FRACTIONAL VEGETATION COVER ESTIMATION FROM

when values extracted from ASTER spectral library areconsidered, with errors slightly better than the previouscase, lower than 12%.Table 2. Errors on the fractional vegetation cover,e(FVC), depending on the NDVI value consideredand obtained according to the sensibility analysisgiven by Equations (5) to (8). The methodologyproposed by Gutmand and Ignatov (1998) has .100.100.100.10δNDVI 0.010.030.060.080.10δNDVI00.090.070.050.030.00FVC (%) 84.75VARI green 22.78e(FVC)0.140.130.130.130.143.2. Spectral Mixture AnalysisNDVIδNDVIδNDVI ite the NDVI has been widely used forassessment and monitoring of changes in canopybiophysical properties such FVC, this vegetation indexshows problems of saturation for high vegetationcovers, as has been pointed out by [13]. The authorsfound that for FVC higher than 60% the NDVI is almostinsensitive to FVC changes, mainly due to the NIRreflectance behaviour. In order to solve this problemand using the concept of ARVI (AtmosphericallyResistant Vegetation Index) [14], [12] propose anempirical relationship between FVC and a new indexdenoted as VARIgreen (Variable AtmosphericallyResistant Index) and given by:ρ green ρredρgreen ρred ρblue(10)with a standard error of estimation less than 10%.Table 3. Similar to Table 2, but using the NDVI andNDVI0 values extracted from the ASTER spectrallibrary.VARI green green reflectance (ρ green ) and including the bluereflectance (ρ blue) in order to compensate theatmospheric effects. Despite TOA (Top Of Atmosphere)reflectances can be also used in order to calculate theNDVI, better results can be achieved by using at-surfaceor atmospherically corrected reflectances. However,VARIgreen is calculated using TOA reflectances, so theatmospheric correction is included in Equation (9) bymeans of the blue reflectivity. Using 71 samples [12]obtained the following linear relationship between FVCand VARI:(9)The expression of the VARIgreen is similar to the NDVI(Equation 1), but substituting the NIR reflectance by theThe Spectral Mixture Analysis (SMA)technique has been developed in recent years to extractland-cover information at a sub-pixel level. SMAdivides each ground resolution element into itsconstituent materials using endmembers (EMs), whichrepresent the spectral characteristics of the cover types.When applied to multispectral satellite data, the result isa series of images each depicting the abundance of acover type. The basic physical assumption is that thereis not a significant amount of photon multiple scatteringbetween the macroscopic materials, in such a way thatthe flux received by the sensor represents a summationof the fluxes from the cover types (macroscopicmaterials) and the fraction of each one is proportional toits covered area [5]. This assumption complies with theproperties of the considered CHRIS/PROBA data sets,collected over a flat area and dominated byhomogeneous crop fields. As a result, most of theendme mber substances are sitting side-by-side withinthe field of view of the imager, and minimal secondaryreflections or multiple scattering effects can beassumed. In this paper a simple linear mixing modelLSU (Linear Spectral Unmixing) has been used, inwhich only a few EMs are used to describe the surfacecomposition in each pixel of an image [15]. Thereflectivity spectra for each endmember have beenautomatically extracted from the image using theAMEE (Automated Morphological EndmemberExtraction) method. The input to AMEE method is thefull spectral data cube, with no previous dimensionalityreduction. The method is based on two parameters: aminimum S min and a maximum S max spatial kernelsize. Firstly, a minimum kernel K S min is considered.This structuring element (SE) is moved through all thepixels of the image, defining a spatial context aroundeach hyperspectral pixel [16]. A morphologicaleccentricity index (MEI) [17] is then obtained bycalculating the SAD distance between the two above

signatures. This operation is repeated for all the pixelsin the scene, using SEs of progressively increased size,and the resulting scores are used to evaluate each pixelin both spatial and spectral terms. The algorithmperforms as many iterations as needed until K S max .The associated MEI value of selected pixels atsubsequent iterations is updated by means of newlyobtained values, as a larger spatial context is considered,until a final MEI image is generated. Endmemberselection is performed by a fully automated approachconsisting of two steps: 1) autonomous segmentation ofthe MEI image, and 2) spatial/spectral growing ofresulting regions [16].3.3. Algorithms testingThe expressions shown in section 3.1 havebeen tested by comparing the estimated values and theones measured in situ according to the methodologyexplained in section 2. For this purpose thePROBA/CHRIS image acquired at near nadir view(27.60º) has been used in order to extract the pixelvalues included in each plot. As has been commentedbefore, in order to obtain FVC from the scaled NDVI isnecessary to establish the values of NDVI0 and NDVI .Values proposed by [11] or [18] are very useful forglobal studies, but they could not be appropriate for aparticular agricultural area, as is the case of the Barraxtest site. In order to estimate the FVC from NDVIvalues over the Barrax test site, a linear relationshipbetween the NDVI values extracted from thePROBA/CHRIS image and the FVC measured in situhas been obtained:FVC 1.1101NDVI 0.0857(11)with a correlation coefficient r 0.91 and a standarderror of estimation of 13%. The NDVI0 and NDVI values can be obtained from Equation (11) by choosingFVC 0 and FVC 1, respectively. In this way oneobtains NDVI 0 0.08 and NDVI 0.98, which agreeswith the results obtained by [19] over the same test site.These values lead to an error on FVC less than 8%,according to the sensitivity analysis carried out insection 3.1. The root square addition between this resultand the standard error of estimation leads to a final errorof 15%.The direct application of Equation (10), inwhich FVC is obtained from VARIgreen values assuggested by [13], does not provides good results andtends to underestimate de FVC in comparison with thein situ values, with a root mean square error of 28%,which can be explained by taking into account thatEquation (10) was obtained over a particular area.Recalculating this expression according to the FVCvalues measured in situ and the VARIgreen valuesextracted from the CHRIS data, it is possible to obtain:FVC (%) 113.30VARIgreen 43.40(12)with a correlation coefficient r 0.97 and a standarderror of estimation of 8%, in more accordance with theresults obtained by [13].In order to obtain the FVC applying a LSUtechnique the AMEE method described in section 3.2has been used in order to extract the EMs from theimage without a priori knowledge of the field site. TheAMEE method extracted a total amount of 10 EMs, inwhich 2 EMs for green vegetation have been found (therest of the EMs correspond to clouds, shadows, baresoil, etc.). The comparison with in situ measurementsshows a final root mean square error (RMSE) less than12% is obtained. The LSU has been also applied usingonly 2 EMs (bare soil and green vegetation) extractedfrom the image by using the land cover map of the fieldsite. In this way the results are slightly better, with aRMSE less than 9%. The main constraint in this case isthat a priori knowledge of the field site is needed inorder to extract the EMs.According to the results obtained, FVC can beretrieved from satellite data using a linear relationshipwith the NDVI, with errors lower than 15%, and using alinear relationship with the VARIgreen, improving in thisway the results, with errors lower than 10%. FVC canbe also obtained using a LSU method and automaticallyextracted EMs, i.e., without a prior knowledge of thefield site. In this way, errors lower than 12% areobtained.4. ANGULAR EFFECTSIn order to analyze the angular effects on theFVC retrieved from satellite data, NDVI and VARIgreenvalues have been extracted for each plot at the fivePROBA/CHRIS acquisition view zenith angles: -57.40º,-42.53º, 27.60º, 42.44º and 57.29º. Fig. 3 shows theangular variation of the NDVI extracted from the sevenplots considered in this study. Differences between theNDVI at each angle and the NDVI at almost nadir view(27.6º) are generally greater for extreme view angles(-57.4º and 57.3º). Differences are lower than 0.04 inmost cases, except for the corn (C1) plot, reachingvalues of 0.07. These differences are not importantwhen estimating the FVC from NDVI values, as isshown in Fig. 4, in which the FVC estimated fromEquation (11) for each observation angle is graphed.Differences lower than 5% are obtained in all casesexcept for the corn (C1) plot, reaching differences of8%. An increase on FVC when increasing the viewangle is not observed in all cases, as will be expected.This fact could be explained due to the different angularresponse of the NDVI at backward and forwarddirections, as has been pointed out by [19].

0.9Garlic (G1)Corn (C2)Corn (C1)Sugar Beet (B3)Alfalfa (A10)Alfalfa (A1)Potatoes ew angle (º)geological studies. Different satellite methods have beenpublished in the last years in order to retrieve LSE fromthermal infrared satellite data: the reference channelmethod [21], thermal spectral indices [22], alpharesiduals [23], day/nigth methods [24], the Temperatureand Emissivity Separation (TES) method [25], amongothers. LSE can be also obtained from visible and nearinfrared satellite data. These methods estimate LSEfrom NDVI or FVC values [26,27,18]. In this context,LSE can be obtained for a mixed surface composed bysoil and vegetation using the following very simplifiedparameterization:Fig. 3. NDVI versus the CHRIS view angle for differentsamples.0.9Garlic (G1)Corn (C2)Corn (C1)Sugar Beet (B3)Alfalfa (A10)Alfalfa (A1)Potatoes (P1)0.80.7FVC0.6ε ε s (1 Pv ) ε V PV(13)where εs refers to the soil emissivity, εv to the vegetationemissivity and Pv is the FVC. Equation (13) must beparticularized to a certain wavelength o sensor band. Inthis last case, effective instead of spectral emissivitieswill be used according to the following expression:0.50.4ε 0.30.2 ε f dλ f dλλλ(14)λ0.10.0-60-40-200204060View angle (º)Fig. 4. Fractional vegetation cover estimated using alinear relationship with the NDVI versus the CHRISview angle.Low variations on NDVI with the view anglehas been also obtained by [20]. These authors point outthat the higher angular variations on NDVI are due tochanges in the solar zenith angle, and not in the viewangle. It should be noted that angular variations onNDVI and also FVC are higher for heterogeneoussurfaces than for homogeneous ones. In this study theplots selected tend to be homogeneous, with highvegetation cover except for the garlic plot (G1), withvery low vegetation cover. So in these cases negligiblecorrections between the values obtained at each viewangle and the nadir one have been obtained. Negligibleangular variations has been also observed whenanalyzing the results obtained with the LSU analysis.5. EMISSIVITYVALUESRETRIEVALFROMFVCLand surface emissivity (LSE) is a paremeterinvolved in the thermal part of the spectrum and is a keyvariable in order to retrieve land surface temperaturewith a good accuracy over natural surfaces, which arenot perfect blackbodies. LSE is also important inwhere fλ is the spectral function for a given sensor bandor simply filter function. According to error theory, theuncertainty on the emissivity, e(ε), calculated usingEquation (13) is given bye(ε ) δ ε2s δ ε2v δ P2v(15)whereδεs εe(ε s ) (1 Pv )e(ε s ) ε s(16) εe(ε v ) Pv e(ε v ) ε v(17) εe( Pv ) (ε v ε s ) e( Pv ) Pv(18)δεv δ Pv and e(εs ), e(εv ) and e(Pv ) refers to the errors on soilemissivity, vegetation emissivity and FVC, respectively.Therefore, in order to apply Equation (13),effective soil and vegetation emissivities for a certainthermal band are needed. The study will be focused onfield thermal radiometers bands, so these instrumentsare used in order to obtain in situ values. In this case, wewill consider two CIMEL radiometers, model CE 312-1(CIMEL 1), with four bands in the thermal regionbetween 8 and 14 µm, and model CE 312-2 (CIMEL 2),

with six bands in 8-14 µm. It should be noted that bothradiometers include a broadband, and the fivenarrowbands of the model CE 312-2 are in coincidencewith the five ASTER (Advanced Spaceborne ThermalEmission and Reflection radiometer) thermal bands,whereas CIMEL 1 bands 3 and 2 are very similar to theAVHRR (Advanced Very High Resolution Radiometer)bands 4 and 5, respectively. Soil and vegetation bandemissivities have been obtained applying Equation (14)to each CIMEL band and using the spectra included inthe ASTER spectral library (http://speclib.jpl.nasa.gov).As an example, the effective values obtained forCIMEL 1 are shown in Table 4. It should be noted thatvegetation emissivities show low standard deviations forall CIMEL bands, whereas soil emissivities show onlylow standard deviations for CIMEL bands included inthe region 10-12 µm. Bands included in the region 8-9.5µm show high standard deviations, so a general valuefor soil emissivity is not acceptable for these bands.This problem can be solved by considering only theinceptisol class, which is the more common soil class onthe Earth’s surface and also in the Barrax test site.Table 5. Expressions for the land surfaceemissivity retrieval from fractional vegetationcover (Pv ) values and for each CIMEL band.MODELCIMEL 1CIMEL 2BAND1 (10.54 µm)2 (11.96 µm)3 (10.80 µm)4 (8.82 µm)1 (10.54 µm)2 (11.29 µm)3 (10.57 µm)4 (9.15 µm)5 (8.69 µm)6 (8.43 µm)EXPRESSIONε 0.962 0.021 Pvε 0.976 0.008 Pvε 0.969 0.013 Pvε 0.946 0.036 Pvε 0.962 0.021 Pvε 0.970 0.013 Pvε 0.968 0.013 Pvε 0.941 0.038 Pvε 0.949 0.033 Pvε 0.946 0.040 Pv0.020Table 5 shows the final expressions foremissivity retrieval at-nadir view according to Equation(13). The total error on emissivity according toEquations (15) to (18) versus the FVC value is shown inFig. 5 for CIMEL 1. The total error has been calculatedassuming an error for soil and vegetation emissivitiesequal to the standard deviation and an error for FVCequal to 10%. For CIMEL 1 bands, the errors are lowerthan 0.01 except for band 4 (11.96 µm) with low FVCvalues.BAND 1 (8-13 um)BAND 2 (8.2-9.2 um)BAND 3 (10.3-11.3 um)BAND 4 (11.5-12.5 um)0.0180.016Error on emissivityTable 4. Mean effective emissivities obtained from theASTER spectral library for the CIMEL 1 radiometer. Thestandard deviation is shown in brackets.CLASS BAND 1 BAND 2 BAND 3 BAND 8All .60.70.80.91.0Fractional vegetation coverFig. 5. Total error on emissivity versus the fractionalvegetation cover for CIMEL 1 bands.As an example, Figure 6 shows the land surfaceemissivity map obtained for the CIMEL broadband (813 µm) and using the FVC image obtained with theLSU technique.

6. CONCLUSIONSFig. 6. Land surface broadband (8-13 µm) emissivitymap obtained from CHRIS data over the Barrax test site(see land use map in Figure 3).Before ending this section, a comment aboutthe validity of Equation (13) for surface emissivityretrieval should be given. Hence, Equation (13) is onlyvalid for flat surfaces. For rough surfaces the cavityeffect should be taken into account, for example in thefollowing way:ε ε s (1 Pv ) εV PV C (19)where the term C includes the cavity effect. Followingthe methodology proposed by [28], the term C can beobtained by geometrical considerations applied to cropsdistributed in rows. Therefore, it is difficult to obtain thecavity term from satellite data. Reference [29] give aconstant value for a cavity term, not depending on theLAI (Leaf Area Index) or FVC value, which could notbe true in all cases. The retrieval of the cavity term fromsatellite data is not an easy task and requires furtherresearch. For this reason it has not been included in thispaper. Anyway, when the cavity effect could beevaluated (for example from in situ measurements), itshould be added to the emissivity value retrieved fromEquation (13). Additional discussion about the validityof Equation (13) can be found in [30].The fraction of vegetation cover or FVC is akey variable in many environmental studies. Differentapproaches have been published in order to retrieve thisparameter from satellite data. Traditionally, theseapproaches have used relations between FVC andvegetation indices. In this paper relationships betweenthe FVC and the NDVI and VARI indices adapted forCHRIS data have been analyzed and tested. Bothrelations provide good results, especially the FVC vsVARI approach, with errors lower than 10%. Thisapproach has also the advantage of using TOA (Top OfAmosphere) data, so the atmospheric correction is notneeded. The availability of several spectral bands allowsthe application of other techniques for FVC retrieval, asfor example the Spectral Mixture Analysis or, morespecifically, the Linear Spectral Unmixing. Thistechnique has been applied using the AMEE method forextracting the endmembers, with error lower than 12%.The results slightly improve when extracting theendmembers from the image with a priori knowledge ofthe field site, with errors in this case lower than 10%.Angular effects on vegetation indices and FVChave been also analyzed. The results show lowvariations on NDVI with view angle, also leading to lowangular variations when estimating the FVC from NDVIand also with the other techniques. These low a

ABSTRACT/RESUME In this paper, two different methods for fractional vegetation cover (FVC) retrieval from CHRIS (Compact High Resolution Imaging Spectrometer) data based on vegetation indices have been analyzed. The first method uses NDVI (Normalized Difference Vegetation