area surrounding LFMC sampling sites should be as homo-geneous as possible in vegetation and topography over an areaof 100 ha to make the comparison between LFMC and RS datafeasible. In some regions, this might be impossible but sites se-lection should still be optimized.

(iii) Sites should ideally be dominated by one vegetation type, andmixed vegetation sites should be avoided. For each site, all dom-inant species presented should be sampled and their fractional

coverage should be recorded.Databases have been created to collect eld-sampled LFMC data(e.g. the U.S. National Fuel Moisture Database (http://www.wfas.net/nfmd/public/index.php, last accessed February 2013) and theLFMC database developed from 1996 to 2010 by the University ofAlcal (Spain) (Chuvieco et al., 2011)). However a global network toroutinely measure LFMC following a common sampling methodologyor protocol at the appropriate spatial and temporal resolutions hasnot been established. This observational limitation makes it difcultto obtain LFMC at a sufcient number of locations within a shorttime interval to calibrate and validate satellite data products. A recentdevelopment in the application of terrestrial laser scanning to mea-sure vegetation EWT may in future provide a rapid non-destructiveapproach to obtain LFMC in three dimensional space over spatial rangesof 10100 m based on dual-channel laser scanner measurements with-in a sampling plot (Gaulton et al., 2013).

A common sampling protocol should address issues related to (i)plot size, (ii) the best time to acquire the samples, (iii) what to dowhen precipitation occurs, (iv) how many samples per plot to take,(v) which kind of material to harvest (new and old leaves, combineleaves and shoots, include grassland roots, etc.), how much materialand how many samples to take, (vi) information about the weightingprocedure (if the samples were weighted on the eld or in the lab andthe weighing device accuracy), (vii) details about the drying process(drying device, drying time and temperature) and (viii) the mate-rial used to seal the samples for transportation to the lab, (see fore.g. Desbois et al. (1997) and Zahn and Henson (2011)). Furtherwork is needed to understand spatial variability in LFMC across plant,local, and regional scales to better inform sampling strategy.

5. Estimation of LFMC from satellite data

Various methods have been developed to estimate LFMC fromRS data, which may be broadly classied into statistical (empirical)(Caccamo et al., 2012b; Dennison et al., 2005; Garcia et al., 2008;Peterson et al., 2008) and physical model-based approaches (Colomboet al., 2008; Yebra & Chuvieco, 2009a; Yebra et al., 2008b; Zarco-Tejadaet al., 2003). The type of sensor used to acquire RS data is an importantconsideration for either method. Table 1 summarizes some examplesof application of RS data for LFMC estimation. Most of the studies inTable 1 are focused on high temporal resolution sensors with a fewspectral bands, such as MODIS. Spectral vegetation indices (VI) are anefcient means of obtaining empirical information from multispectralsensors; however physical model-based methods also benet fromVI. Typically, hyperspectral sensors are used for physical model-basedmethods, since the numerous narrow spectral bands offer the possibilityof novel VI.

The decision to use any combination of method, sensor and spectralinformation to estimate LFMC depends on the objectives of the research.Before going into detail we will briey review the physical basis for aRS-based estimation of LFMC.

5.1. Physical basis for a RS-based estimation of LFMC

In the solar spectral domain, water has a direct effect on spectralreectance through absorption of radiation within the NIR andSWIR spectral regions. Depending on tissue water content, reec-tance is thus reduced to a varying extent within the water absorptionfeatures centered on 970, 1200, 1450, 1940, and 2500 nm (Knipling,1970; Thomas et al., 1971; Tucker, 1980). RS methods which makeuse of these absorption features are considered direct estimationtechniques. However, changes in leaf pigment concentrations andleaf internal structure co-vary with LFMC, and produce changes invisible and NIR reectance that may be correlated with LFMC(Fig. 1). When plants are under water stress, depletion of chlorophyllmay produce a decrease in reectance in the visible bands, especially

in the red end of the visible spectrum. When leaves wilt during

cillary information used, and vegetation types are listed, alongwith the corresponding reference.UT is look-up table, # sites is the number of sites used in the study, ST is the surface temperature,a Landsat Thematic Mapper-like broad waveband and Ccov is the canopy coverage.

Vegetation type Locations #sites

Independentvalidation

Reference

Mediterranean grasslandsand shrublands

Spain 7 Yes Chuvieco et al.(2004c)

Chaparral California, U.S.A. 3 No Stow et al.(2005)

Chaparral and coastalshrubland

California, U.S.A. 12 No Dennison et al.(2005)

Chaparral California, U.S.A. 4 Yes Stow et al.(2006)

Chaparral California, U.S.A. 14 No Roberts et al.(2006)

Pine stands and hard wooddominated areas

Georgia, U.S.A. 2 No Hao and Qu(2007)

Chaparral California, U.S.A. 11 Yes Stow andNiphadkar(2007)

459M. Yebra et al. / Remote Sensing of Environment 136 (2013) 455468Table 1Examples of studies using RS data to estimate LFMC.Methodologies, sensors, spectral and anIn the case of physical model-based (RTM)methodologies the inversemethod is indicated. LDOY is a function of the day of the year, LCT is the land cover type, TM is the reectance in

Method Sensor Spectral information Ancillaryinformation

Empirical AVHRR NDVI ST, DOY, LCT

Empirical MODIS NDWI, VARI None

Empirical MODIS NDVI, NDWI None

Empirical MODIS B1, B2, B3, B4, B5, B6, B7, NDVI,SAVI, EVI, NDWI, NDII6, NDII7,NDGR, VARI

None

Empirical MODIS NDVI, EVI, VARI, VIg NoneAVIRIS NDII6, NDII7, NDWI, WI, EWT

RTM-Simulations MODIS B1, B2, B3, B4, B5, B6, B7 None

EmpiricalEmpirical MODIS NDII6, NDWI, VARI Maxmin

scaled NDII6, NDWI, VARINonedehydration and senescence, many of the reective interfaces ofleaves are eliminated as internal air space is reduced and cell wallscome together, which reduces NIR reectance (Knipling, 1970).

Additionally, LFMC changes affect plant canopy temperature becausewater availability is a critical parameter in plant evapotranspiration.When the plant dries, transpiration and latent heat transport is reduced,which increases sensible heat (Kozlowski et al. 1991). As a result of thisrelation, the difference between air and canopy surface temperature (ST)should be clearly related to plant water content and to water stress(Boulet et al., 2007; Gonzalez-Dugo et al., 2012; Moran et al., 1994;Rahimzadeh-Bajgiran et al., 2012; Vidal et al., 1994).

Finally the microwave region has also been found to be sensitiveto water content of plant and soil moisture. Changes in dielectricconstant associated with the water content of biomass components(leaves, branches, and stems) impact the radar backscatter measure-ments and introduce larger variability when compared with dry biomass(Way et al., 1991). However the use of microwave images for retrieval ofplant water content is more complex than with optical sensors, and pre-sents different factors of potential confusion, such as vegetation biomass,height, topographic position or roughness (Beaudoin et al., 1995).

Empirical MODIS NDVI, EVI, VARI, VIgreen,NDII6, NDII7, NDWI

None Chaparral and coastalsagescrub

California, U.S.A. 14 No Peterson et al.(2008)

RTM-LUT MODIS NDVI, SAVI, EVI, GEMI, VARI,NDII6, NDWI, GVMI

LAI, DMC, LCT Mediterranean Grasslandsand shrublands

CabaerosNational Park,Spain

5 Yes Yebra et al.(2008b)

EmpiricalEmpirical AVHRR NDVI ST, Function of

day of year, LCTMediterranean Shrublands Spain 7 Yes Garcia et al.

(2008)Mediterranean Grasslands

RTM-LUT MODIS B1, B2, B3, B4, B6, B7 LAI, plantecologicalinformation

Mediterranean Shrublands Spain 12 Yes Yebra andChuvieco(2009b)

RTM-LUT MODIS B1, B2, B3, B4, B6, B7, NDII6 None Mediterranean Woodlands Spain 5 Yes Yebra andChuvieco(2009a)

Empirical MODIS VARImaxmin and NDII6maxmin None Shrubland, heathland andsclerophyll forest

South-easternAustralia

8 Yes Caccamo et al.(2012b)

Empirical AISA Eagle andHawk sensors

Reectance, rst derivatives,MSI, WI, NDWI

None Calluna vulgaris andgrassland

Central Pennineuplands, U.K.

10 No Al-Moustafa etal. (2012)

TM5/TM7, NDVI, NDII6Empirical MODIS NDVI, NDWI, CWC None Gambel Oak Sagebrush Utah, U.S.A. 10 No Qi et al. (2012)RTM Simulations NDWI, NDII6, SWRI, RMSI,

NDMI, NDTI, CAI, NDLI, NDNI,LCA, SINDRI, DMCI, ratio ofwater index and dry-matterindex

None Quercus alba, Acer rubrumand Zea mays

None None Yes Wang et al.(2013)

EmpiricalRTM, LUT MODIS B1, B2, B3, B4, B5, B6, B7, NDII6 LAI, Ccov Woodlands Spain 19 Yes Jurdao et al.

(2013)

Fig. 1. Reectance spectra for vegetation canopies with different LFMC values. Spectrawere collected by the Airborne Visible Infrared Imaging Spectrometer (AVIRIS) over aplot dominated byAdenostoma fasciculatum in southern California, U.S.A. The approximatespectral extent of the rst seven MODIS bands is also shown in gray.

5.2. Statistical versus physical model-based approaches

Empirical methods for LFMC estimation are commonly based onstatistical tting between eld-measured LFMC and spectral mea-sures based on reectance data.

The alternatives to empirical approaches are those based on simu-lation models, which use physical models to estimate a parameter(water content in this case), based on a set of simulation scenarios.The estimation is commonly based on comparing the observed reec-tances of every pixel to those simulated in a look-up table, assigningto each pixel the parameters of the most similar simulated spectrum.The concept of similarity is commonly formalized in radiative transfermodel (RTM) inversion approaches using a merit function, whichimplies minimizing the differences between the observed andmodeledreectance. Several merit functions have been formulated and used forRTM inversion (Table 2). The RS inputs for the inversion can be eitherspectral waveband indices or both.

Yebra et al. (2008b) compared the performance of statistical andphysical approaches to derive LFMC of Mediterranean grassland and

460 M. Yebra et al. / Remote Sensing of Environment 136 (2013) 455468shrubland species from MODIS reectance measurements. Both ap-proaches provided good estimations of LFMC for both vegetation func-tional types. The physical approach based on simulated data was morecomplex to parameterize, but offered lower standard errors of estima-tion of LFMC of 29.5% and 12.6% for grasslands and shrublands versus30.1% and 17.5% for the statistical approach. The physical approachalso proved more robust when several calibration samples were select-ed, since the coefcients of the calibrated models did not signicantlyvary with the samples used and consequently were expected to havea greater generalization power. This hypothesis was tested by Yebraet al. (2008a)with groundmeasurements obtained at other areas dom-inated by grasslands in Spain and Australia. The results showed thatboth physical and statistical models offered similar accuracy levelswhen applied to grassland with analogous types of vegetation to thecalibration site with an average RMSE of LFMC of 42% and 36% for statis-tical and physical model approaches respectively. Nevertheless, thephysical model offered greater accuracy than the statistical modelwhen themodels were applied to grasslandswith different characteris-tics to those of the calibration sites with an average RMSE of 50% and15%, for the statistical and physical model, respectively.

5.3. Fine versus coarse spatial and spectral resolution input data

Coarse spatial resolution, globally available AVHRR or MODIS data arethe most common choice for estimating LFMC from RS data, as they pro-vide high enough temporal resolution for operational applications. Therst studies were carried out in the 1980s and 1990s, when strong

Table 2Merit functions formulated in the literature and used for RTM inversion. i,Obs and i,modare the observed and modeled reectivity in each band i respectively, n the number ofbands considered in the comparison and v and w are the observed and modeled spectrarespectively, both of them considered as an n-dimensional feature vector.

Function Formulation Reference

Absolute errorAE

Xni1

i;Obsi;mod Koetz et al. (2005)

Square error SE Xni1

i;Obsi;mod 2 Zarco-Tejada et al.

(2003)

Root meansquare error

RMSEp 1n

Xni1

i;Obsi;mod 2s Combal et al. (2002b)

Relative rootmean square error

RMSEp

1n

Xni1

i;Obsi;modi;Obs

!2vuut Weiss et al. (2000)Spectral angle SA v

;w

cos1 v

wv

w

0@

1A Jurdao et al. (2013) and

Yebra and Chuvieco(2009b)correlations between LFMC and multitemporal series of AVHRR datawere found for herbaceous species using the Normalized DifferenceVegetation Index (NDVI, Table 3) (Chladil & Nunez, 1995; Paltridge &Barber, 1988). Weaker relationships were reported for shrubs andtrees, as chlorophyll absorption and NIR reectance do not correlateas strongly for these vegetation types as for grasses (Yebra et al.,2008b). Improved empirical models have combined NDVI with surfacetemperature, a seasonal trend index (Chuvieco et al., 2004c) and mete-orological data (Garcia et al., 2008).

Since the launch of Terra and Aqua satellites by NASA in 1999 and2002, respectively, MODIS data have been much more commonly usedto estimate LFMC (Caccamo et al., 2012b; Dennison et al., 2005;Peterson et al., 2008; Roberts et al., 2006; Stow & Niphadkar, 2007;Stow et al., 2005, 2006; Yebra & Chuvieco, 2009a,b; Yebra et al., 2008b),as this sensor offers a ner spatial and spectral resolution than AVHRR,along with additional bands in the NIR and SWIR regions (Fig. 1). WhileMODIS and AVHRR will eventually become unavailable, the new VisibleInfrared Imaging Radiometer Suite (VIIRS) is expected to provide datacontinuity with better spatial resolution (Lee et al., 2006).

Medium-spatial-resolution sensors such as Landsat ThematicMapper (TM) and Enhanced Thematic Mapper + (ETM+) sensorshave also been used to estimate LFMC (Chuvieco et al., 2002; Yilmazet al., 2008). Chuvieco et al. (2004b) examined the correlation coef-cients computed between VI derived from AVHRR, SPOT-Vegetation,Landsat TM, and eld-measured LFMC and found consistent trendsamong the three sensors showing similar correlation values in spiteof being from different time periods and having different spatial res-olutions. Even though correlations were slightly stronger for LandsatTM images, the combination of temporal resolution (every 16 daysfor Landsat data) and cloud cover may limit the use of this type of me-dium spatial resolution data for LFMC monitoring applications.

Active microwave images have been used to examine LFMC incloudy areas, because these wavelengths are not interfered with bycloud cover. Leblon et al. (2002) compared radar image intensity valuesfrom European Remote Sensing satellite Synthetic Aperture Radar(ERS-1 SAR) imagery collected over boreal forest in 1994 to LFMC data.The authors reported a signicant relationship between rates of changein LFMC and radar backscatter, which can be largely accounted for inthe relationship between rate of change in the backscatter coefcientand rates of change in the LFMC, and therefore good correlations can befound in areaswhere other forest parameters are relatively stable in time.

Airborne hyperspectral data have proven useful for LFMC retrievaland validation against satellite data (Al-Moustafa et al., 2012; Chenget al., 2008, 2011; Roberts et al., 2006). Roberts et al. (2006) foundstronger correlations between spectral indices calculated fromAirborneVisible Infrared Imaging Spectrometer (AVIRIS) data and LFMC than in-dices calculated from MODIS data. However, while airborne data maybe acquired on demand, they are unlikely to have the temporal resolu-tion required for LFMC monitoring.

Current plans for a space-borne imaging spectrometer mission,the Hyperspectral InfraRed Imager (HyspIRI), specify a 19 day repeatcoverage for the imaging spectrometer instrument (http://hyspiri.jpl.nasa.gov/, last accessed February 2013), which is an insufcient tem-poral resolution for monitoring LFMC, but would provide useful datafor the calibration and validation of LFMC estimated from higher tem-poral resolution sensors. Combinations of coarse and moderate spatialresolution sensors may provide complementary spatial informationfor operational LFMC monitoring.

High spatial resolution sensors, such as Quickbird, Ikonos, GeoEye-1,and WorldView-2, have not previously been used for LFMC analysis. Thelimited spectral range of these sensors (48 bands covering the visibleand near infrared) excludes measurement of spectral regions containingliquid water absorption features, although simple VI based on visibleand near infrared reectance may have some value. Data from theplannedWorldView-3 mission, which includes a 3.7 m spatial resolution

SWIR band, may prove more useful for high resolution LFMC analysis.

l fornm

rmu

VI VI =

I RI =reen

II6 =II7 =

WI

I =

MI

461M. Yebra et al. / Remote Sensing of Environment 136 (2013) 4554685.4. Input data used to estimate LFMC

Atmospheric correction of satellite radiance measurements to re-trieve apparent surface reectance reduces the effects of variation inatmospheric conditions and solar irradiance. However, effects of dif-ferent sensor view angles and differences in slope, aspect, and soilbackground reectance are readily apparent in RS data. Removingsuch factors in spectral data means that any changes observed in thespectra are more likely to be related to leaf biochemical composition,leaf structure, or water content. VI are combinations of reectancemeasured at two or more wavelengths, which tend to suppress thesebackground effects and enhance spectral differences. VI can be directlyor indirectly related to LFMC depending on the wavelengths used(Fig. 1). Liquid water absorbs radiation strongly in the SWIR, so indicesthat include reectance in that region such as theNormalizedDifferenceInfrared Index (NDII) (Table 3) are directly related to LFMC and haveshown signicant correlations with LFMC for grasslands, shrublandsand woodlands (Chuvieco et al., 2002, 2004b; Roberts et al., 2006;Yebra et al., 2008b). MODIS band 5, centered on 1240 nm, also capturesliquid water absorption. The Normalized Difference Water Index(NDWI) (Table 3) has been correlated with LFMC in southern Californiachaparral (Dennison et al., 2005; Roberts et al., 2006). Indices using onlyvisible and NIR wavelengths with minimal water absorption, such asNDVI (Table 3), are indirectly related to LFMC. Visible-NIR indices aresensitive to both changes in pigment concentrations and changes inLAI that may correspond with changing LFMC. Multiple studies havefound strong empirical relationships between the Visible Atmospheri-cally Resistant Index (VARI, Table 3) and chaparral LFMC (Petersonet al., 2008; Roberts et al., 2006; Stow et al., 2005, 2006), as well as

Table 3Spectral indices used to estimate LFMC including their shortened acronym, mathematicain Fig. 1, except in the case of Water Index, where the subscripts refer to wavelength in

Index Fo

Normalized Difference Vegetation Index ND

Soil Adjusted Vegetation Index SA

Enhanced Vegetation Index EV

Visible Atmospherically Resistant Index VAVegetation Index Green (also Normalized Green Red Difference) VIgNormalized Difference Infrared Index (with band 6) NDNormalized Difference Infrared Index (with band 7) ND

Normalized Difference Water Index ND

Water Index W

Global Vegetation Moisture Index GVre-prone vegetation types (shrubland, heathland and sclerophyll forest)in south-eastern-Australia (Caccamo et al., 2012b). NDVI and othervisible-NIR indices are reliable indicators of vegetation phenology, andmeasures of phenological changes may lead to useful empirical relation-ships with LFMC (Hardy & Burgan, 1999).

The rst derivative of canopy reectance has been shown to be in-sensitive to variations caused by changes in illumination intensity,which may be related to variations in sun angle, cloud cover, atmo-spheric attenuation or topography (Blackburn, 2007; Imanishi et al.,2004). Another advantage of derivative analysis is that it can beused to determine the location of key spectral features such as thered edge, and chlorophyll absorption and, most importantly here,water absorption peaks in the near- and shortwave infrared.Al-Moustafa et al. (2012) successfully used airborne hyperspectral im-agery to estimate LFMC in a Calluna vulgaris-dominated semi-naturalupland area in the United Kingdom, but found that a broad wavebandspectral index provided a very similar correlation with LFMC comparedto that derived using rst derivatives at selected wavelengths between400 and 2500 nm derived from the hyperspectral data.More complex measures based on full range reectance spectrahave also successfully been used to estimate LFMC. Roberts et al.(2006) found correlations between chaparral LFMC and both greenvegetation and non-photosynthetic vegetation fractions calculatedfrom AVIRIS and MODIS data using a spectral mixture model. Finally,several authors have used thermal infrared (TIR) data to estimateplant water content, mainly for crops (Gonzalez-Dugo et al., 2012;Jackson et al., 1981; Moran et al., 1994). Forest and shrub canopiesare more complex, but relationships between the differences in air andsurface temperature and LFMC have been also found (Chuvieco et al.,2004c). TIR radiation can be used alone or in combination with green-ness vegetation indices such NDVI. The combined use of TIR and VI hasshown statistically stronger relationships with LFMC than either of thetwo variables alone (Chuvieco et al., 2004c) because the differences inair and surface temperature are closely related to the density of vegeta-tion coverage.

A breakthrough in TIR analyses was made using a synthesis ofpolar-orbiting and geostationary satellite RS data (Anderson et al.,2007, 2011; Kustas & Anderson, 2009). Geostationary RS data are spa-tially coarser than MODIS or AVHRR data, but geostationary sensorsacquire a daily prole of land surface temperature, so that dailyevapotranspiration can be more accurately estimated. Evapotranspi-ration estimates may reveal coarse scale drought stress, but plantphysiological adaptations to drought stress (and thus LFMC) vary byvegetation type.

6. Challenges in estimating FMC from satellite data

The main challenges for estimating LFMC from satellite data fall

mulation and reference. is reectance and the subscripts refer to MODIS bands shown.

lation Reference212 1

Rouse et al. (1974)

(1 + 0.5) (2 1) / (2 + 1 + 0.5) Huete (1988)2:5 21

2 6 17:5 3 1Huete et al. (2002)

(4 1) / (4 + 1 3) Gitelson et al. (2002)= (4 1) / (4 + 1) Tucker (1979)(2 6) / (2 + 6) Hardisky et al. (1983)(2 7) / (2 + 7)

252 5

Gao (1996)

R900/R970 Peuelas et al. (1993, 1997)2 0:1 6 0:02 2 0:1 6 0:02

Ceccato et al. (2002a)into two categories: theoretical and methodological. The former chal-lenges are common to all methods and are mainly derived from thestrength of the theoretical link between LFMC variations, the detectedsignal and other factors affecting spectral variation and data quality.The latter challenges are those derived from the specic method orapproach used to link spectral information to LFMC.

6.1. Theoretical challenges

6.1.1. Decoupling the impact of EWT from DMC on reectanceThe rst theoretical challenge is related to the complexity of

decoupling the impact of water from other factors affecting reec-tance, temperature, or backscatter. For instance, at the leaf level, esti-mation of LFMC from NIR and SWIR data requires discriminating thecontribution of DMC from EWT. Sensitivity analyses with RTM haveproved that variations in reectance when plants are drying are mainlydue to both EWT and DMC (Ceccato, 2001; Danson & Bowyer, 2004).Riao et al. (2005) showed that DMC of fresh samples could not beappropriately estimated from RTM because (i) the higher specic

absorption coefcient of water overmost of the solar-reected spectrum(Fig. 2) and (ii) EWT is usually greater than DMC. However, wherechanges in leaf water content occur over a short time period and thereis little change in DMC, we can expect strong relationships betweenleaf LFMC and spectral reectance (Bowyer & Danson, 2004). Trombettiet al. (2008) inverted the PROSPECT-SAILH model optimized with arti-cial neural networks to determine CWC over the continental UnitedStateswithMODIS data and validated their approach using CWC estima-tions obtained from AVIRIS airborne hyperspectral data. MODIS-CWCestimates gave consistent results throughout the continental USA butthe algorithm does not retrieve DMC and thus does not directly estimateLFMC.

A few studies have attempted to measure canopy DMC with RSdata. Accurate DMC retrieval would allow conversion of CWC to LFMC,assuming that LAI can also be accurately retrieved. Since the spectral re-sponse to DMC is masked in fresh samples but not in dry samples, Riaoet al. (2005) proposed obtaining annual estimates of spatially distributedDMC near the end of the driest season. They assume that this biomassestimate will remain relatively constant over time without a change inland cover.

Additionally, spectral indexes with absorption features sensitiveto DMC could improve LFMC estimates (Ustin et al., 2012; Wang etal., 2011, 2013). Dry matter exhibits several absorption features inthe reected SWIR (Asner, 1998; Feret et al., 2008; Jacquemoud etal., 1995, 1996) which could be measured with narrow-band sensors.

40

60

80

100

120

140

rptio

n co

effic

ien

t (cm

2 g-

1 ) EWTDMC

a)

462 M. Yebra et al. / Remote Sensing of Environment 136 (2013) 4554680

0.1

0.2

0.3

0.4

0.5

0.6

0.7

R2

Wavelength (nm)

EWT FreshDMC FreshEWT Dry DMC Dry

0

20

400 900 1400 1900 2400

400 900 1400 1900 2400

Abso

b)

Fig. 2. a) Spectral absorption coefcients for water and dry matter (cm2 g1) used inthe PROSPECT leaf model (version 4 (Feret et al., 2008)) and b) correlation betweenEWT and DMC for each wavelength. DMC had little effect on reectance of fresh leaves;therefore, radiative transfer modeling may be limited in providing a good estimation of

DMC for fresh leaves. Figure adapted from Riao et al. (2005).To retrieve DMC, several multispectral indexes have been developedlike the Cellulose Absorption Index (Nagler et al., 2000), theASTER-dened Lignin-Cellulose Absorption index and the improvedShortwave-Infrared Normalized Difference Residue Index (Serbin et al.,2009). These three spectral indices were designed to estimate the cover-age of dry non-photosynthetic vegetation (plant litter or crop residue)over bare soil, and not DMC of living foliage.

Two dry matter indices were developed based on the spectral dif-ferences between DMC and EWT (Romero et al., 2012; Wang et al.,2011, 2013). The Normalized Dry Matter Index (NDMI) is based onone narrow absorption feature of dry matter (1722 nm) at whichthe specic absorption coefcient of dry matter is greater than thatof water (Wang et al., 2011, 2013). The effect of DMC on simulatedleaf reectance may be seen by a change in the slope of reectancespectra at 17101750 nm (Fig. 1). Romero et al. (2012) found that thenormalized difference ratio (2305 1495)/(2305 + 1495), where is reectance and the subscript is wavelength in nm, also estimatedDMC, because the absorption coefcient of dry matter at 2305 nm isgreater than that of liquid water. While the absorption coefcients ofdry matter may be greater than water at 1722 and 2305 nm, it is notcertain that these indices could be used to measure canopy DMC be-cause there is often more liquid water than dry matter in leaves.

Wang et al. (2013) hypothesized that the ratio of a spectral waterindex and a spectral dry matter index would be related to LFMC sinceLFMC is the ratio of EWT/DMC. The relationships between index ra-tios and LFMC were signicant but non-linear (Wang et al., 2013).The ratio of the NDII6 (Table 3) to NDMI produced particularly strongcorrelations. Experimental data at the leaf scale validated this approach(Wang et al., 2013), but future research needs to be done at the canopyscale with narrow-band sensors. It is currently unclear how uncertaintyin retrieved DMC and EWT will affect accuracy of combined LFMCestimation.

6.1.2. Confounding effects of LAI and other canopy variablesAt the canopy level, the retrieval of LFMC from RS data needs to

discriminate the inuence of leaf structure, LAI and fraction coverageon plant reectance (Ceccato et al., 2002b; Zarco-Tejada et al., 2003).CWC is the product of leaf EWT and LAI (Eq. 5), but LFMC is indepen-dent from LAI and may remain constant while LAI varies spatially ortemporally. Weak relationships between LFMC and spectral reectanceare expected when there is also variation in LAI, as this factor has astronger inuence on canopy reectance (Bowyer & Danson, 2004),but if LAI changes slowly over space and time, strong relationships be-tween spectral reectance and LFMC may be expected (Al-Moustafaet al., 2012). This was conrmed in the physical model-based study ofDanson and Bowyer (2004) which showed that when a canopy reec-tance model was driven by site-specic biophysical data with a narrowrange of leaf DMCand LAI, statistically signicant relationships betweenVIs and LFMC were obtained.

LAI signicantly affects NDII, an index sensitive to water absorption,for a constant EWT value. But if the EWT LAI product (CWC) is con-stant, NDII also remains quite stable. PROSAIL simulations in Fig. 3demonstrate the above statement. NDII6 varies broadly (0.1050.389)for the same EWT (0.01 g/cm2) when LAI (15) and therefore CWC(0.0120.060 g/cm2) is varied (Fig. 3, left). On the other hand,NDII6 would have similar values (0.3260.404) for the same CWC(0.05 g/cm2) even though EWT changes (Fig. 3, right). In thiscase, CWC is kept constant by simultaneously changing EWT andLAI.

Spatial heterogeneity of vegetation within a pixel is a major issuefor estimating LFMC and for comparing estimated LFMC with groundmeasurements (Qi et al., 2012). For heterogeneous areas such assavannah-like environments, LFMC estimated from RS data will notrepresent LFMC for a single species, but rather an aggregate LFMC forthe area measured by the sensor. There is potential for using spectral

mixture analysis (SMA, Shimabukuro and Smith (1991)) to estimate

LFMC at the subpixel level and to improvemethods to aggregate samplesfrom different endmembers (species). When species vary spatially, adifferent set of endmembers could be appropriate to estimate LFMC atthe subpixel level for each species, inwhat is calledmultiple endmemberSMA (MESMA, Roberts et al. (1998)). Relative SMA (RSMA, Okin (2007),Okin et al. (2013)) might be another option when fraction of eachendmember changes over time within the same pixel. Another possibil-ity is that two pixels might have the same species composition with thesame LFMC but different fraction of soil cover. To account for this, Huanget al. (2009) isolated the vegetation surface reectance from the soil sig-nal using SMA to estimate in their case corn and soybean water contentindependent of the soil fraction. Finally, operational LFMC from MODISor other coarse spatial resolution optical image data is most affected byheterogeneous vegetation type composition within the ground resolu-tion element. For example, natural forest could contain in the samepixel irrigated agriculture and each vegetation type would have a differ-

ships between remotely sensed variables and LFMC. Peterson et al.

differences in site-specic properties and calibrated a statistical modelbased on the maximumminimum normalization of VARI and NDII6.

Researchers have also investigated new statistical methods to im-prove statistical model performance. Stow and Niphadkar (2007)normalized time series MODIS VIs using rescaling based on time se-ries maximum and minimum VI values within each pixel. Maxminscaling reduced the effects of spatial and interannual varying vegetationcover, in order to relate changes of LFMC to meteorological drivers andalso reduced errors in LFMC estimates. Kogan et al. (2003) used a simi-lar approach assessing vegetation health in response to drought usingAVHRR NDVI time series. Li et al. (2008) demonstrated the potentialof genetic algorithms coupled with partial least squares (GA-PLS) in re-trieving EWT and CWC. Zhang et al. (2011) introduced orthogonalsignal correction-partial least square regression (OSC-PLSR) to extractEWT and LFMC from lab-measured reectance spectra. The OSC-PLSRmodel showed good prediction for LFMC and reduced model complex-

ereff int

463M. Yebra et al. / Remote Sensing of Environment 136 (2013) 455468variable. This allowed the regression model to be used for a given func-tional type. Similarly, Caccamo et al. (2012b) aimed to account for

0.000.050.100.150.200.250.300.350.400.45

0.000 0.050 0.100

ND

II6

CWC (g cm-2)Fig. 3. PROSAIL simulations. Constant EWT = 0.01 g/cm2 and variant LAI = 15 and thsimultaneously EWT = 0.010.04 g/cm2 and LAI = 1.255 (Right). For both cases lea

2 2(2008) estimated LFMC using MODIS reectance data by calibratingempirical equations which accounted for site-specic and inter-annualdifferences in vegetation amount and condition. Two new independentvariables were added to amultiple linear regression analysis: an additiveVI summary statistic variable, and a multiplicative VI summary statisticent seasonal behavior. LFMC of the natural forest would be impacted bythe irrigated agriculture, so it would also need to be spectrally unmixed.

6.2. Methodological challenges

Both physical model-based (RTM) and statistical (empirical) methodsto estimate LFMC from RS data have considerable limitations.

6.2.1. Site-specicity of statistical methodsDue to the site-specicity of statistical methods, empirical rela-

tionships cannot be applied to regional or global scales due to spatialdifferences in leaf and canopy characteristics, soil background, sensorcharacteristics and observation conditions (Dennison et al., 2005;Garcia et al., 2008; Riao et al., 2005; Yebra et al., 2008a). Dennisonet al. (2005) demonstrated that the best t linear regressions ofLFMC versus NDVI were site dependent (Fig. 4). Garcia et al. (2008)combined AVHRR and meteorological data for estimating LFMC.Their model showed good performance in grassland and shrublandbut the authors stated that the application of the models to areaswith different vegetation types/species may provide an unreliable es-timation of LFMC.

Although statistical methods may be site specic, models based onheterogeneous sites or on multiple sites may capture broad relation-33 mg/cm , DMC = 0.02 g/cm , hot spot = 0.001, plagiophile leaves, sun angle = 25, vieity compared to simple PLSR. Additional surface measurements such assoil moisture may also provide opportunities for improving remoteestimation of LFMC (Qi et al., 2012).

6.2.2. Constraining parameters of physical model-based approachesSince RTM are based on physical relationships that are independent

of sensor or site conditions, they should be more universal than empir-icalmodels. However, the selection and parameterization of RTM are farmore complex than empirical models, since they require as inputs plantphysiological and structural variables that are not always available, andare based on assumptions that may not accurately resemble conditionsfound in nature (e.g., Lambertian broad-at leaves, semi-innite hori-zontally homogeneous plant canopies), especially when complex cano-pies are involved. Finally, physical models (i) present uncertaintiesin the inversion mode, since very similar reectances can be derivedfrom a different set of input parameters, which is the well-knownill-posed inverse problem (Garabedian & Paul, 1964) and (ii) do nottake into account ecophysiological relations, and therefore they mightprovide poor estimations when unrealistic combinations of input pa-rameters are considered.

A priori knowledge of plant biophysical parameters should be usedto constrain the input parameters of the RTM to model conditions asclosely as possible to the actual canopy state (Combal et al., 2002a).Some authors have chosen to include data derived from satellite im-agery as input parameters (Zarco-Tejada et al., 2003). Others have re-lied upon experimental data in controlled conditions (Riao et al.,2005; Yebra & Chuvieco, 2009a). Yebra and Chuvieco (2009a) andYebra et al. (2008b) proposed using ecological rules observed onthe eld to avoid simulating unrealistic spectra derived from combina-tions of parameters never met in nature. Using these ecological rules insimulation models signicantly decreased the residual estimationerror (RMSE = 19.77%) when compared to models run randomly

0.000.050.100.150.200.250.300.350.400.45

ND

II6

EWT (g cm-2)0.000 0.020 0.040 0.060

ore variant CWC = 0.0120.060 g/cm2 (Left). Constant CWC = 0.05 g/cm2 by varyingernal structure parameter was held constant at 1.5, leaf chlorophyll a + b content =

w angle = 5, relative azimuth angle = 30.

%)

464 M. Yebra et al. / Remote Sensing of Environment 136 (2013) 455468combining the input parameters (RMSE = 64.93%) (Yebra & Chuvieco,2009b).

Recently Jurdao et al. (2013) made an exhaustive exploration ofdata sources to include ecological criteria in an RTM and estimatedLFMC in woodlands of two different climatic regions of Spain. In an at-tempt to nd the maximum accuracy, the authors xed LAI and canopycoverage of vegetation (Ccov) in the inversion using different remotesensing products. However, these steps did not improve their results.

Since all these RTM studies are based on a priori knowledge of plantbiophysical parameters to represent realistic situations, the RTMmodelscalibrated this way can be applied to other areas without affecting theaccuracy of the estimations.

However, Pingheng and Quan (2011) pointed out that all theseapproaches of alleviating the ill-posed problem based on a priori knowl-edge of plant biophysical parameters can restrict the operational utility

Fig. 4. Slope, y-intercept and R2 values for the best t linear relationship between LFMC (County Fire Department. Data from Dennison et al. (2005).ofmodel inversion. To overcome this issue, the authors applied a separatemerit function for each of the RTM input parameters based on specicwavelengths determined by sensitivity analysis. Extensive validationsbased on both in situ measured data sets and a RTM-simulated data setsuggested that these procedures can substantially improve the inversionmodel performance and strongly reduce the ill-posed problem. How-ever, this procedure needs to be applied and tested at the canopy level.

6.2.3. Dependence of physical model-based accuracies on inversionprocedure

Another critical aspect of retrieval of LFMC is that RTM perfor-mance has a strong dependency on the inversion procedure used.Yebra (2008) compared two different merit functions of similarity be-tween observed and simulated spectra (relative minimum quadraticdistance (RMSEp) and minimum spectral angle (SA), Table 2) andconcluded that the use of the SA criteria provides a more consistentmeasure of spectral similarity which led to a more accurate estimationof LFMC, with RMSE and R2 of 14.6% and 0.63 versus 22.43% and 0.36using RMSEp (Fig. 5).

7. Main problems for the operational use of RS derived LFMCmodels

7.1. LFMC estimation error

Errors in LFMC estimation may directly affect safety and resourcecosts associated with prescribed re and wildre suppression (Weiseet al., 1998), depending on the sensitivity of the re behavior modelto LFMC. The most widely used re spread model (Rothermel) is highlysensitive to LFMC in most fuel types and Jolly (2007) for example foundthat a 10% difference in LFMC could produce up to 1200% difference inpredicted ROS. Fire managers using LFMC models must be aware of theuncertainties involved in LFMC estimation, and how this uncertaintycan affect the accuracy of re spread predictions and re risk estimation.Additionally, a formal evaluation program for LFMC estimation methodsshould better guide end-users to decide which model-product theyshould use in accordance with the accuracy needed.

Several published studies (Table 1) report LFMC estimation accura-cies against eld measured LFMC using the coefcient of determination(R2), the root mean square error (RMSE) or mean absolute error(MAE). However, a conclusive comparison between studies is notpossible due to the many differences in methodology and accuracy

and NDVI for chamise (Adenostoma fasciculatum) at 11 sites sampled by the Los Angelesreporting. For example, these studies have differences in: i) calibrationand validation procedures (e.g. RMSE sometimes reects site-based ormulti-site parameter tting, sometimes independent cross-validation);ii) the input data used (sensor, RS product and collection); iii) the num-ber, location, and type of eld sites considered; and iv) protocol for eldsample collection and processing. Future scientic studies should striveto use independent datasets to assess accuracy and report errors ratherthan R2 values.

In general terms, LFMC of shrublands can be retrieved with higheraccuracy than LFMC of grasslands and woodlands. Focusing on thestudies that performed an independent validation (10 of 18 fromTable 1), errors reported ranged between 8 and 20% for shrublands(Caccamo et al., 2012b; Garcia et al., 2008; Yebra & Chuvieco, 2009b),2561% for grasslands (Chuvieco et al., 2004c; Garcia et al., 2008; Yebraet al., 2008b) and approximately 30% for woodlands (Jurdao et al.,2013; Yebra & Chuvieco, 2009a). Higher errors may be reported forgrassland species due to higher variability in LFMC for grasses over time.

Errors in estimated LFMC are critically important for determiningre risk. An error of 20% added to or subtracted from an estimatedLFMC of 90% would result in shrubland re danger ranging from low(ignition probability (IP) = 19% associated with a LFMC = 110%), tohigh (IP = 60% associated with a LFMC = 70%) when IP is calculatedas in Chuvieco et al. (2004a). Dasgupta et al. (2007) evaluated the impli-cation of uncertainties in LFMC for re spread rate predictions usingFARSITE re behavior model (Finney, 1998). Their results showed thatmodeled re spread rates for the pocosin fuel model when LFMC was

LFMC estimates. Future work on statistical methods should

465M. Yebra et al. / Remote Sensing of Environment 136 (2013) 455468below100%, andmaximumerrors of 56% in LFMC estimation could pro-duce an equivalent error in re spread rate of 47 m h1 for a no wind,no slope re spread simulation.

7.2. LFMC and re behavior

Some operational re spread models (e.g. 1998) use fuel modelscontaining spatially constant descriptions of LFMC. Use of spatiallyvariable LFMC estimates provided by RS data can produce differencesin modeled re spread (Bossert et al., 2000). A challenge for re be-havior models that incorporate more complex descriptions ofthree-dimensional fuel distributions (e.g. Burgan, 1979; Linn et al.,2002; Mell et al., 2007; Morvan & Dupuy, 2004; Zylstra, 2011a,b) isestimation of LFMC within multiple fuel strata. In these situations,two-layer RTM, such as the Kuusk Markov Chain Canopy ReectanceModel (MCRM (Kuusk, 2001)) may prove useful. The Kuusk modelconsiders that the vegetation is homogeneously distributed for eachlayer and uses leaf optical properties derived from the PROSPECT

0.00

0.05

0.10

0.15

0.20

0.25

B3 B4 B1 B2 B6 B7

Refle

ctan

ce

MODIS Bands

Selected spectrum (RMSE*p); LFMC=85.84%Selected spectrum (SA); LFMC=47.15%Observed espectra; LFMC=45.54%

Fig. 5. Example of observed and simulated spectra using relative mean square error(RMSE) and spectral angle (SA) as merit functions. The simulated spectrum selectedwith the RMSE is more similar in absolute terms to the observed spectrum than thesimulated spectrum selected by the SA function. However the spectrum selected by theSA function corresponds to LFMC estimates closer to the observed value (Yebra, 2008).model (Jacquemoud, 1990). The canopy directional reectance isgenerated based on the single scattering and diffuse uxes fromeach layer, using direct and diffuse solar irradiance. The Kuuskmodel is more easily parameterized than other geometric modelsdue to its assumptions of homogenous canopies, while still providingsufcient complexity through its inclusion of two vegetation layerswith independent input conditions. On account of that, this modelhas been successfully used for re severity modeling (Chuvieco etal., 2007; De Santis & Chuvieco, 2007). The use of 3D models (e.g.DART (Gastellu-Etchegorry et al., 2004) or FLIGHT (North, 1996))should also be considered, although their parameterization wouldbe signicantly more challenging due to the large number of inputparameters that are needed.

7.3. LFMC and re risk

The conceptual denition of a re risk assessment system shouldinclude the most relevant components associated with the re pro-cess (Chuvieco et al., 2012). In relation to LFMC, another challengeis translating LFMC estimated using RS data into re risk units toallow integration into a comprehensive re risk assessment system.LFMC can be converted into re risk using empirical models relatingLFMC to re occurrence, as discussed in Section 2. LFMC can also betranslated into ignition probability (IP), where moisture of extinction

focus on making them more robust across larger areas. Physical

methods should improve retrieval of DMC and provide realisticranges of parameters for solving the ill-posed model inversionproblem.

iii) Most prior work has examined LFMC in Mediterranean ecosys-tems in Europe and Western North America. Further researchis needed to assess the full utility of LFMC estimation acrossother re-prone ecosystems.

iv) Reduction of error in LFMC estimates is needed to improve rerisk estimation and adoption by the end-user community. In-tegration of LFMC into re behavior and re risk modelsshould include uncertainty produced by error in LFMC esti-mates.

v) An international effort is needed to create a network of eldmeasurements for use in improving and validating LFMC esti-mates. The global database should use consistent methodolo-gy and be properly documented, georeferenced and publiclyaccessible.

vi) Research projects should go beyond scientic papers andshould search for long-term operational viability of productsand interaction with end-users.

vii) A formal evaluation program for methods of mapping LFMCshould be organized to better guide end-users to assess which(ME) identies the minimum water content that prevents re igni-tion. Fuelswith LFMC values aboveMEwould have low (or no) probabil-ity of being burned (Dimitrakopoulos & Papaioannou, 2001). This wasthe basis of using ME to convert FMC to IP in a re risk assessment byChuvieco et al. (2004a). The authors used an inverse linear function toobtain IP from ME, based on critical thresholds (105% for shrub speciesand 30% for grasslands) derived from experimental analysis. Jurdaoet al. (2012) explored several methods to convert LFMC into IP consider-ing climate (Mediterranean and Eurosiberian regions) and vegetationfunctional types (grasslands and shrublands). Non-parametric signi-cance tests, histograms and percentiles, classication trees, and logisticregression models were used for estimating the IP from ve variablesbased on LFMC. Logistic regression analysis was found to be the mostadvantageousmodelingmethodology, since it uses several predictor var-iables to compute a continuous probability of IP. Chuvieco et al. (2004a)and Jurdao et al. (2012) present regional solutions to estimating IPfrom LFMC, but methods that can be scaled globally could potentiallybe connected to global observations of re occurrence (Justice et al.,2002). Additionally, an elevated probability of high re occurrencedoes not necessarily imply a large number of res or an extensive burnedarea, because when causal agents are absent, few or no re events occur(Chuvieco et al., 2009).

8. Concluding remarks

Developing operational LFMC estimation is useful for improvingre risk assessment. Improved spatial and temporal monitoring ofLFMC can potentially result in better allocation of re protectiveresources and increase vigilance for re hazard to people and property.This review demonstrates that recent research has addressed many is-sues constraining operational LFMC monitoring. Further advancementtowards operational adoption of LFMC estimation can be realized inthe following areas:

i) Improvements in spatial resolution and spectral dimensionali-ty which have the potential to provide more accurate LFMC es-timation. VIIRS represents an advance in spatial resolution overMODIS, while the planned HyspIRI mission could provide com-plementary improvements in the spectral domain.

ii) Progress in algorithm development may enhance the utility ofproduct or method for LFMC estimation best ts their needs.

466 M. Yebra et al. / Remote Sensing of Environment 136 (2013) 455468Acknowledgments

Portions of this work were supported by funding from the UnitedStates National Aeronautics and Space Administration grants#NNX11AF93G and #NNX09AN51G. Marta Yebra was supportedby a CSIRO OCE postdoctoral scholarship. USDA is an equal opportunityprovider and employer. Thanks to Dr. Stuart Matthews and Dr. MattPlucinski (CSIRO Ecosystem Sciences) for their helpful comments thatimproved this paper.

References

Al-Moustafa, T., Armitage, R. P., & Danson, F. M. (2012). Mapping fuel moisture contentin upland vegetation using airborne hyperspectral imagery. Remote Sensing of Envi-ronment, 127, 7483.

Alexander, M. E., & Cruz, M. G. (2012). Assessing the effect of foliar moisture on thespread rate of crown res. International Journal of Wildland Fire. http://dx.doi.org/10.1071/WF12008.

Anderson, M. C., Hain, C., Wardlow, B., Pimstein, A., Mecikalski, J. R., & Kustas, W. P.(2011). Evaluation of drought indices based on thermal remote sensing of evapo-transpiration over the continental United States. Journal of Climate, 24, 20252044.

Anderson, M. C., Norman, J. M., Mecikalski, J. R., Otkin, J. A., & Kustas, W. P. (2007). Aclimatological study of evapotranspiration and moisture stress across the continentalUnited States based on thermal remote sensing: 1. Model formulation. Journal ofGeophysical Research-Atmospheres, 112.

Anderson, S. A. J., & Anderson, W. R. (2010). Ignition and re spread thresholds in gorse(Ulex europaeus). International Journal of Wildland Fire, 19, 589598.

Asner, G. P. (1998). Biophysical and biochemical sources of variability in canopy reec-tance. Remote Sensing of Environment, 64, 234253.

Beaudoin, A., Vidal, A., Desbois, N., & Debaux-Ros, C. (1995). Monitoring the water sta-tus of Mediterranean forests using ERS-1 to support re risk prevention. IEEETransactions on Geoscience and Remote Sensing, 2, 963966.

Blackburn, G. A. (2007). Wavelet decomposition of hyperspectral data: A novel ap-proach to quantifying pigment concentrations in vegetation. International Journalof Remote Sensing, 28, 28312855.

Bossert, J. E., Linn, R. R., Reisner, J. M., Winterkamp, J. L., Dennison, P. E., & Roberts, D. A.(2000). Coupled atmosphere re behavior model sensitivity to spatial fuels charac-terization. Third symposium on re and forest meteorology, Amer. Meteor. Soc. 80thAnnual Meeting Long Beach, CA (pp. 2126).

Boulet, G., Chehbouni, A., Gentine, P., Duchemin, B., Ezzahar, J., & Hadria, R. (2007).Monitoring water stress using time series of observed to unstressed surface tem-perature difference. Agricultural and Forest Meteorology, 146, 159172.

Bowyer, P., & Danson, F. M. (2004). Sensitivity of spectral reectance to variation in livefuel moisture content at leaf and canopy level. Remote Sensing of Environment, 92,297308.

Bradstock, R. A., Cohn, J. S., Gill, A. M., Bedward, M., & Lucas, C. (2009). Prediction of theprobability of large res in the Sydney region of south-eastern Australia using reweather. International Journal of Wildland Fire, 18, 932943.

Burgan, R. E. (1979). Estimating live fuel moisture for the 1978 National Fire DangerRating System. Research paper. INT-RP-226. Intermountain Forest and Range Exper-iment Station, Ogden, Utah: USDA Forest Service (16 pp.).

Burgan, R. E., Klaver, R. W., & Klaver, J. M. (1998). Fuel models and re potential fromsatellite and surface observations. International Journal of Wildland Fire, 8, 159170.

Burrows, N. D. (1999). Fire behaviour in jarrah forest fuels. CALMScience, 3, 3184.Caccamo, G., Chisholm, L. A., Bradstock, R. A., & Puotinen, M. L. (2011). Assessing the

sensitivity of MODIS to monitor drought in high biomass ecosystems. Remote Sensingof Environment, 115, 26262639.

Caccamo, G., Chisholm, L. A., Bradstock, R. A., & Puotinen, M. L. (2012a). Usingremotely-sensed fuel connectivity patterns as a tool for re danger monitoring.Geophysical Research Letters, 39, L01302.

Caccamo, G., Chisholm, L. A., Bradstock, R. A., Puotinen, M. L., & Pippen, B. G. (2012b).Monitoring live fuel moisture content of heathland, shrubland and sclerophyll for-est in south-eastern Australia using MODIS data. International Journal of WildlandFire, 21, 257269.

Castro, F. X., Tudela, A., Gabriel, E., Montserrat, D., Canyameres, E., & Segarra, M. (2006).Evolution of live fuel moisture in Mediterranean forest. In D. X. Viegas (Ed.),V international conference on forest re research. Figuera da Foz, Amsterdam:Elsevier BV.

Ceccato, P. (2001). Estimation of vegetation water content using remote sensing for theassessment of re risk occurrence and burning efciency. Phylosophy (pp. 168).Greenwich: University of Greenwich.

Ceccato, P., Flasse, S., & Gregoire, J. M. (2002a). Designing a spectral index to estimatevegetation water content from remote sensing data: Part 2. Validation and applica-tions. Remote Sensing of Environment, 82, 198207.

Ceccato, P., Flasse, S., Tarantola, S., Jacquemoud, S., & Grgoire, J. M. (2001). Detectingvegetation leaf water content using reectance in the optical domain. Remote Sens-ing of Environment, 77, 2233.

Ceccato, P., Gobron, N., Flasse, S., Pinty, B., & Tarantola, S. (2002b). Designing a spectralindex to estimate vegetation water content from remote sensing data: Part 1. The-oretical approach. Remote Sensing of Environment, 82, 188197.

Cheney, N. P., Gould, J. S., & Catchpole, W. R. (1998). Prediction of re spread in grass-lands. International Journal of Wildland Fire, 8, 113.Cheng, T., Rivard, B., & Sanchez-Azofeifa, A. (2011). Spectroscopic determination of leafwater content using continuous wavelet analysis. Remote Sensing of Environment,115, 659670.

Cheng, Y. -B., Ustin, S. L., Riano, D., & Vanderbilt, V. C. (2008). Water content estimationfrom hyperspectral images and MODIS indexes in Southeastern Arizona. RemoteSensing of Environment, 112, 363374.

Chladil, M. A., & Nunez, M. (1995). Assessing grassland moisture and biomass in Tasmania.The application of remote sensing and empirical models for a cloudy environment.International Journal of Wildland Fire, 5, 165171.

Chuvieco, E., Aguado, I., Cocero, D., & Riao, D. (2003). Design of an empirical index toestimate fuel moisture content from NOAA-AVHRR analysis in forest re dangerstudies. International Journal of Remote Sensing, 24, 16211637.

Chuvieco, E., Aguado, I., & Dimitrakopoulos, A. (2004a). Conversion of fuel moisture contentvalues to ignition potential for integrated re danger assessment. Canadian Journal ofForest Research-Revue Canadienne de Recherche Forestiere, 34(11), 22842293.

Chuvieco, E., Aguado, I., Jurdao, S., Pettinari, M. L., Yebra, M., Salas, J., et al. (2012). Inte-grating geospatial information into re risk assessment. International Journal ofWildland Fire. http://dx.doi.org/10.1071/WF12052.

Chuvieco, E., Cocero, D., Aguado, I., Palacios-Orueta, A., & Prado, E. (2004b). Improvingburning efciency estimates through satellite assessment of fuel moisture content.Journal of Geophysical Research-Atmospheres, 109(D14S07), 18. http://dx.doi.org/10.1029/2003JD003467.

Chuvieco, E., Cocero, D., Riao, D., Martn, M. P., Martnez-Vega, J., de la Riva, J., et al.(2004c). Combining NDVI and surface temperature for the estimation of live fuelmoisture content in forest re danger rating. Remote Sensing of Environment, 92,322331.

Chuvieco, E., De Santis, A., Riao, D., & Halligan, K. (2007). Simulation approaches forburn severity estimation using remotely sensed images. Journal of Fire Ecology, 3,129150.

Chuvieco, E., Gonzlez, I., Verd, F., Aguado, I., & Yebra, M. (2009). Prediction of re oc-currence from live fuel moisture content measurements in a Mediterranean eco-system. International Journal of Wildland Fire, 18, 430441.

Chuvieco, E., Riao, D., Aguado, I., & Cocero, D. (2002). Estimation of fuel moisture con-tent from multitemporal analysis of Landsat Thematic Mapper reectance data:Applications in re danger assessment. International Journal of Remote Sensing,23, 21452162.

Chuvieco, E., Yebra, M., Jurdao, S., Aguado, I., Salas, F. J., Garca, M., et al. (2011). Fieldfuel moisture measurements on Spanish study sites. In U.o.A. Department ofGeography, Spain (Ed.), Alcala de Henares: Department of Geography.

Colombo, R., Meroni, M., Marchesi, A., Busetto, L., Rossini, M., Giardino, C., et al. (2008).Estimation of leaf and canopy water content in poplar plantations by means ofhyperspectral indices and inverse modeling. Remote Sensing of Environment, 112,18201834.

Combal, B., Baret, F., & Weiss, M. (2002a). Improving canopy variables estimation fromremote sensing data by exploiting ancillary information. Case study on sugar beetcanopies. Agronomie, 22, 205215.

Combal, B., Baret, F., Weiss, M., Trubuil, A., Mace, D., Pragne're, A., et al. (2002b). Retrievalof canopy biophysical variables from bidirectional reectance Using prior informa-tion to solve the ill-posed inverse problem. Remote Sensing of Environment, 84, 115.

Danson, F. M., & Bowyer, P. (2004). Estimating live fuel moisture content from remote-ly sensed reectance. Remote Sensing of Environment, 92, 309321.

Dasgupta, S., Qu, J. J., Hao, X., & Bhoi, S. (2007). Evaluating remotely sensed live fuelmoisture estimations for re behavior predictions in Georgia, USA. Remote Sensingof Environment, 108, 138150.

Davies, G. M., Legg, C. J., Smith, A. A., & MacDonald, A. J. (2009). Rate of spread of res inCalluna vulgaris-dominated moorlands. Journal of Applied Ecology, 46, 10541063.

De Santis, A., & Chuvieco, E. (2007). Burn severity estimation from remotely senseddata: Performance of simulation versus empirical models. Remote Sensing of Envi-ronment, 108, 422435.

Deeming, J. E., Burgan, R. E., & Cohen, J. D. (1978). The National Fire-Danger Rating System1978. General technical report INT-39. Intermountain Forest and Range Experiment Sta-tion, Ogden, Utah: USDA Forest Service (63 pp.).

Dennison, P. E., & Moritz, M. A. (2009). Critical live fuel moisture in chaparral ecosystems:A threshold for re activity and its relationship to antecedent precipitation. Interna-tional Journal of Wildland Fire, 18, 10211027.

Dennison, P. E., Moritz, M. A., & Taylor, R. S. (2008). Evaluating predictive models ofcritical live fuel moisture in the Santa Monica Mountains, California. InternationalJournal of Wildland Fire, 17, 1827.

Dennison, P. E., Roberts, D. A., Peterson, S. H., & Rechel, J. (2005). Use of NormalizedDifference Water Index for monitoring live fuel moisture. International Journal ofRemote Sensing, 26, 10351042.

Desbois, N., Deshayes, M., & Beudoin, A. (1997). Protocol for fuel moisture contentmeasurements. In E. Chuvieco (Ed.), A review of remote sensing methods for the studyof large wildland res (pp. 6172). Alcal de Henares: Departamento de Geografa,Universidad de Alcal.

Dimitrakopoulos, A., & Papaioannou, K. K. (2001). Flammability assessment of Mediter-ranean forest fuels. Fire Technology, 37, 143152.

Dimitrakopoulos, A. P., & Bemmerzouk, A. M. (2003). Predicting live herbaceous mois-ture content from a seasonal drought index. International Journal of Biometeorology,47, 7379.

Feret, J. -B., Franois, C., Asner, G. P., Gitelson, A. A., Martin, R. E., Bidel, L. P. R., et al.(2008). PROSPECT-4 and 5: Advances in the leaf optical properties model separat-ing photosynthetic pigments. Remote Sensing of Environment, 112, 30303043.

Finney, M. A. (1998). FARSITE: Fire area simulatorModel development and evaluation.USDA Forest Service, Rocky Mountain Research Station, Research Paper RMRS-RP-4 (FortCollins, CO. 51 pp.).

467M. Yebra et al. / Remote Sensing of Environment 136 (2013) 455468Fletcher, T. H., Pickett, B. M., Smith, S. G., Spittle, G. S., Woodhouse, M. M., Haake, E.,et al. (2007). Effects of moisture on ignition behavior of moist California chaparraland Utah leaves. Combustion Science and Technology, 179, 11831203.

Fourty, T., & Baret, F. (1997). Vegetation water and dry matter contents estimated fromtop-of-the atmosphere reectance data: A simulation study. Remote Sensing ofEnvironment, 61, 3445.

Gao, B. C. (1996). NDWI. A normalized difference water index for remote sensing ofvegetation liquid water from space. Remote Sensing of Environment, 58, 257266.

Garabedian, P., & Paul, R. (1964). Partial differential equations. New York: Wiley.Garcia,M., Chuvieco, E., Nieto, H., & Aguado, I. (2008). Combining AVHRR andmeteorolog-

ical data for estimating live fuelmoisture content. Remote Sensing of Environment, 112,36183627.

Garnier, E., & Navas, M. L. (2012). A trait-based approach to comparative functionalplant ecology: Concepts, methods and applications for agroecology. A review.Agronomy for Sustainable Development, 32, 365399.

Gastellu-Etchegorry, J. P., Martin, E., & Gascon, F. (2004). DART: A 3D model for simu-lating satellite images and studying surface radiation budget. Remote Sensing ofEnvironment, 25, 7396.

Gaulton, R., Danson, F. M., Ramirez, F. A., & Gunawan, O. (2013). The potential ofdual-wavelength laser scanning for estimating vegetation moisture content.Remote Sensing of Environment, 132, 3239.

Gellie, N., Gibos, K., & Johnson, K. (2010). Relationship between severe landscapedryness and large destructive res in Victoria. In X. Viegas (Ed.), VI internationalconference on forest re research. Coimbra, Portugal: ADAI.

Giglio, L., Descloitres, J., Justice, C. O., & Kauffmam, J. B. (2003). An enhanced contextualre detection algorithm for MODIS. Remote Sensing of Environment, 87, 273282.

Gillon, D., Dauriac, F., Deshayes, M., Valette, J. C., & Moro, C. (2004). Estimation of foliagemoisture content using near infrared reectance spectroscopy. Agricultural and ForestMeteorology, 124, 5162.

Gitelson, A., Kaufmam, J. Y., Stark, R., & Rundquist, D. (2002). Novel algorithms forremote estimation of vegetation fraction. Remote Sensing of Environment, 80,7687.

Gonzalez-Dugo, V., Zarco-Tejada, P., Berni, J. A. J., Suarez, L., Goldhamer, D., & Fereres, E.(2012). Almond tree canopy temperature reveals intra-crown variability that iswater stress-dependent. Agricultural and Forest Meteorology, 154, 156165.

Gould, J. S., McCaw, W. L., Cheney, N. P., Ellis, P. F., Knight, I. K., & Sullivan, A. L. (2007).Project Vesta Fir