empirical estimation of leaf chlorophyll density in winter wheat canopies using sentinel-2 spectral...

16
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 1 Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution Massimo Vincini, Member, IEEE, Stefano Amaducci, and Ermes Frazzi Abstract—A comparison between the sensitivities to leaf chloro- phyll density at the canopy scale of several vegetation indices (VIs) obtained at different spectral resolutions was carried out using spectral reflectance collected in winter wheat field trials with different nitrogen fertilization levels. A total of 350 spectra were collected from experimental plots at Feekes growth stages 5, 6, and 9 using a portable spectroradiometer (ASD FieldSpec HH), along with Minolta SPAD measurements of leaf optical thickness as a proxy for leaf chlorophyll density. Indices based on visible and near-infrared (NIR) bands were obtained from average reflectance in spectral ranges corresponding to SPOT HRG and Sentinel-2 (S2) bands. Indices requiring a red-edge band were obtained from reflectance at the originally proposed VI wavelengths using the 1.6-nm nominal spectral resolution bandwidth of the spec- troradiometer and from average reflectance in the S2 red-edge bands with the closest spectral position to VI originally proposed wavelengths. Among VIs obtained from Sentinel-2 bands MERIS terrestrial chlorophyll index, red-edge position and triangular chlorophyll index/optimized soil adjusted VI ratio (TCI/OSAVI) indices, obtainable at 20-m spatial resolution from future S2 red-edge bands, and chlorophyll VI (CVI), obtainable at 10 m from visible and NIR bands, were the best estimators of winter wheat leaf chlorophyll density. The sensitivity of the best-performing indices obtained from S2 bands to winter wheat with other con- ditions was addressed by the analysis of a large synthetic data set obtained using the PROSPECT–SAILH model in the direct mode. Analysis of the synthetic data set using Sentinel-2 spectral resolu- tion indicates that the two leaf area index normalized (TCI/OSAVI and CVI) indices are better leaf chlorophyll estimators. Index Terms—Leaf chlorophyll, Sentinel-2, variable-rate fertil- ization, vegetation indices (VIs), winter wheat. I. I NTRODUCTION N ITROGEN fertilization is a key factor for productivity, energy balance, and environmental impact of crops. Ef- fective plant nitrogen status diagnostic tools are required to guide spatially variable in-season N applications, a promising means to improve crops nitrogen use efficiency in precision agriculture [1]–[3]. Leaf chlorophyll density (i.e., chlorophyll Manuscript received November 12, 2012; revised February 27, 2013 and May 7, 2013; accepted May 29, 2013. M. Vincini and E. Frazzi are with the Centro Ricerca Analisi geoSpaziale e Telerilevamento (CRAST), Università Cattolica del Sacro Cuore, 29122 Piacenza, Italy (e-mail: [email protected]; [email protected]). S. Amaducci is with the Istituto di Agronomia Generale, Genetica e Colti- vazioni erbacee, Università Cattolica del Sacro Cuore, 29122 Piacenza, Italy (e-mail: [email protected]). Digital Object Identifier 10.1109/TGRS.2013.2271813 mass per unit leaf area in micrograms per square centimeter) is sensitive to soil N availability and is probably the most effective biophysical indicator of N deficiency [4]–[6]. Handheld leaf chlorophyll meters, such as SPAD-502 (Minolta Ltd., Japan), are being used for rapid nondestructive assessment of leaf chlorophyll density [3], [7]–[10] and in-season optimum N dose determination in traditional (i.e., not spatially variable) N application. To avoid the effect of genotype and environ- ment [2], chlorophyll meter measurements are being used in relative terms, e.g., as the ratio of the SPAD readings of an area with limited N to the SPAD reading where N was not limited. However, to collect leaf chlorophyll spatial data required by spatially variable N application by handheld chlorophyll meters is expensive and time-consuming. Hence, proximal and remote optical sensors (i.e., on-the-go active sensors mounted on N fertilizer applicators [3], [10], [11] and airborne or spaceborne optical sensors [1], [12]) are seen as important tools to provide timely leaf chlorophyll spatial estimations from canopy reflectance for in-season site-specific N prescriptions. Due to the fact that in-season N applications are often carried out at early growth stages [3], such spectral indicators, to be effective for this application, should be sensitive to leaf chlorophyll density for low leaf area index (LAI) values. Effec- tive spectral estimators should be sensitive to leaf chlorophyll density before canopy closure, when bare soil optical properties are still affecting canopy reflectance. For instance, intensive management studies conducted on winter wheat have shown that split topdressings of N fertilizer after spring green-up improve N efficiency and increase yields [7]. Spectral indicators specifically sensitive to chlorophyll density (i.e., insensitive to LAI variations) are desirable for this application because LAI variations are often linked to soil water stress rather than plant available N [13]. N-rich regions can be misidentified as N-limited regions when water stress mimics N stress by lower- ing LAI [14]. Two intrinsically different main approaches exist in remote sensing for estimating vegetation biophysical parameters for precision agriculture applications from optical reflectance mea- surements: using spectral vegetation indices (VIs) or inverting physically based canopy reflectance models. Since canopy re- flectance depends on a number of scene- and sensor-specific factors, the portability of the relationship between the different VI and crop biophysical parameters is limited [1], [15], [16]. The reliability of VI as leaf chlorophyll estimators, although 0196-2892/$31.00 © 2013 IEEE

Upload: ermes

Post on 16-Dec-2016

212 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 1

Empirical Estimation of Leaf Chlorophyll Density inWinter Wheat Canopies Using Sentinel-2

Spectral ResolutionMassimo Vincini, Member, IEEE, Stefano Amaducci, and Ermes Frazzi

Abstract—A comparison between the sensitivities to leaf chloro-phyll density at the canopy scale of several vegetation indices(VIs) obtained at different spectral resolutions was carried outusing spectral reflectance collected in winter wheat field trials withdifferent nitrogen fertilization levels. A total of 350 spectra werecollected from experimental plots at Feekes growth stages 5, 6,and 9 using a portable spectroradiometer (ASD FieldSpec HH),along with Minolta SPAD measurements of leaf optical thicknessas a proxy for leaf chlorophyll density. Indices based on visible andnear-infrared (NIR) bands were obtained from average reflectancein spectral ranges corresponding to SPOT HRG and Sentinel-2(S2) bands. Indices requiring a red-edge band were obtainedfrom reflectance at the originally proposed VI wavelengths usingthe 1.6-nm nominal spectral resolution bandwidth of the spec-troradiometer and from average reflectance in the S2 red-edgebands with the closest spectral position to VI originally proposedwavelengths. Among VIs obtained from Sentinel-2 bands MERISterrestrial chlorophyll index, red-edge position and triangularchlorophyll index/optimized soil adjusted VI ratio (TCI/OSAVI)indices, obtainable at 20-m spatial resolution from future S2red-edge bands, and chlorophyll VI (CVI), obtainable at 10 m fromvisible and NIR bands, were the best estimators of winter wheatleaf chlorophyll density. The sensitivity of the best-performingindices obtained from S2 bands to winter wheat with other con-ditions was addressed by the analysis of a large synthetic data setobtained using the PROSPECT–SAILH model in the direct mode.Analysis of the synthetic data set using Sentinel-2 spectral resolu-tion indicates that the two leaf area index normalized (TCI/OSAVIand CVI) indices are better leaf chlorophyll estimators.

Index Terms—Leaf chlorophyll, Sentinel-2, variable-rate fertil-ization, vegetation indices (VIs), winter wheat.

I. INTRODUCTION

N ITROGEN fertilization is a key factor for productivity,energy balance, and environmental impact of crops. Ef-

fective plant nitrogen status diagnostic tools are required toguide spatially variable in-season N applications, a promisingmeans to improve crops nitrogen use efficiency in precisionagriculture [1]–[3]. Leaf chlorophyll density (i.e., chlorophyll

Manuscript received November 12, 2012; revised February 27, 2013 andMay 7, 2013; accepted May 29, 2013.

M. Vincini and E. Frazzi are with the Centro Ricerca Analisi geoSpazialee Telerilevamento (CRAST), Università Cattolica del Sacro Cuore, 29122Piacenza, Italy (e-mail: [email protected]; [email protected]).

S. Amaducci is with the Istituto di Agronomia Generale, Genetica e Colti-vazioni erbacee, Università Cattolica del Sacro Cuore, 29122 Piacenza, Italy(e-mail: [email protected]).

Digital Object Identifier 10.1109/TGRS.2013.2271813

mass per unit leaf area in micrograms per square centimeter) issensitive to soil N availability and is probably the most effectivebiophysical indicator of N deficiency [4]–[6]. Handheld leafchlorophyll meters, such as SPAD-502 (Minolta Ltd., Japan),are being used for rapid nondestructive assessment of leafchlorophyll density [3], [7]–[10] and in-season optimum Ndose determination in traditional (i.e., not spatially variable)N application. To avoid the effect of genotype and environ-ment [2], chlorophyll meter measurements are being used inrelative terms, e.g., as the ratio of the SPAD readings of anarea with limited N to the SPAD reading where N was notlimited.

However, to collect leaf chlorophyll spatial data requiredby spatially variable N application by handheld chlorophyllmeters is expensive and time-consuming. Hence, proximal andremote optical sensors (i.e., on-the-go active sensors mountedon N fertilizer applicators [3], [10], [11] and airborne orspaceborne optical sensors [1], [12]) are seen as importanttools to provide timely leaf chlorophyll spatial estimations fromcanopy reflectance for in-season site-specific N prescriptions.Due to the fact that in-season N applications are often carriedout at early growth stages [3], such spectral indicators, tobe effective for this application, should be sensitive to leafchlorophyll density for low leaf area index (LAI) values. Effec-tive spectral estimators should be sensitive to leaf chlorophylldensity before canopy closure, when bare soil optical propertiesare still affecting canopy reflectance. For instance, intensivemanagement studies conducted on winter wheat have shownthat split topdressings of N fertilizer after spring green-upimprove N efficiency and increase yields [7]. Spectral indicatorsspecifically sensitive to chlorophyll density (i.e., insensitiveto LAI variations) are desirable for this application becauseLAI variations are often linked to soil water stress rather thanplant available N [13]. N-rich regions can be misidentified asN-limited regions when water stress mimics N stress by lower-ing LAI [14].

Two intrinsically different main approaches exist in remotesensing for estimating vegetation biophysical parameters forprecision agriculture applications from optical reflectance mea-surements: using spectral vegetation indices (VIs) or invertingphysically based canopy reflectance models. Since canopy re-flectance depends on a number of scene- and sensor-specificfactors, the portability of the relationship between the differentVI and crop biophysical parameters is limited [1], [15], [16].The reliability of VI as leaf chlorophyll estimators, although

0196-2892/$31.00 © 2013 IEEE

Page 2: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

they are being used for most operational applications, is limitedby the fact that usually they poorly account for site- and scene-specific conditions, especially soil background when upscaledfrom leaf to canopy level. In addition, they usually are notspecifically sensitive to the characteristic of interest (i.e., struc-tural (LAI) or biophysical (leaf chlorophyll) parameters ofvegetation) [17].

In the last three decades, physically based canopy reflectancemodels describing the interaction and transfer of electromag-netic radiation inside the canopy have been developed. Thesemodels aim to more easily deal with site- and scene-specificconditions. Biophysical vegetation parameters can be spatiallyestimated by the inversion of radiative transfer models (i.e.,by the determination of the parameters that minimize the dif-ferences between the measured and modeled reflectance data).The popular coupled leaf and canopy PROSPECT+SAILHreflectance model [18]–[20] has been widely used in a numberof studies for estimating vegetation biophysical parameters.Model inversion, however, due to the complexity of the de-scription of the radiative transfer within the canopy and ofthe inversion process itself requires significant computationalresources and is impractical for large data sets. Furthermore,model inversions are hampered by the ill-posed nature of theinversion process, which leads to unstable inversion results dueto the counterbalance of some parameters [21], [22]. Differentsets of model input parameters (e.g., a sparse canopy withplanophile leaf orientation and a dense erectophile canopy) mayyield to similar spectra. The radiometric information is notsufficient to identify a unique solution, and the inverse problemneeds to be regularized by exploiting additional informationsuch as prior knowledge on canopy and soil input variables. Forthe same crop and genotype, all the solutions sharing similarspectral reflectances tend to have similar values of canopyintegrated chlorophyll content [12]. The chlorophyll densityper unit crop area (i.e., canopy chlorophyll density) is obtainedas the product between chlorophyll density per unit leaf areaand LAI. Hence, VI’s specific empirical sensitivity to LAI orleaf chlorophyll density can be used to regularize the inver-sion problem. Inversion techniques based on a precomputedreflectance database, typically neural networks and look-uptables, are being used for operational estimation of vegeta-tion biophysical parameters from reflectance data collected bymedium spatial resolution sensors such as Moderate ResolutionImaging Spectroradiometer and Medium Resolution ImagingSpectrometer (MERIS) [23]. Their estimation performancesrely on the training database, which, due to the ill-posed natureof the inverse problem, must be integrated with some a prioriinformation [23], [24].

Due to its simplicity, the empirical approach has provedto be desirable for operational mapping of crop biophysicalparameters. Much effort has been devoted to improve thespecific sensitivity of the various proposed empirical indices toleaf chlorophyll density at the canopy level. Several VIs withimproved sensitivity, obtainable by remote or proximal sensing,have been proposed in the literature for the empirical spatialestimation at the canopy scale of leaf chlorophyll. However,the classical normalized difference VI (NDVI) [25], whichis mainly sensitive to canopy structural parameters (i.e., leaf

area), scarcely sensitive to leaf chlorophyll density, and heavilyaffected by soil reflectance [13], [26]–[28], is still consideredas one of the main optical empirical indicators of crops’ Nnutritional status at the operational level [3], [29]. Indeed, littleagreement exists on which VI or spectral transformation hasthe strongest relationships with leaf chlorophyll at the canopyscale for low LAI values. Studies addressing the comparison ofVI’s sensitivity to leaf chlorophyll are periodically being pub-lished, usually with contrasting results [30]–[34]. For instance,the MERIS terrestrial chlorophyll index (MTCI) developed byDash and Curran [35] and used by the European Space Agencyas a MERIS level-2 product appeared to be the most suitableindex for the retrieval of chlorophyll content from MERIS dataand showed the highest correlation levels with leaf chlorophyllin a comparison based on laboratory and field measurements incorn and wheat canopies [32]. However, it has been reportedto have a nonnegligible root-mean-square error (rmse) relatedto the optical properties of bare soils and to perform poorly asa leaf chlorophyll estimator on wheat canopies reconstructedin the laboratory at early growth stages (i.e., the phenologywindow for in-season N fertilization) [31].

The red edge (680–780 nm) is the region between maximumchlorophyll absorption in the red and maximum reflection in thenear infrared (NIR) caused by leaf cellular structure abundance(i.e., LAI). The spectral behavior of vegetation at the red edgeis mainly affected by chlorophyll density and, to a lesser extent,by LAI. Several VIs requiring one or more high-spectral-resolution bands in the red edge have proved to be sensitivefor different crops to leaf chlorophyll density at the canopyscale. In general, these leaf chlorophyll estimators are basedon measurements of the depth of the chlorophyll absorption inthe visible relative to the reflectance in the red-edge spectralregion [6], [17], [32], or they focus only on the red edge [35]–[37], thus requiring at least two narrow spectral bands in the rededge. Indices based on the calculation of chlorophyll absorptiondepth have been reported to have a nonlinear relationship withthe measured chlorophyll density, whereas indices based on thered-edge region show a linear relationship with the measuredchlorophyll content [32].

A few VIs based on the green and NIR reflectances [38], [39],or including also reflectance in the red [40], have been proposedwith the same objective. Among the few indices used for Nprescriptions by on-the-go active proximal sensors mountedon N applicators [10], [11], there are the “green NDVI” [38]and the “green simple ratio” or its slight modification [39] [thechlorophyll index (CI)].

A further general distinction among VIs proposed as leafchlorophyll estimators can be made between VI normalizedfor LAI (using other VI more sensitive to LAI [17], [32] orreflectance ratios in the visible [40]) and indices that are notnormalized for LAI. In general, indices based on measurementsof the depth of the chlorophyll absorption in the visible areLAI normalized to be specifically sensitive to leaf chlorophyll,whereas indices such as the MTCI that focus on the red edgeare not, even though in literature there are indications that theyare somewhat sensitive to LAI [31], [32].

To briefly summarize the state of the art, little agree-ment exists on whether the empirical estimation of crop leaf

Page 3: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

VINCINI et al.: EMPIRICAL ESTIMATION OF LEAF CHLOROPHYLL DENSITY IN WINTER WHEAT CANOPIES 3

TABLE ISENTINEL-2 SPECTRAL BAND SPECIFICATIONS

IN THE VISIBLE–NIR SPECTRAL RANGE

chlorophyll, at early growth stages, can be effectively based onthe following:

1) indices based on measurements of the depth of the chloro-phyll absorption in the visible relative to reflectance in thered edge, normalized for LAI;

2) indices focusing on the red edge, not LAI normalized;3) indices based exclusively on visible and NIR reflectances.

The Sentinel-2 (S2) constellation is expected to represent asignificant advancement in precision agriculture applications,with the pair of satellites in operation S2 having a revisit timeof 2–3 days at mid-latitudes [41]. Sentinel-2 data will includethree visible and one NIR spectral bands at 10-m (very high)spatial resolution and three spectral bands in the red edge at20-m (high) resolution. Sentinel-2 spectral band specificationsin the visible–NIR spectral range are reported in Table I.

Given the disagreement in the literature, it is not clear whichVI is the best empirical leaf chlorophyll estimator that canbe used for an effective exploitation of future S2 data inprecision agriculture. Besides the innovative availability of red-edge bands at high spatial resolution from a spaceborne sensor,S2 visible band spectral resolution will be much higher thanthose of previous high-spatial-resolution spaceborne multispec-tral sensors. On the other hand, S2 spectral resolution (Table I)will not provide all the bands and wavelengths required by allproposed leaf chlorophyll estimators at the spatial resolutionrequired for precision agriculture. For instance, the normalizedpigment chlorophyll ratio index [42] reported as the best per-forming leaf chlorophyll estimator on winter wheat canopies[31] includes reflectance in the violet (430 nm), but the S2band with the closest spectral position, B1 band centered at443 nm (Table I), has a moderate spatial resolution configuredfor atmospheric applications. The double difference index wasthe best leaf chlorophyll estimator on a large synthetic database[43], but it requires three bands in the red edge (700, 720, and750 nm), not provided by S2 spectral resolution. For other leafchlorophyll estimators, the originally proposed wavelengths arenot closely matched by S2 band center wavelengths.

The availability of high spatial and spectral resolution S2 vis-ible and red-edge bands can significantly improve the effective-ness of VI application to variable-rate fertilization prescriptionsin precision agriculture. However, for an effective exploitationof S2 data for this application, the following issues should beaddressed, at least for main grain cereals (i.e., corn, wheat, andrice), before operational activities.

1) Is it possible to obtain fairly accurate 10-m resolutionempirical spatial estimates of leaf chlorophyll densityover crop canopies at early growth stages using indicesbased on S2 visible and NIR bands?

2) Which is the best-performing VI for the same applicationusing S2 20-m spatial resolution red-edge bands?

Among grain cereals, wheat is grown on more land area thanany other crop and is the main staple food for humans. Thepresent work addresses the comparison between the sensitivityof several VI, obtainable from future Sentinel-2 bands, to leafchlorophyll density at the canopy scale in winter wheat spec-trometric field trials with different nitrogen fertilization levels.In addition, the sensitivity of leaf chlorophyll estimator indicesobtained from S2 bands to winter wheat with other conditions(i.e., soil background, genotype, and sun elevation) is addressedin this paper by the analysis of a large synthetic data set.

II. METHODS

A. Field Trial Spectrometry

Spectral reflectances of wheat canopies in early growthstages in the visible–NIR spectral range were collected byfield spectrometry in the context of a field trial with differentfertilization levels and removal or plow in of off-shoots of thepreceding sorghum crop. Four levels of nitrogen fertilizationto the winter wheat crop (0, 60, 120, and 180 kg · ha−1) wereequally split in two dressings at Feekes [44] growth stages5 (pseudostem) and 6 (first node visible). A total of 350spectral reflectances in the 350–1075-nm range were collectedaround solar noon over experimental plots at Feekes 5, 6,and 9 (127 spectral reflectances at day of year (DOY) 90,128 at DOY 102, and 95 at DOY 133) along with SPADmeasurements of leaf optical thickness as a proxy of leafchlorophyll density. Two spectral reflectances per plot werecollected on a regular grid over 64 experimental plots (5 ×5 m). One spectral reflectance was not correctly acquired forDOY 90, and due to wheat growth, only 95 spectral reflectancescould be collected for DOY 133, less relevant for fertilizationpurposes. As mentioned, Minolta SPAD chlorophyll meter iscommonly used to make nitrogen recommendations on winterwheat for the spring application (i.e., at Feekes phenologicalstages 4–6). Field spectral data were collected using a 512-channel ASD FieldSpec HH portable spectroradiometer witha nominal spectral resolution of 1.6 nm and an FWHM spectralresolution of approximately 3 nm at around 700 nm. A footprintof about 0.35 m2 was obtained by holding the optical fiber, withfield of view of 25◦, about 1.5 m above the canopy. Beforeeach radiometric measurement, the radiance of a referencewhite panel (Spectralon) was measured in order to computethe vegetation spectral reflectance. Three chlorophyll metermeasurements were taken from the uppermost fully expandedleaf (i.e., ligule emerged from the sheath of the preceding leaf)of two plants within the spectroradiometer footprint. SPADmeasurements were converted to chlorophyll (a+ b) contentper unit leaf area (in micrograms per square centimeter) usingthe second-order polynomial calibration equation specificallyproposed for winter wheat by Bannari et al. [31]

Chla+b = 0.3015 · SPAD + 0.0102 · SPAD2. (1)

Equation (1) was shown to agree well with another calibra-tion equation proposed in literature for winter wheat [31].

Page 4: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

TABLE IIALGORITHMS, REFERENCES, AND SENTINEL-2 BANDS USED FOR VI CALCULATION OF VI REQUIRING ONLY VISIBLE AND NIR BANDS

(10-m SPATIAL RESOLUTION) AND VI REQUIRING RED-EDGE BANDS (20 m) ANALYZED IN THIS PAPER

Indices considered as leaf chlorophyll estimators in winterwheat field trials included only VI whose originally proposedwavelengths could be matched, at least roughly, by center wave-lengths of S2 bands at high spatial resolution (i.e., 10–20 m;Table I). Table II reports the algorithms with the originallyproposed bands (for indices obtained from visible and NIRbands, indicated as Vis–NIR VI from hereon), the center wave-lengths (for indices requiring red-edge bands), the references,and the Sentinel-2 bands used for VI calculation. Indices listedin Table II are briefly described hereafter.

B. Vis–NIR Indices

Indices obtained only from red and NIR bands focus onthe contrast between the spectral response of vegetation in thered and NIR parts of the spectrum. The ratio VI [45] or SR,calculated as the ratio of the reflectances in the NIR and red,has been reported as the best estimator of both LAI and canopychlorophyll density (i.e., chlorophyll density per unit crop area)for low LAI values [30].

In order to linearize the relationships with vegetation bio-physical variables, the renormalized difference VI (RDVI) hasbeen developed from NDVI [27].

The high sensitivity of the green band in comparison withother wavelengths in the visible spectral range to photosyntheticpigment content of vegetation was pointed out by early remotesensing studies [46] and later confirmed [4]. Gitelson et al. [38]proposed the “green NDVI” as a leaf chlorophyll estimator atthe canopy scale, using a green band rather than a red band,as in the classic NDVI. Schepers et al. [47] reported strongcorrelations between the narrow band 550-/850-nm reflectance

ratio (i.e., the inverse of the green SR) and the chlorophyllcontent for leaves of corn grown under different nitrogenregimes. Gitelson and Merzlyac [48] found similar results forthe 750-/550-nm reflectance ratio of maple and chestnut leaves.In maize and soybean canopies [39], close relationships werefound between the CI (a slight modification of the green SR;Table II) and chlorophyll content. Hunt et al. [49] proposed thetriangular greenness index (TGI) as a leaf chlorophyll densityestimator calculated as the area of the triangle defined by thegreen peak, the chlorophyll absorption maximum, and the blueband (480 nm).

In order to obtain an index incorporating the spectral in-formation of the green band with enhanced sensitivity toleaf chlorophyll content and insensitivity to LAI variation,Vincini et al. [40] developed the chlorophyll VI (CVI). The CVIindex is obtained from the green SR by introducing the red/green ratio to minimize the sensitivity to differences in thecanopy LAI before canopy closure. The CVI index, obtainedat SPOT HRG spectral resolution, has been reported to bespecifically sensitive to leaf chlorophyll density at the canopyscale in a wide range of soil and scene conditions [40] forplanophile crop canopies [33].

Indices incorporating bands in the green have proven to beeffective LAI and canopy chlorophyll density estimators. Thetriangular VI [30] is calculated as the area of the triangle definedby the green peak (550 nm), the red chlorophyll absorptionmaximum (670 nm), and the NIR shoulder (750 nm) in thespectral space. Haboudane et al. [50] developed, for LAI es-timation, the modified TVI (MTVI) by introducing a scalefactor and by replacing the 750-nm wavelength, sensitive to leafchlorophyll content, by the 800-nm wavelength.

Page 5: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

VINCINI et al.: EMPIRICAL ESTIMATION OF LEAF CHLOROPHYLL DENSITY IN WINTER WHEAT CANOPIES 5

C. Indices Including a Red-Edge Band

Indices requiring a red-edge band included only VI specif-ically proposed as leaf chlorophyll estimators at the canopyscale. The considered red-edge VI included indices normalizedfor LAI by taking the ratio with the optimized soil adjustedVI (OSAVI) index, indices completely or partially focusing onthe spectral behavior in the red edge [i.e., red-edge inflectionposition (REP), MTCI, and simplified red-edge model (R-M)],and one index [the normalized area over the reflectance curve(NAOC)], integrating reflectance over the whole red-NIR spec-tral range. LAI-normalized red-edge indices are based on mea-surement of the depth of the chlorophyll absorption in thevisible relative to reflectance in the red-edge spectral region(Table II).

The chlorophyll absorption reflectance index (CARI) hasbeen proposed as a measure of the depth of the chlorophyllabsorption at 670 nm relative to the green reflectance peak at550 nm [51]. The modified CARI (MCARI), developed to beresponsive to chlorophyll variation, is also sensitive to LAI vari-ations even though no NIR band is considered [6]. The 700-nmwavelength used in the MCARI, as well as in the transformedCARI (TCARI) and triangular chlorophyll index (TCI) algo-rithms (Table II), matches the boundary between the regionwhere vegetation reflectance is dominated by pigments and thebeginning of the red-edge portion where vegetation structuralcharacteristics tend gradually to dominate. The ratio ρ700/ρ670was introduced to minimize the effect of the underlying soilreflectance and nonphotosynthetic materials. Haboudane et al.[17] proposed for chlorophyll estimation the TCARI, anothermodification of the CARI index, and a variant of the TVI[30], the TCI modifying the use of the ρ700/ρ670 ratio or itssquared root as a background-minimizing factor. To decoupleLAI and chlorophyll sensitivity, Haboudane et al. [17] proposedthe combined TCARI/OSAVI ratio and the MCARI/OSAVI andTCI/OSAVI ratios [32], where the OSAVI [28] is introduced tominimize the sensitivity to differences in the canopy LAI.

Different methods have been proposed to calculate the spec-tral position of the inflection point (i.e., the maximum of thefirst derivative of the reflectance spectra) in the red edge or red-edge position (REP), sensitive to leaf chlorophyll at the canopyscale [36], [37]. A simplified procedure based on a linear in-terpolation between NIR and red reflectances was proposed byGuyot and Baret [52] and modified by [53] and [54]. The linearinterpolation method for REP estimation based on Sentinel-2center-band wavelengths (REP in Table II) is considered in thepresent study. This method has been used in a preliminary pub-lication of field trial partial results [55], presenting relationshipsbetween some VI and SPAD measurements not calibrated toleaf chlorophyll density.

The MTCI, obtained from reflectances at 681-, 709-, and754-nm center wavelengths of the MERIS standard band set-ting, has been proposed by Dash and Curran [35] as a leafchlorophyll estimator at the canopy scale. The MTCI indexis strongly correlated with the REP but more sensitive to thechange in chlorophyll content at high chlorophyll contents [56].

Gitelson et al. [39] developed a conceptual method basedon red-edge reflectance for chlorophyll estimation in crop

canopies. A simplified form of the model, the R-M index [32],is considered in the present study (Table II).

The NAOC index [57] estimates the relative area over thereflectance curve in a spectral region extending from the max-imum chlorophyll absorption (red) to the early NIR. The bestresults from NAOC are obtained by setting the integration limitsat 643 and 795 nm [57]. At Sentinel-2 spectral resolution, theNAOC index can be obtained from B4–B5–B6–B7 bands, andthe integration limits can be set at 665- (B4) and 783-nm (B7)center wavelengths.

VIs were obtained from average reflectance in spectral rangescorresponding to SPOT HRG bands (500–590, 610–680, and780–890 nm) and S2 bands (Table II) and from reflectanceat the original spectral resolution of the spectroradiometer(1.6-nm nominal bandwidth). Indices requiring one or morenarrow bands in the red edge and the TGI index, requiring ablue band, could not be obtained at SPOT spectral resolution.Sentinel-2 bands with the closest spectral position to VI’soriginal wavelengths were selected for VI calculation.

D. Comparing the Best-Performing VI Using a SyntheticData Set

The popular PROSPECT+SAILH leaf and canopy coupledreflectance model [18]–[20], [58] was used in the direct modein order to obtain two synthetic data sets representing winterwheat at two different early growth stages, suitable for springdressing, with different LAI and leaf chlorophyll density values,in a wide range of soil, genotype, and illumination geometryconditions. The synthetic data sets were then used to comparethe sensitivity in different soil/genotype/illumination conditionsof VI, obtained at Sentinel-2 spectral resolution, which showedthe strongest leaf chlorophyll correlation in winter wheat fieldtrials as well as in other experiments reported in literature [17],[32], [33], [40].

The soil reflectance database [59] used as model input in-cluded the spectral reflectances of six different soils (Fig. 1)representing the large majority of the spectral variability ofmid-latitude cropland topsoils. For each soil characterized bya large reflectance variability between wet and dry conditions(i.e., Othello, Cecil, Portneuf, and Cordorus in Fig. 1), twospectral reflectances were used, representing wet and dry soilconditions (respectively, solid and dashed lines in Fig. 1),whereas for soils with little variability in soil reflectance relatedto soil moisture (Barnes and Houston Black Clay), a singlespectral reflectance was used, representing intermediate soilmoisture condition.

The acquisition geometries considered in the syntheticdatabase included only nadiral observation (i.e., with zero viewzenith angle) and two solar zenith angles (15◦ and 45◦). Themodel input parameters that were varied, in addition to soilbackground, to obtain the synthetic data sets are summarizedin Table III. Leaf chlorophyll (a+ b) density varied from20 (i.e., leaf chlorosis) to 50 μg · cm−2, with 2.5-μg · cm−2

increments. In earlier stages, the largest changes in wheatdetermined by N availabilty are in growth and not chlorophylldensity, and remotely sensed images quickly lose value. Atthose stages, on-the-go sensors are more effective at managing

Page 6: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 1. Spectral reflectances of six different wet (solid line) and dry (dashedline) cropland topsoils in the 400–1000-nm range used as SAILH model input(modified from [40]).

TABLE IIIPARAMETERS VARIED, IN ADDITION TO SOIL BACKGROUNDS,

TO SYNTHETIZE THE TWO PROSPECT–SAILH DATA SETS

REPRESENTING WINTER WHEAT AT TWO DIFFERENT EARLY

GROWTH STAGES (I.E., WITH TWO DIFFERENT LAI RANGES)

the variable-rate fertilization of large planted areas. On theother hand, remote sensing is probably the most cost-effectivemethod to determine N rates with relatively higher cover andLAI values. Bearing this in mind, two LAI value ranges withLAI variability plausibly determined by variations in N nutri-tional status [13] were considered to represent two winter wheatearly growth stages suitable for topdressings of N fertilizer.The data set representing the earlier growth stage consideredLAI values from 1.4 to 2.4, with 0.2 increments. A wider LAIrange (2.2–4.0 with 0.2 increments) was chosen for the secondstage in order to simulate marked but realistic LAI variability,determined by variations in N availability [13], at stages whenremotely sensed image data could be more effective at manag-ing N fertilization.

Within the theoretical framework of the PROSPECT model,the leaf is considered to be a series of flat homogeneous plateswhose number is accounted for by the mesophyll structureindex value (N). The lowest value accepted by the model is

N = 1. Given the relevant contribution of N structure indexvalue input to the PROSPECT model output in the NIR spectralregion [60], [61], in the present work, N = 1 and N = 2 areconsidered as inputs to the PROSPECT model. The two valuesrepresent the minimum and maximum of the N leaf structureparameter range typically used in literature to model wheatleaves [22], [32], [62], [63]. Similarly, in order to representplausible winter wheat genotype variability, two leaf orien-tations encompassing the typical range used in literature tomodel wheat canopies were used as models input, with averageleaf angle (ALA) values of 60◦ and 70◦, respectively. Thereflectance for a given wavelength in the visible–NIR spectralrange will monotonically increase or decrease, respectively,for increasing N values and ALA values [60] in the selectedparameter ranges. Hence, if the same VI relative effective-ness is found for the minimum and maximum winter wheatpossible N values or ALA angle, it can be assumed that agenotype variation of the mesophyll structure parameter orof the ALA will not affect VI relative effectiveness as leafchlorophyll estimators. The hot spot is the solid angle (cone)where the solar and viewing directions are close together. Thecanopy hot-spot size parameter, depending on the mean sizeand shape of leaves and on canopy height, was introduced inthe SAILH model to properly reproduce the behavior of canopyreflectance especially in the hot spot. The hot-spot size param-eter was set at 0.1, a typical value for erectophile leaf orienta-tions. For other model input parameters that are less relevantfor practical remote sensing applications in the visible–NIRrange, suggested typical values were used: water content of0.012 g · cm−2, dry matter content of 0.005 g · cm−2, brownpigment content of 0 (arbitrary unit), and leaf surface roughnessangle of 59◦.

The two resulting synthetic data sets for the two LAIranges included simulated soil-canopy spectral reflectance datafor, respectively, 6240 and 10 400 different soil/genotype/illumination conditions (2 N leaf structural parameter values,2 leaf orientations, 10 soil types and moisture levels, 6 or10 LAI values, 13 leaf chlorophyll densities, and 2 sun zenithangles).

A power regression function was used in the present workto describe the relationship between VI and leaf chlorophyll(a+ b) density (Chl) values obtained from the synthetic datasets for different models’ input parameters:

VI = a · Chlb + c. (2)

The power model was selected from the commonly usedregression functions based on its monotonic behavior, suited toconsistently analyze the sensitivity of spectral indices to leafchlorophyll density. The chosen power model also has highsignificance levels for both the linear and nonlinear relation-ships between chlorophyll density and the different VI tested.In order to evaluate the effect of the regression function choice,the correlation levels obtained by the power function (2) weresystematically compared to those obtainable by using othercommon monotonic (linear and exponential) and nonmono-tonic regression models (quadratic and cubic polynomials) forOthello and Cecil soils’ subsets (i.e., for two soils with

Page 7: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

VINCINI et al.: EMPIRICAL ESTIMATION OF LEAF CHLOROPHYLL DENSITY IN WINTER WHEAT CANOPIES 7

markedly different spectral reflectances; Fig. 1). The results ofsuch comparison showed that the selected power model wasthe best monotonic regression model for both LAI ranges andfor all soil wetness/sun zenith/ALA/N leaf structural parametervalue combinations considered. In addition, results showedthat a negligible increase of regression R2 values could beobtained using nonmonotonic functions. In particular, for thetwo soil subsets, the R2 values of the REP index could beincreased using a cubic polynomial function by 0.011 and 0.017at maximum, respectively, for the lower and higher LAI rangesconsidered, and the R2 values of the other indices increased byless than 0.010 for both LAI ranges.

In addition to traditional regression-based statistics (R2 co-efficient of determination and rmse), changes in sensitivity of aVI over the range of leaf chlorophyll density were analyzed byusing a sensitivity function obtained according to the methodproposed by Ji and Peters [64]. The sensitivity function (3) iscalculated as the ratio of the first derivative of the regressionfunction (2) [using leaf chlorophyll density as the independentvariable (x) and the VI values as the dependent variable (y)]and the standard error σy of the predicted value (y)

s =dy/dx

σy. (3)

The sensitivity function, rather than providing a singlegoodness-of-fit value, can describe the changes in VI sensitivityover the range of biophysical variables and, being independenton the unit or magnitude of VI, can be used for a direct compar-ison of the performance of the various VIs [63]. The absolutevalue of s was considered in order to compare VI character-ized by positive (e.g., MTCI) and negative (e.g., TCI/OSAVI)correlations with leaf chlorophyll density.

III. RESULTS

A. Field Trial Spectrometry

Table IV gives the R2 and rmse values of power regres-sions (2) between leaf chlorophyll density per unit leaf area(in micrograms per square centimeter) versus VI obtained atdifferent spectral resolutions. The regressions’ R2 and rmse arereported separately for SPAD measurements collected at Feekesgrowth stage 5 (i.e., pseudostem) and aggregated measurementscollected at stages 5–6 [i.e., 6 (first node visible)] and 5–6–9[i.e., 9 (ligule of the last leaf just visible)]. A distinction isalso made in Table IV between VI obtainable from S2 bands at10-m [i.e., VI based on visible and NIR bands (Vis–NIR VI)] or20-m [i.e., VI requiring one or more spectral bands in thered edge (red-edge VI)] spatial resolution. The limited repro-ducibility of SPAD measurements should be taken into accountconsidering the R2 values reported in Table IV. SPAD measure-ments are collected on a few leaves in the spectroradiometerfootprint, and the mean SPAD measurements in the footprintare affected by plant, leaf [8], [9], and point of the leaf laminachosen.

At pseudostem (Feekes 5), before the first application ofdifferent nitrogen rates, the chlorophyll content showed a lim-ited variability, and as a consequence, both VI based on red-

TABLE IVREGRESSIONS’ R2 (TOP ROW, SIGNIFICANT AT p < 0.001) AND RMSE (IN

MICROGRAMS PER SQUARE CENTIMETER, BOTTOM ROW) VALUES ∗ OF

CHLOROPHYLL (a+ b) DENSITY VERSUS VIS–NIR VI (TOP) AND

VI REQUIRING RED-EDGE BANDS (BOTTOM) OBTAINED AT SPOT,SENTINEL-2 (IN BRACKETS), AND ASD SPECTRORADIOMETER

(IN SQUARE BRACKETS) SPECTRAL RESOLUTIONS FOR WHEAT

FEEKES GROWTH STAGES 5, 5–6, AND 5–6–9 AGGREGATED

edge bands (0.23 < R2 < 0.44; Table IV, bottom) and VIbased on visible and NIR bands (0.17 < R2 < 0.41; Table IV,top) show low correlation levels. The correlations of bothred-edge (0.36 < R2 < 0.72) and Vis–NIR VIs (0.14 < R2 <0.65) markedly improve after the aggregation of measurementscollected at Feekes growth stage 6 (i.e., first node visible)when the first differentiated nitrogen application had caused alarger variability in leaf chlorophyll density. A further marginalimprovement of VI correlation levels is obtained by the ag-gregation of data collected at Feekes growth stage 9, whichis, however, characterized by much higher LAI values and isless relevant for crop nutrient management. Albeit the highercorrelation levels are obtained by the aggregation of data of thedifferent growth stages, given the lower slopes of the regressionlines over the wider leaf chlorophyll content ranges, rmse also

Page 8: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

tends to increase with such aggregation from a range of 2.1 <rmse < 2.5 μg · cm−2 (Table IV) at Feekes stage 5 to a range of2.8 < rmse < 4.4 μg · cm−2 at Feekes stage 5–6–9 aggregated.

Among Vis–NIR VIs, the strongest correlations (maxi-mum R2 and minimum rmse values in bold in Table IV) atSPOT spectral resolution were obtained by the RDVI (Feekesstages 5–6 and 5–6–9 aggregated) or by the CVI (Feekesstages 5 and 5–6–9 aggregated) indices and by the CVI in-dex, for all considered growth stage data sets at S2 spectralresolution.

At ASD spectroradiometer resolution, the CVI index ob-tained the strongest correlations for Feekes stages 5–6 and5–6–9 aggregated, and green NDVI and green SR showed acorrelation slightly higher than CVI at Feekes stage 5. Withthe exception of the CVI index, the other Vis–NIR VI in-dices proposed specifically as leaf chlorophyll estimators atthe canopy scale (i.e., the green NDVI and the green SR orCI) appear to be just slightly more effective estimators ofleaf chlorophyll density than traditional NDVI and SR indices.Moreover, the green NDVI and the green SR are outperformedby VI nonspecifically proposed as chlorophyll estimators suchas the RDVI or OSAVI and by the MTVI index specificallyproposed as a LAI estimator.

Due to their common dependence on nitrogen availability[65], LAI and leaf chlorophyll density are, in general, directlyrelated in field nitrogen fertilization trials [17]. Specific ex-perimental treatments such as a temporary exclusion of light(i.e., etiolating) have been used in literature [40] to decoupleLAI and leaf chlorophyll density in studies addressing VIspecific sensitivity to the photosynthetic pigment. Among theconsidered Vis–NIR VIs, the SR, NDVI, RDVI, and OSAVIindices are sensitive both to vegetation structural parameters(i.e., LAI) and leaf chlorophyll density. As expected, LAI andleaf chlorophyll density are also correlated in the field trialexperimental data sets of the present study. It is interestingto note in Table IV that, for all considered growth stage datasets, only Vis–NIR VIs proposed as leaf chlorophyll estimators,particularly the CVI index, improve their correlation with leafchlorophyll density by using the Sentinel-2 higher spectralresolution. In contrast, due to their sensitivity to LAI NDVI,RDVI, OSAVI, and MTVI obtain higher correlations withleaf chlorophyll density at SPOT HRG lower spectral resolu-tion. The higher S2 spectral resolution allows the exploitationof the specific sensitivity of chlorophyll estimators. In particular,the high Sentinel-2 spectral resolution in the visible spectralrange enhances the effectiveness of the red/green ratio as a LAInormalization factor in the CVI algorithm, which is stronglydependent on spectral resolution [40]. The CVI R2 value atgrowth stages 5–6 is increased by the higher S2 spectral res-olution from 0.53 to 0.64, whereas by comparison, the R2 ofthe green SR index, from which the CVI is obtained by theintroduction of the red/green ratio, increases only from 0.46to 0.50. A similar trend is also shown by aggregated data ofFeekes growth stages 5–6–9 (Table IV). The CVI index, ob-tained at SPOT HRG spectral resolution, has been shown to besensitive to leaf chlorophyll density, especially for planophilecrop canopies [33]. However, it appears that the increased S2spectral resolution enhances the effectiveness of the red/green

ratio as a LAI normalization factor in the CVI (i.e., the CVIspecific sensitivity to leaf chlorophyll density) for erectophilecrops such as winter wheat.

The MTCI index was the best leaf chlorophyll estimatoramong red-edge VIs obtained both at S2 and ASD spectrometerspectral resolutions for all considered growth stage data sets(Table IV, bottom), with the exception of Feekes stage 5 dataset at S2 spectral resolution for which REP showed a slightlyhigher correlation level. Besides the MTCI index, the REPand TCI/OSAVI indices were the best leaf chlorophyll densityestimators, with the values of R2 and rmse very close to thoseof the MTCI (Table IV), especially for REP.

In comparison with the ASD 1.6-nm nominal spectral resolu-tion, the lower Sentinel-2 resolution does not affect, or slightlyaffects, the strength of the correlation versus leaf chlorophylldensity of the best red-edge leaf chlorophyll estimators MTCI,REP, and TCI/OSAVI, whereas in general, the other red-edgeVIs tend to show better correlation levels when obtained atthe ASD highest spectral resolutions (Table IV, bottom). Inparticular, the effectiveness as leaf chlorophyll density esti-mator appears to be markedly relying on the highest spectralresolution for the NAOC, TCARI/OSAVI, and simplified R-Mmodel indices.

Fig. 2 shows the scatterplots and power fittings of leaf chloro-phyll density versus the best Vis–NIR or red-edge leaf chloro-phyll density estimators CVI, MTCI, REP, and TCI/OSAVI(i.e., indices showing R2 values higher than 0.6 for Feekesgrowth stages 5–6 aggregated; Table IV), obtained from theaverage reflectance in Sentinel-2 bands. Scatterplots and fit-tings visible in Fig. 2 confirm that the MTCI index, the bestchlorophyll density estimator, is nonlinearly related to REP andis more sensitive to the changes in high chlorophyll contents[56], as shown by the slope of the regression lines in Fig. 2.The more linear relationship of the REP index with chlorophylldensity [32] is confirmed as well. Although indices based on thecalculation of chlorophyll absorption depth have been reportedto have a nonlinear relationship with the measured chlorophyllcontent, the fittings of chlorophyll density versus TCI/OSAVIratio in Fig. 2 do not seem to noticeably deviate from linearity[32]. As visible in Fig. 2, the aggregation of the data of Feekesgrowth stage 9 does not seem to change markedly the relation-ship between VI and leaf chlorophyll density, especially for theMTCI and CVI indices. Feekes growth stage 9 is characterizedby much higher LAI values and is less relevant for nutrientmanagement. It is interesting to note the similarity between theCVI and MTCI scatterplots reported in Fig. 2.

Fig. 3 shows the scatterplots and linear fittings of the MTCIversus CVI indices and of the REP versus TCI/OSAVI indices,obtained at Sentinel-2 spectral resolution for data collected atFeekes growth stages 5–6–9 aggregated. Despite their differenttheoretical backgrounds and different considered bands, indicesfocusing on the red edge tend to be linearly related to LAI-normalized indices: The MTCI index is strongly linearly relatedto the CVI, whereas the REP shows an inverse linear rela-tionship with TCI/OSAVI index. Conversely, the two indicesstrictly focusing on the red edge (i.e., REP and MTCI) andthe two LAI-normalized indices (i.e., TCI/OSAVI and CVI) arenonlinearly related to each other.

Page 9: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

VINCINI et al.: EMPIRICAL ESTIMATION OF LEAF CHLOROPHYLL DENSITY IN WINTER WHEAT CANOPIES 9

Fig. 2. Scatterplots and power fittings (y = a · xb + c) of chlorophyll (a+ b) density (in micrograms per square centimeter) versus the best leaf chlorophylldensity estimators (CVI, top-left; MTCI, top-right; REP, bottom-left; and TCI/OSAVI, bottom-right) obtained at Sentinel-2 spectral resolution for data collectedat Feekes growth stages 5–6 and 5–6–9 aggregated (Table IV).

Fig. 3. Scatterplot and linear fitting of the MTCI versus CVI indices (left) and of the REP versus TCI/OSAVI indices (right) obtained at Sentinel-2 spectralresolution for the data collected at Feekes growth stages 5–6–9 aggregated.

Page 10: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

TABLE VCOEFFICIENT OF DETERMINATION (R2, TOP ROW) AND RMSE (IN MICROGRAMS PER SQUARE CENTIMETER, BOTTOM ROW)

VALUES OF VI’S POWER FITTINGS VERSUS LEAF CHLOROPHYLL (a+ b) CONTENT FOR THE DATA SYNTHETIZED BY

DIFFERENT INPUT PARAMETERS: 60◦ (LEFT) AND 70◦ (RIGHT) ALAS, DIFFERENT SOIL TYPES AND SOIL MOISTURE

CONDITIONS [DRY, WET, OR INTERMEDIATE (INT.)], 15◦ (TOP) OR 45◦ (BOTTOM) SUN ZENITH ANGLES, AND NLEAF STRUCTURAL PARAMETER VALUES OF 1 OR 2; LAI RANGE OF 1.4–2.4—MAXIMUM R2 VALUE IN BOLD

The very close relationship (R2 = 0.96) between the MTCIand CVI indices seems to indicate that reflectance in the visibleand NIR Sentinel-2 bands is almost as effective as high-spectral-resolution reflectance data in the red edge at enhancing the VIspecific sensitivity to leaf chlorophyll density. This finding hasrelevant implications for the application of future Sentinel-2visible and NIR bands, obtainable at 10 m, a spatial resolutionmore suitable for variable-rate N fertilization prescriptions thanthe 20-m spatial resolution of the red-edge bands.

B. Comparing the Best-Performing VI in Winter Wheat FromSynthetic Data

Tables V and VI give the R2 and rmse values, respectivelyfor LAI ranges 1.4–2.4 and 2.2–4.0, of the power functionregression (2) of the four best chlorophyll density estimators(i.e., MTCI, REP, TCI/OSAVI, and CVI) obtained at S2 spectralresolution. Regressions are obtained from the synthetic dataset representing plausible PROSPECT+SAILH model inputranges for winter wheat genotypes (i.e., N leaf structural pa-rameter values of 1 or 2 and ALAs of 60◦ or 70◦) for differentsoil types, soil wetness conditions, and two sun zenith angles(15◦ and 45◦).

For almost all soil/soil wetness/sun zenith/ALA/N leaf struc-tural parameter value combinations considered and for bothLAI ranges, the strongest correlations (R2 values in bold inTables V and VI) between VI and leaf chlorophyll contentwere obtained by the TCI/OSAVI ratio (0.861<R2<0.999 and0.31< rmse<3.56 μg·cm−2) or by the CVI index (0.802<R2<0.999 and 0.35< rmse<4.21 μg·cm−2). Out of 160 con-

sidered LAI range/soil/soil wetness/sun zenith/ALA/N leafstructural parameter combinations, the TCI/OSAVI obtainedthe strongest correlation with leaf chlorophyll density in 101cases and the CVI in 58. In only one case out of 160 consideredcombinations (LAI range of 1.4–2.4, Othello soil wet, sunzenith of 15◦, ALA of 60◦, and N = 2), the REP index obtainedthe highest correlation level (R2 value of 0.933 versus 0.924 forthe CVI index; Table V).

Soil spectral properties, regardless of sun zenith/ALA/Nleaf structural parameter value conditions, appear to determinewhich index between the TCI/OSAVI and CVI is the bestleaf chlorophyll density estimator, with the former, in general,obtaining the strongest correlations for brighter/dry soils andthe latter for darker/wet soils (Fig. 1; Tables V and VI). Theworst performances of the CVI index (0.802 < R2 < 0.982and 1.30 < rmse < 4.21 μg · cm−2) were obtained with theCecil soil, characterized by the highest value of the red/greenreflectance ratio (Fig. 1), the ratio used by the CVI as a LAInormalization factor. Besides soil spectral properties, also theLAI range affects, to a lesser extent, the relative effectivenessof the considered VI as chlorophyll estimators: In six casesout of 80 considered combinations, the order of the relativeeffectiveness of the TCI/OSAVI (or REP in one case) and theCVI indices is reversed between the two LAI ranges, with noclear prevalence of one of the indices for lower or higher LAIconditions.

The two indices strictly focusing on the red edge (i.e., REPand MTCI), not LAI normalized, show correlation levels appre-ciably lower than LAI-normalized TCI/OSAVI and CVI indicesfor both LAI ranges considered (Tables V and VI). In general,

Page 11: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

VINCINI et al.: EMPIRICAL ESTIMATION OF LEAF CHLOROPHYLL DENSITY IN WINTER WHEAT CANOPIES 11

TABLE VICOEFFICIENT OF DETERMINATION (R2, TOP ROW) AND RMSE (IN MICROGRAMS PER SQUARE CENTIMETER, BOTTOM ROW) VALUES OF VI’S POWER

FITTINGS VERSUS LEAF CHLOROPHYLL (a+ b) CONTENT FOR DATA SYNTHETIZED BY DIFFERENT INPUT PARAMETERS: 60◦ (LEFT) AND 70◦ (RIGHT)ALAS, DIFFERENT SOIL TYPES AND SOIL MOISTURE CONDITIONS [DRY, WET, OR INTERMEDIATE (INT.)], 15◦ (TOP) OR 45◦ (BOTTOM) SUN ZENITH

ANGLES, AND N LEAF STRUCTURAL PARAMETER VALUES OF 1 OR 2; LAI RANGE OF 2.2–4.0—MAXIMUM R2 VALUE IN BOLD

Fig. 4. Average and standard deviation values of the rmse (in micrograms persquare centimeter) of power fittings between the best leaf chlorophyll densityestimators (VIs) obtained at Sentinel-2 spectral resolution and leaf chlorophyll(a+ b) content for the data synthetized from different combinations of inputparameters (n = 80; Tables V and VI) and for different LAI ranges: LAI of1.4–2.4 (top, Table V) and LAI of 2.2–4.0 (bottom, Table VI).

for the lower LAI range (Table V), the REP index (0.843 <R2 < 0.950 and 2.14 < rmse < 3.78 μg · cm−2) tends to bea better chlorophyll estimator than MTCI (0.820 < R2 <0.940 and 2.33 < rmse < 4.05 μg · cm−2), whereas for thehigher LAI range (Table VI), MTCI (0.830 < R2 < 0.937 and2.37 < rmse < 3.90 μg · cm−2) generally shows stronger cor-relations than REP (0.815 < R2 < 0.913 and 2.79 < rmse <4.07 μg · cm−2).

Fig. 4 shows the average and standard deviation values of thermse (in micrograms per square centimeter) of power fittingsbetween the four VIs, obtained at Sentinel-2 spectral resolution,and leaf chlorophyll density for data synthetized from the 80different combinations of input parameters and for the twoconsidered LAI ranges (Tables V and VI).

Fig. 4 clearly shows the higher effectiveness and moreconsistent behavior in different conditions of the two LAI-normalized indices (TCI/OSAVI and CVI) as leaf chlorophyllestimators compared to the two indices focusing on the red edge(REP and MTCI). Considering the whole variability of the twosynthetic data sets, for both LAI ranges, the TCI/OSAVI is thebest leaf chlorophyll density estimator (average rmse of 1.63 ±0.86 and 1.42± 0.63 μg · cm−2, respectively, for the lowerand higher LAI ranges), closely followed by the CVI index(average rmse of 1.88 ± 0.81 and 1.69± 0.90 μg · cm−2). TheMTCI (average rmse of 3.26 ± 0.38 and 3.12± 0.40 μg · cm−2,respectively, for the lower and higher LAI ranges) and REP(average rmse of 2.94 ± 0.37 and 3.31± 0.60 μg · cm−2)indices, which were the best chlorophyll estimators in the fieldtrials, seem to be relatively less effective when applied to asynthetic data set where LAI and chlorophyll variability isintrinsically uncorrelated. As mentioned, due to their commondependence on nitrogen availability, LAI and leaf chlorophyll

Page 12: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 5. Sensitivity functions (s) of the best leaf chlorophyll density estimators (VIs) obtained at Sentinel-2 spectral resolution plotted against leaf chlorophyll(a+ b) density (in micrograms per square centimeter) for experimental data collected on winter wheat at Feekes growth stages 5–6 aggregated (top-left) and forsynthetic data obtained for some soil conditions and for different LAI ranges: LAI of 1.4–2.4 (middle) and LAI of 2.2–4.0 (top-right, bottom).

density are, in general, directly related in field nitrogenfertilization trials, and some evidences indicate that they arecorrelated in the experimental data set of the present study as

well. The different performances as leaf chlorophyll estimatorsof LAI-normalized indices (TCI/OSAVI and CVI) comparedto the two indices strictly focusing on the red edge (REP

Page 13: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

VINCINI et al.: EMPIRICAL ESTIMATION OF LEAF CHLOROPHYLL DENSITY IN WINTER WHEAT CANOPIES 13

and MTCI) in field trials and in the comparison based onsynthetic data can be explained by the intrinsic differencesof the two data sets. In field trials, LAI and leaf chlorophylldensity are correlated, whereas in synthetic data, they are not.The higher correlations of the indices focusing on the rededge, which are not completely insensitive to LAI [31], [32],are probably due to the LAI–chlorophyll correlation in theexperimental data and do not represent in themselves an in-dication of a higher specific sensitivity to leaf chlorophylldensity. Literature indications of a nonnegligible rmse ofMTCI chlorophyll estimates at early growth stages, relatedto the optical properties of bare soils [31], are confirmed byresults of the analysis of synthetic data. Such results alsoseem to indicate a relatively poor accuracy of the REP indexobtained by linear interpolation of reflectances in Sentinel-2spectral bands.

Similar to the field trials, it is worth noting that the perfor-mances of CVI as an estimator of leaf chlorophyll density areconsistent with different soil/genotype/illumination conditions(Tables V and VI; Fig. 4). The values of the CVI R2 reportedin Tables V and VI varied from 0.802 to 0.999, with anaverage rmse value of 1.88 ± 0.81 and 1.69± 0.90 μg · cm−2,respectively, for the two considered LAI ranges (Fig. 4). Theseresults indicate that the CVI index, obtainable at 10-m spatialresolution from future Sentinel-2 data, could be used as a leafchlorophyll estimator in winter wheat crop canopies at growthstages suitable for N fertilizer topdressings, for most genotypesand soil conditions.

The sensitivity functions of the four VIs, obtained atSentinel-2 spectral resolution, are calculated by using (3) frompower regressions (2) with VI and leaf chlorophyll densityconsidered, respectively, as dependent and independent vari-ables. In Fig. 5, the sensitivity functions (s) are reported forfield experimental data collected on winter wheat at Feekesgrowth stages 5–6 aggregated (top-left) and for synthetic dataobtained for some combinations of LAI range, soil, genotype(ALA of 60◦ and N = 1), and illumination (sun zenith of 15◦)conditions. As expected by considering the strength of theircorrelations with leaf chlorophyll density, for synthetic data, thesensitivity of LAI-normalized TCI/OSAVI and CVI indices isgenerally higher than that of red-edge REP and MTCI indicesover the whole chlorophyll density range (Fig. 5). Respectivelyfor brighter/dry and darker/wet soils, the TCI/OSAVI (Othellodry soil; Fig. 5, middle-left) and the CVI index (Barnes soil;Fig. 5, top-right) are the best leaf chlorophyll estimators overthe whole chlorophyll density range, with the former decreasingits sensitivity for higher chlorophyll densities and the latterincreasing it.

For comparable correlation levels, the CVI tends to bemore sensitive for larger values of leaf chlorophyll content,which is more realistic for the crop nutritional status, thanthe TCI/OSAVI. For instance, the two VIs have similar R2

values, for both LAI ranges, for Cordorus wet soil, sunzenith angle of 15◦, ALA of 60◦, and leaf structure N = 1(TCI/OSAVI R2 = 0.987 and CVI R2 = 0.978—Table V;TCI/OSAVI R2 = 0.981 and CVI R2 = 0.971—Table VI).However, the sensitivity functions (Fig. 5, middle-right andbottom-right, respectively, for the lower and higher LAI ranges)

Fig. 6. Scatterplots and power fittings (y = a · xb + c) of the narrow-band TCI/OSAVI (top) and broad-band CVI (bottom) indices, obtained atSentinel-2 spectral resolution, versus chlorophyll (a+ b) density (in micro-grams per square centimeter) for Cordorus wet soil, sun zenith angle of 15◦,ALA of 60◦, N = 1, and LAI of 2.2–4.0.

indicate that, for such conditions, TCI/OSAVI is a more ef-fective estimator of leaf chlorophyll in the 20–30-μg · cm−2

range (i.e., severe chlorotic conditions), whereas CVI becomesprogressively more sensitive in the 35–50-μg · cm−2 range.Fig. 6 shows the scatterplots and power fittings of the two LAI-normalized indices for the model input combination from whichthe sensitivity functions reported in Fig. 5 bottom-right wereobtained (LAI of 2.2–4.0, Cordorus wet soil, sun zenith angle of15◦, ALA of 60◦, and leaf structure N = 1). As visible in Fig. 6,the relative sensitivity of the TCI/OSAVI and CVI indicesover the chlorophyll density range is mainly due to the changesof the slope of their regression lines, respectively, decreasing orincreasing for higher chlorophyll densities.

The same higher sensitivity for higher leaf chlorophyll den-sity is shown by the sensitivity functions of MTCI comparedto the REP, thus confirming the literature indication that theMTCI index is strongly correlated with the REP but more sen-sitive to the change in chlorophyll content at high chlorophyllcontents [56].

Comparing the shapes of the VI sensitivity functions reportedin Fig. 5 between field trial experimental data (top-left) andsynthetic data (all the other graphs in Fig. 5), a different

Page 14: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

general behavior of the sensitivity of the REP and TCI/OSAVIindices can be highlighted. While MTCI and CVI show aconsistent behavior of the sensitivity functions in field andsynthetic data (i.e., increasing more or less markedly over thechlorophyll density range), the REP and TCI/OSAVI indicesdo not. Rather than decreasing their sensitivities for higherchlorophyll densities as they do for synthetic data, the REPand TCI/OSAVI sensitivity functions tend to slightly increaseover the chlorophyll density range in field data (Fig. 5, top-left).Such different behavior of the sensitivity functions of REP andTCI/OSAVI indices for field data could possibly be due to eitherthe limited reproducibility of SPAD measurements affecting theestimate of the slopes of regression functions (Fig. 2, bottom) orto the mentioned LAI–chlorophyll correlation in the field trialexperimental data set.

On the other hand, field and synthetic data agree in showingsubstantially similar shapes of the sensitivity functions overthe chlorophyll density range for the REP and TCI/OSAVI(relatively more sensitive for lower densities) and for the MTCIand CVI indices (relatively more sensitive for higher densities).The sensitivity functions of REP and MTCI and of TCI/OSAVIand CVI tend to intersect around mid-chlorophyll density val-ues both for field and synthetic data (Fig. 5). As for fieldtrial experimental data, also synthetic data seem to indicatethe existence of a functional correlation as leaf chlorophyllestimators between REP and TCI/OSAVI and between MTCIand CVI even though both VI pairs include one red-edge indexfocusing on the red-edge spectral region (i.e., REP or MTCI)and one LAI-normalized index (i.e., TCI/OSAVI and CVI)considering green and red visible bands in their algorithm.

IV. CONCLUSION

The results of a field spectrometric experiment on winterwheat canopies with different nitrogen fertilization rates ad-dressing the sensitivity to leaf chlorophyll density of severalindices, based on visible–NIR or red-edge bands and obtainablefrom future Sentinel-2 spectral bands, indicate the following.

1) Two indices exclusively based on red-edge bands, theMTCI and REP indices, closely followed by the LAI-normalized TCI/OSAVI index, also based on visiblebands, were the best leaf chlorophyll estimators amongindices requiring Sentinel-2 red-edge bands.

2) As a chlorophyll estimator, CVI outperformed the otherVI requiring only Sentinel-2 visible and NIR bands andshowed a correlation level comparable to those of thebest-performing VI requiring red-edge bands.

The high sensitivity (0.64 < R2 < 0.71 and 2.3 < rmse <2.6 μg · cm−2) of the four best leaf chlorophyll estimatorsis obtained in the field spectrometric experiment on winterwheat canopies at growth stages suitable for N fertilization(i.e., Feekes growth stages 5–6). In particular, the highSentinel-2 spectral resolution in the visible spectral range ap-pears to markedly improve the effectiveness of the red/greenratio as a LAI normalization factor in the CVI algorithm.

To test the consistency as leaf chlorophyll estimators indifferent conditions, the MTCI, REP, TCI/OSAVI, and CVIindices were obtained at Sentinel-2 spectral resolution from a

large synthetic data set simulating winter wheat canopies atgrowth stages suitable for N fertilization with different soilbackground, winter wheat genotypes, and illumination con-ditions. The results of the analysis of the synthetic data setusing Sentinel-2 spectral resolution indicate that, compared tothe two indices focusing on the red edge (REP and MTCI),the two LAI-normalized (TCI/OSAVI and CVI) indices arebetter leaf chlorophyll estimators. The TCI/OSAVI index, usingreflectance both in the visible and red-edge spectral regions,in general, obtains the strongest correlations for brighter/drysoils, whereas the CVI index, based exclusively on visible andNIR bands, has the best correlations for darker/wet soils. Soilspectral properties appear to determine which index betweenthe TCI/OSAVI and CVI is the best leaf chlorophyll densityestimator regardless of sun zenith and winter wheat genotypevariability (ALA and N leaf structural parameter values).

On the basis of both field trials and synthetic data analysis,it can be concluded that the TCI/OSAVI index, requiring bandsin the red edge and obtainable at 20-m spatial resolution fromfuture Sentinel-2 data, and the CVI index, relying only onvisible bands and obtainable at 10-m spatial resolution, canbe used as leaf chlorophyll estimators in winter wheat cropcanopies at growth stages suitable for N fertilizer topdressings,for most genotypes and soil conditions. On the other hand,results of the analysis with synthetic data seem to confirm anappreciable residual sensitivity to LAI of MTCI [31] and REP,as obtained by S2 bands, affecting the accuracy of their leafchlorophyll density estimates at early growth stages (i.e., thephenology window for in-season N fertilization).

A further comparison of the four VIs based on the use of asensitivity function, describing the changes in their sensitivityover the range of leaf chlorophyll density, seems to indicate arelatively higher sensitivity of the REP and TCI/OSAVI indicesfor lower leaf chlorophyll densities and of the MTCI and CVIindices for higher densities. For soil conditions with compara-ble overall chlorophyll sensitivity of the TCI/OSAVI and CVIindices, the former seems to be a more effective estimatorof leaf chlorophyll in the 20–35-μg · cm−2 range (i.e., severechlorotic conditions), whereas the latter becomes progressivelymore sensitive in the 35–50-μg · cm−2 range.

REFERENCES

[1] M. S. Moran, Y. Inoue, and E. M. Barnes, “Opportunities and limitationsfor image-based remote sensing in precision crop management,” RemoteSens. Environ., vol. 61, no. 3, pp. 319–346, Sep. 1997.

[2] G. Le Maire, M. Jeuffroy, and F. Gastal, “Diagnosis tool for plantand crop N status in vegetative stage. Theory and practices forcrop N management,” Eur. J. Agron., vol. 28, no. 4, pp. 614–624,May 2008.

[3] S. M. Samborski, N. Tremblay, and E. Fallon, “Strategies to make useof plant sensors-based diagnostic information for nitrogen recommenda-tions,” Agron. J., vol. 101, no. 4, pp. 800–816, Jul. 2009.

[4] T. M. Blackmer, J. S. Schepers, and G. E. Varvel, “Light reflectancecompared with other nitrogen stress measurements in corn leaves,” Agron.J., vol. 86, no. 6, pp. 934–938, Nov. 1994.

[5] T. M. Blackmer, J. S. Schepers, G. E. Varvel, and E. A. Walter-Shea,“Nitrogen deficiency detection using reflected shortwave radiation fromirrigated corn canopies,” Agron. J., vol. 88, no. 1, pp. 1–5, Jan. 1996.

[6] C. S. T. Daughtry, C. L. Walthall, M. S. Kim, E. B. De Colstoun, andJ. E. McMurtrey, III, “Estimating corn leaf chlorophyll concentrationfrom leaf and canopy reflectance,” Remote Sens. Environ., vol. 74, no. 2,pp. 229–239, Nov. 2000.

Page 15: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

VINCINI et al.: EMPIRICAL ESTIMATION OF LEAF CHLOROPHYLL DENSITY IN WINTER WHEAT CANOPIES 15

[7] L. Murdock, S. Jones, C. Bowley, P. Needham, J. James, and P. Howe,“Using a Chlorophyll Meter to Make Nitrogen Recommendations onWheat,” Univ. of Kentucky Coop. Extension Service, Lexington, KY,USA, Bull. AGR-170, 1997.

[8] D. G. Bullock and D. S. Anderson, “Evaluation of the Minolta SPAD-502 chlorophyll meter for nitrogen management in corn,” J. Plant Nutr.,vol. 21, no. 4, pp. 741–755, Apr. 1998.

[9] E. Varvel, W. W. Wilhelm, J. F. Shanahan, and J. S. Schepers, “Analgorithm for corn nitrogen recommendations using a chlorophyll me-ter based sufficiency index,” Agron. J., vol. 99, no. 3, pp. 701–706,May 2007.

[10] F. Solari, J. F. Shanahan, R. B. Ferguson, and V. I. Adamchuk, “Anactive sensor algorithm for corn nitrogen recommendations based on achlorophyll meter algorithm,” Agron. J., vol. 102, no. 4, pp. 1090–1098,Jul. 2010.

[11] J. P. Schmidt, A. E. Dellinger, and D. B. Beegle, “Nitrogen rec-ommendations for corn: An on-the-go sensor compared with currentrecommendation methods,” Agron. J., vol. 101, no. 4, pp. 916–924,Jul. 2009.

[12] F. Baret, V. Houles, and M. Gue, “Quantification of plant stress usingremote sensing observations and crop models: The case of nitrogen man-agement,” J. Exp. Bot., vol. 58, no. 4, pp. 869–880, Feb. 2007.

[13] J. U. H. Eitel, D. S. Long, P. E. Gessler, and E. R. Hunt, Jr., “Combinedspectral index to improve ground-based estimates of nitrogen status indryland wheat,” Agron. J., vol. 100, no. 6, pp. 1694–1702, Nov. 2008.

[14] I. B. Strachan, E. Pattey, and J. B. Boisvert, “Impact of nitrogen and en-vironmental conditions on corn as detected by hyperspectral reflectance,”Remote Sens. Environ., vol. 80, no. 2, pp. 213–224, May 2002.

[15] F. Baret and G. Guyot, “Potentials and limits of vegetation indices for LAIand APAR assessment,” Remote Sens. Environ., vol. 35, no. 2/3, pp. 161–173, Feb./Mar. 1991.

[16] N. Gobron, B. Pinty, and M. M. Verstraete, “Theoretical limits to theestimation of the leaf area index on the basis of optical remote sensingdata,” IEEE Trans. Geosci. Remote Sens., vol. 35, no. 6, pp. 1438–1445,Nov. 1997.

[17] D. Haboudane, J. R. Miller, N. Tremblay, P. J. Zarco-Tejada, andL. Dextraze, “Integrated narrow-band vegetation indices for prediction ofcrop chlorophyll content for application to precision agriculture,” RemoteSens. Environ., vol. 81, no. 2/3, pp. 416–426, Aug. 2002.

[18] S. Jacquemoud, “Inversion of the PROSPECT+SAIL canopy reflectancemodel from AVIRIS equivalent spectra: Theoretical study,” Remote Sens.Environ., vol. 44, no. 2/3, pp. 281–292, May/Jun. 1993.

[19] S. Jacquemoud, F. Baret, B. Andrieu, F. M. Danson, and K. Jaggard,“Extraction of vegetation biophysical parameters by inversion of thePROSPECT+SAIL models on sugar beet canopy reflectance data. Appli-cation to TM and AVIRIS sensors,” Remote Sens. Environ., vol. 52, no. 3,pp. 163–172, Jun. 1995.

[20] S. Jacquemoud, C. Bacour, H. Poilve, and J. P. Frangi, “Comparisonof four radiative transfer models to simulate plant canopies reflectance:Direct and inverse mode,” Remote Sens. Environ., vol. 74, no. 3, pp. 471–481, Dec. 2000.

[21] B. Combal, F. Baret, M. Weiss, A. Trubuil, D. Mace, A. Pragnere,R. Myneni, Y. Knyazikhin, and L. Wang, “Retrieval of canopy biophysicalvariables from bi-directional reflectance data. Using prior information tosolve the ill-posed inverse problem,” Remote Sens. Environ., vol. 84, no. 1,pp. 1–15, Jan. 2003.

[22] C. Bacour, S. Jacquemoud, M. Leroy, O. Hautecoeur, M. Weiss, L. Prévot,L. Bruguier, and H. Chauki, “Reliability of the estimation of vegetationcharacteristics by inversion of three canopy reflectance models on air-borne POLDER data,” Agron., Agr. Environ., vol. 22, no. 6, pp. 555–565,Sep./Oct. 2002.

[23] C. Bacour, F. Baret, D. Béal, M. Weiss, and K. Pavageau, “Neural networkestimation of LAI, fAPAR, fCover and LAI × Cab, from top of canopyMERIS reflectance data: Principles and validation,” Remote Sens. Envi-ron., vol. 105, no. 4, pp. 313–325, Dec. 2006.

[24] C. Atzberger, “Object-based retrieval of biophysical canopy variablesusing artificial neural nets and radiative transfer models,” Remote Sens.Environ., vol. 93, no. 1/2, pp. 53–67, Oct. 2004.

[25] J. W. Rouse, R. H. Haas, J. A. Schell, D. W. Deering, and J. C. Harlan,“Monitoring the Vernal Advancements and Retrogradation of NaturalVegetation,” NASA, Greenbelt, MD, USA, pp. 1–137, 1974, NASA/GSFC, Final Rep.

[26] A. R. Huete, “A soil vegetation adjusted index (SAVI),” Remote Sens.Environ., vol. 25, no. 3, pp. 295–309, Aug. 1988.

[27] J. L. Roujean and F. M. Breon, “Estimating PAR absorbed by vegetationfrom bidirectional reflectance measurements,” Remote Sens. Environ.,vol. 51, no. 3, pp. 375–384, Mar. 1995.

[28] G. Rondeaux, M. Steven, and F. Baret, “Optimization of soil-adjustedvegetation indices,” Remote Sens. Environ., vol. 55, no. 2, pp. 95–107,Feb. 1996.

[29] J. B. Solie, A. D. Monroe, W. R. Raun, and M. L. Stone, “Generalizedalgorithm for variable-rate nitrogen application in cereal grains,” Agron.J., vol. 104, no. 2, pp. 378–387, Mar. 2012.

[30] N. H. Broge and E. Leblanc, “Comparing prediction power and stabilityof broadband and hyperspectral vegetation indices for estimation of greenleaf area index and canopy chlorophyll density,” Remote Sens. Environ.,vol. 76, no. 2, pp. 156–172, May 2000.

[31] A. Bannari, K. S. Khurshid, K. Staenz, and J. W. Schwarz, “A compari-son of hyperspectral chlorophyll indices for wheat crop chlorophyll con-tent estimation using laboratory reflectance measurements,” IEEE Trans.Geosci. Remote Sens., vol. 45, no. 10, pp. 3063–3074, Oct. 2007.

[32] D. Haboudane, N. Tremblay, J. Miller, and P. Vigneault, “Remote esti-mation of crop chlorophyll content using spectral indices derived fromhyperspectral data,” IEEE Trans. Geosci. Remote Sens., vol. 46, no. 2,pp. 423–437, Feb. 2008.

[33] M. Vincini and E. Frazzi, “Comparing narrow and broad-band vegetationindices to estimate leaf chlorophyll content in planophile crop canopies,”Precis. Agr., vol. 12, no. 3, pp. 334–344, Jun. 2011.

[34] J. Verrelst, L. Alonso, G. Camps-Valls, J. Delegido, and J. Moreno,“Retrieval of vegetation biophysical parameters using Gaussian processtechniques,” IEEE Trans. Geosci. Remote Sens., vol. 50, pt. 2, no. 5,pp. 1832–1843, May 2012.

[35] J. Dash and P. J. Curran, “The MERIS Terrestrial Chlorophyll Index,” Int.J. Remote Sens., vol. 25, no. 23, pp. 5403–5413, Dec. 2004.

[36] D. N. H. Horler, M. Dockray, and J. Barber, “The red-edge of plant leafreflectance,” Int. J. Remote Sens., vol. 4, no. 2, pp. 273–288, Jan. 1983.

[37] F. Baret, S. Jacquemoud, G. Guyot, and C. Leprieur, “Modeled analysis ofthe biophysical nature of spectral shifts and comparison with informationcontent of broad bands,” Remote Sens. Environ., vol. 41, no. 2/3, pp. 133–142, Aug./Sep. 1992.

[38] A. A. Gitelson, J. Y. Kaufman, and M. N. Merzlyak, “Use of a green chan-nel in remote sensing of global vegetation from EOS-MODIS,” RemoteSens. Environ., vol. 58, no. 3, pp. 289–298, Dec. 1996.

[39] A. A. Gitelson, A. Vina, V. Ciganda, D. C. Rundquist, and T. J. Arkebauer,“Remote estimation of canopy chlorophyll content in crops,” Geophys.Res. Lett., vol. 32, no. 8, pp. L08403-1–L08403-4, Apr. 2005.

[40] M. Vincini, E. Frazzi, and P. D’Alessio, “A broad-band leaf chlorophyllvegetation index at the canopy scale,” Precis. Agr., vol. 9, no. 5, pp. 303–319, Oct. 2008.

[41] ESA Sentinel-2 Team. (2010). GMES Sentinel-2 Mission RequirementsDocument, Paris, France, (accessed on 11 October 2012). [Online]. Avail-able: http://esamultimedia.esa.int/docs/GMES/Sentinel-2_MRD.pdf

[42] J. Peñuelas, J. A. Gamon, A. L. Fredeen, J. Merino, and C. B. Field,“Reflectance indices associated with physiological changes in nitrogenand water-limited sunflower leaves,” Remote Sens. Environ., vol. 48, no. 2,pp. 135–146, May 1994.

[43] G. Le Maire, C. Francois, and E. Dufrene, “Towards universal broad leafchlorophyll indices using PROSPECT simulated database and hyperspec-tral reflectance measurements,” Remote Sens. Environ., vol. 89, no. 1,pp. 1–28, Jan. 2004.

[44] E. C. Large, “Growth stages in cereals,” Plant Pathol., vol. 3, no. 4,pp. 128–129, Dec. 1954.

[45] C. F. Jordan, “Derivation of leaf area index from quality of light on theforest floor,” Ecology, vol. 50, no. 4, pp. 663–666, Jul. 1969.

[46] J. R. Thomas and H. W. Gausman, “Leaf reflectance versus leaf chloro-phyll and carotenoid concentrations for eight crops,” Agron. J., vol. 69,no. 5, pp. 799–802, Sep./Oct. 1977.

[47] J. S. Schepers, T. M. Blackmer, W. W. Wilhelm, and M. Resende, “Trans-mittance and reflectance measurements of corn leaves from plants withdifferent nitrogen and water supply,” J. Plant Physiol., vol. 148, no. 5,pp. 523–529, Nov. 1996.

[48] A. A. Gitelson and M. N. Merzlyak, “Signature analysis of leaf reflectancespectra: Algorithm development for remote sensing of chlorophyll,”J. Plant Physiol., vol. 148, no. 3/4, pp. 494–500, 1996.

[49] E. R. Hunt, Jr., C. S. T. Daughtry, J. U. H. Eitel, and D. S. Long, “Remotesensing leaf chlorophyll content using a visible band index,” Agron. J.,vol. 103, no. 4, pp. 1090–1099, Jul. 2011.

[50] D. Haboudane, J. R. Miller, E. Pattey, P. J. Zarco-Tejada, P. J. Strachan,and I. Strachan, “Hyperspectral vegetation indices and novel algorithmsfor predicting green LAI of crop canopies: Modeling and validation in thecontext of precision agriculture,” Remote Sens. Environ., vol. 90, no. 3,pp. 337–352, Apr. 2004.

[51] M. S. Kim, C. S. T. Daughtry, E. W. Chappelle, J. E. McMurtrey, III, andC. L. Walthall, “The use of high spectral resolution bands for estimating

Page 16: Empirical Estimation of Leaf Chlorophyll Density in Winter Wheat Canopies Using Sentinel-2 Spectral Resolution

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

16 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

absorbed photosynthetically active radiation (APAR),” in Proc. 6th Symp.Phys. Meas. Sign. Remote Sens., Val D’Isere, France, Jan. 17–21, 1994,pp. 299–306.

[52] G. Guyot and F. Baret, “Utilisation de la haute resolution spectrale poursuivre l’état des couverts vegetaux (use of the high spectral resolutionfor monitoring the status of vegetation covers),” in Proc. 4th Int. Colloq.Spectr. Sign. Obj. Remote Sens., T. D. Guyenne and J. J. Hunt, Eds.,Aussois, France, Jan. 18–22, 1988, pp. 279–286, Paris: ESA SpecialPublication SP-287.

[53] J. G. P. W. Clevers, “Imaging spectrometry in agriculture, plant vitalityand yield indicators,” in Imaging Spectrometry—A Tool for EnvironmentalObservations, J. Hill and J. Megier, Eds. Dordrecht, The Netherlands:Kluwer, 1994, pp. 193–219.

[54] M. A. Cho and A. K. Skidmore, “A new technique for extracting the rededge position from hyperspectral data: The linear extrapolation method,”Remote Sens. Environ., vol. 101, no. 2, pp. 181–193, Mar. 2006.

[55] M. Vincini, S. Amaducci, and E. Frazzi, “Sensitivity of Sentinel-2 red-edge bands to leaf chlorophyll concentration in winter wheat,” in Proc.1st ESA-ESRIN Sentinel-2 Preparatory Symp., Frascati, Italy, Apr. 23–27,2012, pp. 1–15, ESA Special Publication SP-707, Jul. 2012.

[56] J. Dash and P. J. Curran, Algorithm Theoretical Basis Document ATBD2.22 Chlorophyll Index, ESA, Paris, France, (accessed on 11 October2012). [Online]. Available: http://envisat.esa.int/instruments/meris/atbd/atbd_2.22.pdf

[57] J. Delegido, L. Alonso, G. González, and J. Moreno, “Estimating chloro-phyll content of crops from hyperspectral data using a normalized areaover reflectance curve (NAOC),” Int. J. Appl. Earth Observ. Geoinf.,vol. 12, no. 3, pp. 165–174, Jun. 2010.

[58] S. Jacquemoud, W. Verhoef, F. Baret, C. Bacour, P. J. Zarco-Tejada,G. P. Asner, C. François, and S. L. Ustin, “PROSPECT + SAIL models:A review of use for vegetation characterization,” Remote Sens. Environ.,vol. 113, no. S1, pp. S56–S66, Sep. 2009.

[59] C. S. T. Daughtry, J. E. McMurtrey, III, M. S. Kim, and E. W. Chappelle,“Estimating crop residue cover by blue fluorescence imaging,” RemoteSens. Environ., vol. 60, no. 1, pp. 14–21, Apr. 1997.

[60] C. Bacour, S. Jacquemoud, Y. Tourbier, M. Dechambre, and J.-P. Frangi,“Design and analysis of numerical experiments to compare four canopyreflectance models,” Remote Sens. Environ., vol. 79, no. 1, pp. 72–83,Jan. 2002.

[61] P. Ceccato, S. Flasse, S. Tarantola, S. Jacquemoud, and J. M. Gregoire,“Detecting vegetation leaf water content using reflectance in the opticaldomain,” Remote Sens. Environ., vol. 77, no. 1, pp. 22–33, Jul. 2001.

[62] G. J. Newnham and T. Burt, “Validation of a leaf reflectance and transmit-tance model for three agricultural crop species,” in Proc. IEEE IGARSS,Sydney, Australia, Jul. 9–13, 2001, pp. 2976–2978.

[63] S. Jacquemoud, S. Flasse, J. Verdebout, and G. Schmuck, “Comparison ofseveral optimization methods to extract canopy biophysical parameters,”in Proc. 6th CNES Int. Symp. Phys. Meas. Sign. Remote Sens., Val d’Isère,France, Jan. 17–21, 1994, pp. 291–298.

[64] L. Ji and A. J. Peters, “Performance evaluation of spectral vegetationindices using a statistical sensitivity function,” Remote Sens. Environ.,vol. 106, no. 1, pp. 59–65, Jan. 2007.

[65] G. Walburg, M. E. Bauer, C. S. T. Daughtry, and T. L. Housley, “Effectsof nitrogen nutrition on the growth, yield and reflectance characteristicsof corn canopies,” Agron. J., vol. 74, no. 4, pp. 677–683, Jul. 1982.

Massimo Vincini (M’09) was born in Milan, Italy,in 1962. He received the M.S. degree in agriculturalsciences from the Università Cattolica del SacroCuore (UCSC), Piacenza, Italy, in 1988.

From 1988 to 1999, he was with the UCSC Radio-chemistry Center, Faculty of Agricultural Sciences.He has been a Technical Manager with the UCSCResearch Centre of Spatial Analysis and RemoteSensing (CRAST) since 1999. From 2002 to 2005,he was a Lecturer in agricultural systems with theUniversità degli Studi di Pavia, Pavia, Italy. He has

been a member of the scientific board of the European Journal of RemoteSensing (formerly Italian Journal of Remote Sensing) since 2007. His researchinterests include the application of optical remote sensing to agriculture andforestry. In this field, he has been the Principal Investigator of Italian SpaceAgency (ASI) projects and has coordinated the scientific activities related tothe acquisitions over some Italian research sites of earth observation missions(Shuttle Radar Topography Mission and ESA hyperspectral CHRIS/Probaobservations).

Stefano Amaducci was born in Ponte dell’Olio,Italy, in 1968. He received the M.S. degree in agri-cultural sciences and the Ph.D. degree in field cropsfrom the Università degli Studi di Bologna, Bologna,Italy, in 1994 and 1998, respectively.

From 1998 to 2000, he held a research positionwith the Wageningen University and Research Cen-tre, Wageningen, The Netherlands. Since 2000, hehas been a Researcher with the Institute of Agron-omy, Genetics and Field Crops, Università Cattolicadel Sacro Cuore, Piacenza, Italy, where he teaches

courses on field crops, nonfood production chains, and bioenergy. His mainresearch interests focus on agronomic evaluation of industrial crops and onmanagement strategies to reduce GHG emission in agriculture.

Dr. Amaducci is member of the Italian Society of Agronomy and is currentlythe coordinator of the EC Project Multihemp.

Ermes Frazzi was born in Soragna, Italy, in 1949.He received the M.S. degree in agricultural sci-ences from the Università Cattolica del Sacro Cuore(UCSC), Piacenza, Italy, in 1973.

From 1978 to 1985, he was a Lecturer in ruralbuildings with UCSC, where he has been an As-sociate Professor of rural buildings and topographysince 1985. From 1986 to 2004, he was the Headof the Institute of Rural Engineering, UCSC Facultyof Agricultural Sciences. He is the current ScientificDirector of the UCSC Research Centre for Spatial

Analysis and Remote Sensing (CRAST). Among principal research fieldsin which he has been involved are rural land and agricultural environmentmanagement and remote sensing application to precision agriculture.