Richter 2011

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

458 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 4, NO.

2, JUNE 2011

Evaluation of Sentinel-2 Spectral Sampling for


Radiative Transfer Model Based LAI Estimation of
Wheat, Sugar Beet, and Maize
Katja Richter, Clement Atzberger, Francesco Vuolo, and Guido D’Urso

Abstract—The present study aimed at testing the potential of the for land surface characterization. This is of particular interest in
future E.O. mission Sentinel-2 (European GMES programme) for the context of precision farming, where information of crop and
the operational estimation of the Leaf Area Index (LAI) of three soil characteristics must be obtained at a fine scale, in a rapid
contrasting agricultural crops (wheat, sugar beet, and maize). Re-
trieval of LAI was achieved by using a look-up table (LUT) based and cost-effective way, and with a stable and high accuracy.
inversion of a radiative transfer model (SAILH+PROSPECT). In the framework of the Global Monitoring for Environment
Analyses were mainly carried out using hyperspectral data ac- and Security (GMES) initiative, the European Space Agency
quired by the optical airborne instrument CASI, simulating the (ESA) initiated the Sentinel-2 superspectral mission [1], aiming
future Sentinel-2 band setting. Estimated LAI was evaluated
at replacing and improving the current generation of satellite
using measurements of effective Plant Area Index (PAIe )
collected during the ESA AgriSAR 2006 campaign. Additionally, sensors. The mission intends to provide continuity to services
measurements from two other experiments were tested to enrich relying on multi-spectral high-resolution optical observations
the validation database. The GMES targeted precision of 10% for over global terrestrial surfaces, such as the adequate quantifica-
green LAI estimation was met for sugar beet (8%), at the limit tion of geo-biophysical variables. The future system Sentinel-2
for wheat (11%) but not for maize (19%). For the three crops
the RMSE was in the range 0.4–0.6. The results demonstrate the
is scheduled to be launched in the year 2013. Spectral sampling
importance of using crop specific radiative transfer models. For is based on sensors that have been used for vegetation moni-
row crops with incomplete coverage and strong leaf clumping, toring in the last decades, such as SPOT, Landsat or MODIS.
such as maize at early stage, the standard SAILH+PROSPECT With a spatial resolution of 10–60 m, Sentinel-2 is designed to
model does not appear suitable. However, results must be taken address high to medium resolution applications. A detailed tech-
cautiously in view of possible uncertainties of the PAIe measure-
ments. Within a future Sentinel-2 product validation framework, nical characterization of the mission can be found in [1], [2].
a standard protocol for reference measurements is required to Together with a number of other vegetation surface variables,
assure consistency of validation data sets. the implementation of Leaf Area Index (LAI) as Level 2b main
product is considered in the context of the mission. To ensure
Index Terms—Sentinel-2, leaf area index, PROSAILH, look-up
table, model inversion, radiative transfer model. that the final product can meet the user requirements, a goal
accuracy of 10% for the “maps with the green leaf area per unit
soil area”, i.e., LAI, was targeted [1].
I. INTRODUCTION Intense work has been done in the last decades to improve
the traditional empirical methods for LAI estimation from E.O.
data based on broad-band vegetation indices (VI). This included
HE current and future Earth Observation (E.O.) database
T offers the opportunity to exploit the potential of remote
sensing in a fully operational context. It also provides the possi-
the development of hyperspectral VIs [3], [4], spectral unmixing
approaches [5], red-edge position analysis [6], or angle indices
[7]. However, the need for collecting in situ calibration data sets
bility to test the performance of existing and new methodologies implies high costs and labor intensive measurement programs to
cover a wide range of species, canopy conditions and view/sun
Manuscript received April 08, 2009; revised February 02, 2010, June 10, constellations.
2010; accepted November 03, 2010. Date of publication December 17, 2010; For these reasons, many studies focused on the more complex
date of current version May 20, 2011. This work was supported in part by
PLEIADeS project (Participatory multi-Level EO-assisted tools for Irrigation approach of physically based variable estimation by means of
water management and Agricultural Decision-Support; Contract 037095 radiative transfer model (RTM) inversion (e.g., [8]–[11]). These
financed by U.E.-VI F.P.). RTMs permit to use the full spectrum acquired by multi- to hy-
K. Richter is with the Department of Geography, Faculty of Geosciences,
Ludwig-Maximilians Universitaet Munich, 80333 Munich, Germany (e-mail: perspectral sensors in contrast to VIs that generally use only
katja.rich@gmail.com). 2–3 spectral bands. In addition, RTMs can also exploit the di-
C. Atzberger is with University of Natural Resources and Life Sciences, Vi- rectional signature of multi-angle sensors. Nevertheless, some
enna Institute of Surveying, Remote Sensing and Land Information, 1190 Vi-
enna, Austria (e-mail: clement.atzberger@boku.ac.at). shortcomings of these models, such as the need of an extensive
F. Vuolo is with School of Geography, University of Southampton, SO171BJ, parameterization, as well as the high computational demand,
UK (e-mail: F.Vuolo@soton.ac.uk). have to be considered. Moreover, some RTMs may be too sim-
G. D’Urso is with DIAAT, Faculty of Agraria, University of Naples “Federico
II”, 80055 Portici (Na), Italy (e-mail: durso@unina.it). plistic to cope with complex canopies such as row crops, which
Digital Object Identifier 10.1109/JSTARS.2010.2091492 are often affected by foliage clumping [12], [13]. Furthermore,
1939-1404/$26.00 © 2010 IEEE
RICHTER et al.: EVALUATION OF SENTINEL-2 SPECTRAL SAMPLING 459

the ill-posed problem has to be taken into account when per- these three crops using the widely applied SAILH+PROSPECT
forming model inversion: different variable combinations may canopy reflectance model, originally developed for homo-
produce almost identical spectra, resulting in significant uncer- geneous canopies. The retrieval accuracy is assessed against
tainties in the estimated vegetation characteristics [14], [15]. standard LAI-2000 measurements without correction for leaf
To retrieve canopy biophysical variables from radiative clumping or any of the before mentioned problems. For RTM
transfer models, three inversion methods are commonly used inversion, the fast and simple LUT approach has been selected
[8], [14]: iterative optimization techniques [16], look-up tables since it presents an objective repeatable approach that has
(LUT), e.g., [10], [11], [15], [17], [18], and neural networks demonstrated potential in operational applications, e.g., for
(NN) [14], [19]. Recently, the new approach of Support Vector MODIS LAI products [26].
Machines Regression (SVR) has been applied to estimate bio-
physical variables from E.O. imagery, e.g., [20]. Several studies
II. MATERIAL AND METHODS
found that LUT and NN were performing best in the inversion
of the RTMs in terms of accuracy and speed [8], [11], [21]. A. Campaign and Study Area
To validate canopy characteristics retrieved from E.O. data,
different devices and sampling schemes have been proposed in The present research was mainly done in the context of the
the last decades. A prominent device for LAI measurements is ESA AgriSAR 2006 campaign [27], designed and performed on
for example the LAI-2000 Plant Canopy Analyzer (LI-COR, the consolidated long-term test site DEMMIN, Mecklenburg-
Lincoln, Nebraska USA) instrument. Measurements with such Vorpommern, Germany. The campaign focused on the Goermin
devices presume an accuracy of the in situ measurements (i.e., farm (with central coordinates N 53 59 24 and E 13 16 10 )
‘actual LAI’ [22]) substantially higher than of the E.O. based es- and included a range of ground measurements, airborne and
timates. This is often difficult as the characteristics of a (pixel) satellite acquisitions. Principal crops of the area comprise winter
area need to be measured to be comparable with remote sensing wheat, sugar beet, winter barley, winter rape, and maize, grown
data. In addition, the instruments themselves rely on some as- on very large parcels (in average 80 ha). Details of the campaign
sumptions in order to allow fast and easy data acquisitions com- and data acquisitions can be found in [27].
pared to destructive sampling. With the LAI-2000 device, for
instance, if no corrections were applied to account for stems or B. Imagery: Acquisition and Processing
senescent leaves, the measured variable corresponds to ‘Plant Hyperspectral images were acquired with the Compact
Area Index’ (PAI), e.g., [23]. Moreover, due to the assump- Airborne Spectrographic Imager (CASI 1500, ITRES Research
tion of a random distribution of foliage, clumping is only par- Ltd., 2006) on June 6 and July 5, 2006. CASI provides a
tially regarded by the instrument and corresponding software nearly-continuum spectrum over its entire spectral range, with
[24], giving therefore effective LAI as output rather very fine observation channels [28]. Therefore, the sensor offers
than actual LAI (or PAI, respectively). For these reasons, vari- a good opportunity to mimic the future multi-spectral system
ables measured by the LAI-2000 instrument will also be called Sentinel-2.
‘ ’ throughout the present study [22], [23]. The area of interest was scanned around 10:00 UTC, corre-
The influence of clumping is especially pronounced for het- sponding to a solar zenith angle of 35 . The sensor (fly al-
erogeneous crops (e.g., maize and sugar beet). The measured titude 3100 m asl) acquired hyperspectral data at 1.5 m spatial
values of these crops will be therefore more exposed to bias than resolution in 288 bands between 370 and 1050 nm, with a band-
homogeneous crops such as wheat. width (FWHMs) of 2.2 nm and a field of view (FOV) equal to
Other error sources related to the use of optical devices can 23.6 . Spectral calibration and atmospheric correction of the im-
affect the measurements, including illumination conditions, agery were carried out by the Laboratory for Earth Observation
saturation effects, or instrument’s simplifications [22]. Recent of the University of Valencia, by using an optimized version of
studies propose therefore the use [22], [23] or the combination the MODTRAN4 code [28]. The specific sensitivity functions
[25] of Digital Hemispherical Photographs (DHPs) with the of Sentinel-2 wavebands were applied to configure the CASI
LAI-2000, which may help to correct for these issues and measurements. To imitate the planned spatial resolution of Sen-
thus providing more reliable estimates of than from tinel-2 (4 channels with 10 m and 4 with 20 m, [1]), the high
LAI-2000 (alone). spatial resolution CASI data (1.5 m) were degraded to a coarser
For assessing clumping and thus actual LAI, it is also impor- resolution of 20 m.
tant that all samples are estimated in a reliable and con- To enlarge the present validation data base, remotely sensed
sistent way [24]. However, inconsistency may occur between data from two other experiments were consulted: 1) hyperspec-
different estimates especially for multiple samples in het- tral field measurements using an ASD FieldSpec UV-VNIR
erogeneous canopies, as highlighted by a recent study [24]. A field spectrometer acquired during the PLEIADeS 2007 field
quantification of according to [24] is therefore desirable campaign in Sardinia, Italy [29], and 2) imagery of the Com-
before using the data for further analyses. pact High Resolution Imaging Spectrometer (CHRIS) on the
The main objective of the present work is the experimental Proba platform collected during SPARC 2004 campaign, in
assessment of the future Sentinel-2 band setting for RTM based Barrax, Spain [30]. The remotely sensed data were configured
LAI retrieval of wheat, sugar beet, and maize. We evaluated if according to the Sentinel-2 spectral bands. The experiments are
the targeted retrieval accuracy of 10% [1] can be reached for described in detail in [29], [30]. As none of the three sensors
460 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 4, NO. 2, JUNE 2011

covered the SWIR region only the visible and near infrared TABLE I
bands of Sentinel-2 could be analyzed. DESCRIPTION OF CROPS OF THE AGRISAR CAMPAIGN (CORRESPONDING TO
TIME OF SENSOR OVERPASS) AND FROM ADDITIONAL DATASET OF PLEIADES
AND SPARC CAMPAIGNS, MEAN VALUES (STANDARD DEVIATIONS, SD)
C. In Situ Measurements
Leaf area index ( , respectively) measurements used in
this study were carried out—simultaneously to the sensor over-
passes—on three fields at the DEMMIN site: wheat in June
2006 (ID 250: 15 samples), sugar beet (ID 102: 8 samples),
and maize (ID 222: 16 samples) in July 2006. Measurements
were performed with the LAI-2000 Plant Canopy Analyzer in-
strument, widely and routinely applied in different studies, e.g.,
[10], [31]–[33].
Each of the 39 AgriSAR in situ values corresponded
to the average of 24 measurements (i.e., 3 repetitions of 1 above
and 8 below canopy readings). This scheme represented an El-
ementary Surface Unit (ESU) of approximately 20 m 20 m.
Measurements of the rather homogeneous wheat canopy were algorithms where the canopy is described in a more complex
taken randomly within the ESU. In case of maize and sugar way (for a review of these models see [12]), i.e., 3-D hybrid
beet, however, the row structure had to be taken into account. radiative transfer [43] or Monte Carlo ray tracing models [44].
Therefore, the sampling was composed of an alternation of be- Such models may be more accurate, but require larger parame-
tween- and within-rows measurements to capture the spatial het- terization efforts [8], [12]. In particular, geometric information
erogeneity of the canopies. The standard deviation around the about the specific canopy architecture of the different crop types
mean has been kept as a measure of uncertainty. Field measured is needed, rendering such models less suitable for operational
values ranged from 3.8 to 5.8 for wheat, from 1.0 to 2.0 applications in large heterogeneous areas.
for sugar beet, and from 0.9 to 2.3 for maize.
To consider a wider range of canopy densities, 21 measure- E. LUT-Based Inversion Procedure
ments of maize from the PLEIADeS 2007 field campaign and The PROSAILH model was inverted by using a look-up table
8 measurements of maize and 6 of sugar beet from the SPARC (LUT). Three principal steps have to be performed to realize this
2004 campaign were included. In both experiments, measure- approach:
ments were performed by using the same protocol as described First, an appropriate number of canopy realizations must be
above for the AgriSAR campaign. generated. For this, a LUT size of 100000 cases of canopy vari-
Table I gives an overview about the monitored crops. As the able combinations was chosen according to the results of [11].
present study uses from different campaigns, only lim- The bounds and distributions of the selected canopy variables
ited information about raw data was available. This made it are depicted in Table II. Variable bounds were taken from mea-
impossible to carry out further elaboration such as uncertainty surement campaigns and/or other studies working with the same
analyses as the comparability of the final estimates would be crops. They were chosen in order to describe the characteris-
reduced. Generally it can be expected that for wheat LAI tics of all crop types used in the study. Gaussian distributions
actual LAI, but for maize and sugar beet LAI actual LAI. were generated for LAI and chlorophyll content in order
Hence, there is the possibility that some of these measurements to put more emphasis on the variable values being present in
are within an error range of above 10%, which would be crit- the actual growth stages of the crops [45]. Instead, the average
ical in view of the study objective. However, LAI products de- leaf angle (ALA), the hot spot factor (HotS), a soil reflectance
rived from satellite or airborne sensor data correspond better to factor , dry matter content and leaf mesophyll pa-
the measured by LAI-2000 (or other optical instruments) rameter were sampled from a uniform distribution since no
than actual LAI [34]. information from the campaigns measurements was available.
As the absorption of leaf water is not influencing the spectral
D. Radiative Transfer Model range used in this study ( nm), leaf water thickness
The Scattering from Arbitrarily Inclined Leaves (SAILH) was fixed to an arbitrary value (Table II).
model [35], [36], later extended to take into account the The soil reflectance spectrum, another essential input into the
“hot-spot” effect [37], was exploited in this study. The SAILH model, was extracted from the CASI imagery (mean of several
model is based on the turbid medium assumption and describes bare soil pixels). A simple multiplicative soil reflectance factor
the canopy structure in a fairly simple way. To account for was introduced, which was assumed to mimic soil water
variations in leaf structure and composition, the canopy re- content and surface roughness induced variations of reflectance
flectance model was combined with a leaf optical properties [46]. Within the distributions defined in Table II, a random
model (PROSPECT) [38] to ‘PROSAILH’ (e.g., [11], [39]). sampling scheme was applied to all variables. To demonstrate
The PROSAILH model, widely used and validated, produces the general applicability of the method and to render the results
realistic results of bidirectional reflectance spectra for different comparable, the same LUT size and variable settings were
crops as reported by several studies, e.g., [9], [16], [40]–[42]. implemented for PLEIADeS and SPARC campaigns data
For this reason, it was preferred over other radiative transfer [29], [30].
RICHTER et al.: EVALUATION OF SENTINEL-2 SPECTRAL SAMPLING 461

Fig. 1. PAI and estimated LAI values for (a) wheat, (b) sugar beet, (c) maize, and (d): all crops. For (a)–(c) point symbols ‘.’ represent AgriSAR data; stars ‘*’
SPARC data, and circles ‘o’ PLEIADeS data. For (d): circles ‘o’ represent wheat, stars ‘*’ sugar beet, and point ‘.’ maize. LAI estimates were obtained from LUT
inversion approach and using the Sentinel-2 band configuration. Standard deviations of measurements and simulations are indicated by error bars in Fig. (a)–(c).
Accuracy is described by root mean square error (RMSE) and relative RMSE (rel.% error).

TABLE II
RANGE AND DISTRIBUTION OF INPUT VARIABLES USED TO ESTABLISH THE calculating the root mean square error (RMSE) between mea-
SYNTHETIC CANOPY REFLECTANCE DATA BASE FOR USE IN THE LUT sured and simulated spectra [41] over all
wavelengths as e.g., [11], [17], [47].
Analyses demonstrated that the sole use of the variable com-
bination corresponding to the smallest RMSE decreases the esti-
mation accuracy [10]. Therefore, for the final solution, all vari-
able combinations within 10% of the lowest RMSE value (1)
were averaged, similar to [21] and [41]. From all possible solu-
tions (i.e., located within the 10% of lowest RMSE), the stan-
dard deviations were calculated and kept as a measure of re-
trieval uncertainty.
Including prior information about the crop status into the al-
gorithm has proven to give more accurate estimates, e.g., [15],
[18]. However, only little information is usually available about
an agricultural area. Thus, for the operational retrieval of bio-
physical variables from E.O. data, applied algorithms have to
fulfill either the requirement to be universally valid or to exploit
local/regional information of the vegetation monitored. The im-
plementation of such prior knowledge about crop’s phenolog-
ical status or plant species can be deduced for instance from
land cover mapping, as it is already done for medium/coarse
resolution sensor data in University of Toronto (UofT) GLOB-
In a second step, the model was applied in direct mode to sim- CARBON LAI algorithm [48], [49]. Assuming availability of
ulate reflectance signatures corresponding to the proposed spec- such information within the variable retrieval algorithms of the
tral sampling of the Sentinel-2 system. Table 3.10.1 in [1] gives future satellite mission, denser samplings around the most prob-
the specifications of the sensors, with central wavebands, spec- able and LAI values were applied (Table II) in the construc-
tral width, bandwidth, purpose and heritage of each channel. All tion of the LUT.
wavebands interesting for LAI or vegetation studies were in-
cluded in the simulation. Hence, the calculations comprised the III. RESULTS AND DISCUSSIONS
visible (490, 560, 665 nm), the red edge (705 and 740 nm) and
Scatter plots between and estimated LAI values for the
the near infrared (775, 842, and 865 nm) parts of the spectrum.
different crops are shown in Fig. 1 (wheat (a), sugar beet (b),
To take into account sensor and model uncertainties, a Gaussian
maize (c), and for all (d)), including error bars of standard
(white) noise was added to the simulated canopy reflectance.
deviation of measurements and retrievals.
The noise was assumed to be proportional to the reflectance
In absolute terms (RMSE), the comparison of and LAI
(mean of zero and standard deviation of 0.04) [19], [39]. The
estimation based on the spectral band configuration of the fu-
simulated reflectance spectra were resampled according to the
ture Sentinel-2 system performed best for sugar beet and maize
Sentinel-2 sensitivity.
(RMSE: 0.42). For wheat, on the contrary, only a RMSE of 0.64
In the final step the LUT was sorted along with a simple cost
could be reached, which has to be seen, however, with respect
function (1)
to the high LAI level of this crop. Consequently, in terms of
relative accuracies, i.e., RMSE divided by the mean of PAI
(1) (rel. RMSE), the best performing crops were sugar beet with
rel. RMSE of 8.5%, and wheat, with rel. RMSE of 10.9%. For
462 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 4, NO. 2, JUNE 2011

maize, only a rel. RMSE of 18.8% was obtained. Hence, in view Sentinel-2. The targeted GMES goal accuracy of % could
of the GMES target, the required accuracy was met for sugar be achieved for sugar beet 8% , marginal for wheat 11%
beet and marginal for wheat. In case of maize the estimation ac- but not for maize 19% . However, these accuracies have to
curacy failed to reach the targeted level of rel. RMSE % be interpreted carefully considering potential bias introduced by
despite an absolute RMSE of 0.4. However, it must be kept the measurements of the reference data set.
in mind that these results are only valid for the present sample The results of the study confirm that the 1-dimensional
data set. In some cases, measurement uncertainty of due SAILH+PROSPECT model may be sufficient for homoge-
to bias may also rise above 10%, as discussed in the introduc- neous structured canopies, even though saturation can occur
tion. for values higher than 4. For heterogeneous row crops
In general, the inversion results reveal an underestimation of instead, such as maize, a higher complexity of modeling may
high values (except maize): Relationships between mea- be required. Such kind of 3-dimensional radiative transfer mod-
sured and simulated variables have the following slope values: eling is for instance implemented in the MODIS-15 operative
0.45 for wheat, 0.96 for sugar beet, 1 for maize and 0.93 for all algorithm for broadleaf crops [26]. However, the MODIS-15
crops. Evaluation uncertainties may for example arise from the approach requires information of canopy architectural types,
different signals obtained and associated with LAI by the remote derived from classification of vegetation in different biomes
sensor ( [31]) and by the LAI-2000 instrument . and multi-angular observations, which are not always avail-
That implies that reference values could be too high, if non- able from high spatial resolution sensors needed for smaller
photosynthetic canopy components, such as wheat ears—which field-scale management purposes. Hence, for operational appli-
were present in this growth stage—reduced the measured gap cations, the PROSAILH model can be a reasonable compromise
fractions of the LAI-2000 instrument. between simple empirical VI-based methods and very complex
The observed underestimation for high values of RTMs for the retrieval of LAI in most crop systems, when
wheat and sugar beet may also result from the nonlinearity of using observations of limited spectral dimension and with one
the LAI-reflectance relationship, leading to saturation effects viewing angle only. The model is for instance applied in the
[19], [39]. The canopy architecture of maize and sugar beet CYCLOPES algorithm for globe wide biophysical products
deviates from the turbid medium assumption of the PROSAILH from VEGETATION [39]. Though, limited accuracy of the
model [40]: maize, but also sugar beet, is typically row-planted approach has been shown by a recent study [49].
and affected by leaf clumping (i.e., leaves are grouped together Conclusively, the coupled PROSPECT SAILH model with
rather than distributed uniformly). Hence, the signature origi- an implemented LUT algorithm can be a suitable tool to support
nating from the vegetation is underrepresented, leading to the the elaboration of Sentinel-2 mission data to ensure the deliv-
observed underestimation of actual LAI ( , respectively). ering of quality products. The simulated Sentinel-2 configura-
Similar findings were also described in other studies applying tion was very encouraging in terms of absolute errors (RMSE:
the same model, e.g., [49]–[51]. However, these uncertainties 0.4–0.6 LAI), but measurement uncertainties must be taken into
may not only arise due to the modeling approach but can be account. However, when accuracies of 10% are required for not
the result of inadequate geolocation or atmospheric correc- yet well established row crops, more complex RTMs should be
tions of the E.O. data. A possible explanation of the better applied. In order to render such approaches suitable for future
results achieved for sugar beet may also be related to the ad- operational applications, further validation exercises of existing
vanced growth stages, exhibiting already a more homogeneous complex 3-D models are proposed.
coverage (50–80%). Consequently, the influence of the soil Regarding a future operational system for the retrieval of bio-
background (which is further enhanced by the row structure physical variables from the Sentinel-2 sensors, the implementa-
of this crop) is probably stronger for maize fields compared tion of contextual [14], [53] and multi-temporal retrieval strate-
to sugar beet. In a future study it has to be tested if a simple gies [54], i.e., using prior knowledge of the crops, should be
mixture model, such as the one proposed by [39] at landscape tested. This could be a valuable alternative to very complex
level or GeoSail [50], [52] is capable of reducing this effect. models requiring high parameterization loads.
These considerations are also supported by [13], who proposed Finally, we strongly suggest the definition of a standard pro-
the use of a row structure model for the early growth stage cedure for the generation of in situ LAI (or , respectively)
(before elongation) of maize canopies, and a homogeneous one records, including suitable instruments (DHP, LAI-2000), ap-
for later growth stages (after elongation). propriate sampling schemes, and accompanying corrections as
In addition, we tested the case that the (average) prior guesses proposed by [24]. This should assure the generation of a consis-
( and ) present the estimated LAI. Relative tent and reliable reference database within a future Sentinel-2
RMSE values increased markedly for sugar beet (30%) some- products validation framework.
what for maize (20%) and little for wheat (13%). This proves
that the RTM retrieval is significantly more accurate than the a
priori guess. ACKNOWLEDGMENT

The authors thank L. Guanter and colleagues from the Lab-


IV. CONCLUSION
oratory for Earth Observation, University of Valencia (Spain),
Leaf area index was estimated for wheat, sugar beet, and for radiometric and spectral calibrations and atmospheric cor-
maize based on the spectral sampling of the future E.O. satellites rection of CASI and CHRIS/Proba imagery.
RICHTER et al.: EVALUATION OF SENTINEL-2 SPECTRAL SAMPLING 463

REFERENCES [19] C. Bacour, F. Baret, D. Béal, M. Weiss, and K. Pavageau, “Neural


network estimation of LAI, fAPAR, fCover and LAIxCab, from top
[1] GMES Sentinel-2 Mission Requirements Document, issue 2 revision of canopy MERIS reflectance data: Principles and validation,” Remote
0-30/01/2007, 31 pp. European Space Agency (ESA), 2007 [Online]. Sens. Environ., vol. 105, no. 4, pp. 313–325, 2006.
Available: http://esamultimedia.esa.int/docs/GMES/GMES_Sen- [20] G. Camp-Valls, J. Muñoz-Marí, L. Gómez-Chova, K. Richter, and J.
tinel2_MRD_issue_2.0_update.pdf Calpe-Maravilla, “Biophysical parameter estimation with a semisuper-
[2] GMES Sentinel-2 performance and products summary, issue 1 revision vised support vector machine,” IEEE Geosci. Remote Sens. Lett., vol.
1, 20/08/2008, 9 pp. European Space Agency (ESA), 2007. 6, pp. 248–252, Apr. 2009.
[3] R. Darvishzadeh, C. Atzberger, S. van Wieren, and A. Skidmore, “Es- [21] K. Richter, C. Atzberger, F. Vuolo, P. Weihs, and G. D’Urso, “Exper-
timation of vegetation LAI from hyperspectral reflectance data: Effects imental assessment of the Sentinel-2 band setting for RTM-based LAI
of soil type and plant architecture,” Int. J. Appl. Earth Obs., vol. 10, no. retrieval of sugar beet and maize,” Can. J. Remote Sens., vol. 35, no. 3,
3, pp. 358–373, 2008. pp. 230–247, 2009.
[4] D. Haboudane, J. R. Miller, E. Pattey, P. J. Zarco-Tejada, and I. B. [22] S. Garrigues, N. V. Shabanov, K. Swanson, J. T. Morissette, F. Baret,
Strachan, “Hyperspectral vegetation indexes and novel algorithms for and R. B. Myneni, “Intercomparison and sensitivity analysis of leaf
predicting green LAI of crop canopies: Modeling and validation in the area index retrievals from LAI-2000, AccuPAR, and digital hemispher-
context of precision agriculture,” Remote Sens. Environ., vol. 90, pp. ical photography over croplands,” Agr. Forest Meteorol., vol. 148, no.
337–352, 2004. 8–9, pp. 1193–1209, 2008.
[5] B. Hu, J. R. Miller, J. M. Chen, and A. B. Hollinger, “Retrieval of [23] S. G. Leblanc, J. M. Chen, R. Fernandes, D. W. Deering, and A.
the canopy leaf area index in the BOREAS flux tower sites using Conley, “Methodology comparison for canopy structure parameters
linear spectral mixture analysis,” Remote Sens. Environ., vol. 89, pp. extraction from digital hemispherical photography in boreal forests,”
176–188, 2004. Agr. Forest Meteorol., vol. 129, pp. 187–207, 2005.
[6] M. Cho, A. Skidmore, and C. Atzberger, “Towards red-edge positions [24] Y. Ryu, T. Nilson, H. Kobayashi, O. Sonnentag, and D. D. Baldocchi,
less sensitive to canopy biophysical parameters for leaf chlorophyll es- “On the correct estimation of effective leaf area index: Does it reveal
timations using properties optique spectrales des feuilles (PROSPECT) information on clumping effects?,” Agr. Forest Meteorol., vol. 150, pp.
and scattering by arbitrarily inclined leaves (SAILH) simulated data,” 463–472, 2010.
Int. J. Remote Sens., vol. 29, no. 7–8, pp. 2241–2255, 2008. [25] Y. Ryu, O. Sonnentag, T. Nilson, R. Vargas, H. Kobayashi, R. Wenk,
[7] S. Khanna, A. Palacios-Orueta, M. L. Whiting, D. Riaño, J. Litago, and and D. D. Baldocchi, “How to quantify tree leaf area index in a het-
S. L. Ustin, “Development of angle indexes for soil moisture estima- erogeneous savanna ecosystem: A multi-instrument and multi-model
tion, dry matter detection and land-cover discrimination,” Remote Sens. approach,” Agr. Forest Meteorol., vol. 150, pp. 63–76, 2010.
Environ., vol. 109, pp. 154–165, 2007. [26] R. B. Myneni, S. Hoffman, Y. Knyazikhin, J. L. Privette, J. Glassy,
[8] F. Baret and S. Buis, “Estimating canopy characteristics from Remote Y. Tian, Y. Wang, X. Song, Y. Zhang, G. R. Smith, A. Lotsch, M.
Sensing observations. review of methods and associated problems,” Friedl, J. T. Morisette, P. Votava, R. R. Nemani, and S. W. Running,
in Advances in Land Remote Sensing: System, Modeling, Inversion “Global products of vegetation leaf area and fraction absorbed PAR
and Application, S. Liang, Ed. Dordrecht, The Netherlands: Springer, from year one of MODIS data,” Remote Sens. Environ., vol. 83, no.
2008, pp. 172–301. 1–2, pp. 214–231, 2002.
[9] S. Jacquemoud, C. Bacour, H. Poilve, and J. P. Frangi, “Comparison [27] I. Hajnsek et al., “AgriSAR 2006: Airborne SAR and optics campaigns
of four radiative transfer models to simulate plant canopies reflectance: for an improved monitoring of agricultural processes and practices,” in
Direct and inverse mode,” Remote Sens. Environ., vol. 74, pp. 417–481, Proc. AGRISAR and EAGLE Campaigns Final Workshop, Noordwijk,
2000. The Netherlands, 2007, p. 8.
[10] R. Darvishzadeh, A. Skidmore, M. Schlerf, and C. Atzberger, “Inver- [28] L. Guanter, V. Estellés, and J. Moreno, “Spectral calibration and atmo-
sion of a radiative transfer model for estimating vegetation LAI and spheric correction of ultra-fine spectral and spatial resolution Remote
chlorophyll in a heterogeneous grassland,” Remote Sens. Environ., vol. Sensing data. Application to CASI-1500 data,” Remote Sens. Environ.,
112, no. 5, pp. 2592–2604, 2008. vol. 109, pp. 54–65, 2007.
[11] M. Weiss, F. Baret, R. B. Myneni, A. Pragnère, and Y. Knyazikhin, [29] K. Richter, F. Vuolo, and G. D’Urso, “Leaf area index and surface
“Investigation of a model inversion technique to estimate canopy albedo estimation: Comparative analysis from vegetation indexes to
biophysical variables from spectral and directional reflectance data,” radiative transfer models,” in Proc. IEEE IGARSS, Boston, MA, Jul.
Agronomie, vol. 20, pp. 3–22, 2000. 2008, vol. 3, pp. 736–739.
[12] W. A. Dorigo, R. Zurita-Milla, A. J. W. de Wit, J. Brazile, R. Singh, [30] K. Richter and W. Timmermans, “Physically based retrieval of crop
and M. E. Schaepman, “A review on reflective remote sensing and data characteristics for improved water use estimates,” Hydrol. Earth Syst.
assimilation techniques for enhanced agroecosystem modeling,” Int. J. Sci., vol. 13, pp. 663–674, 2009.
Appl. Earth Obs., vol. 9, no. 2, pp. 165–193, 2007. [31] M. C. González-Sanpedro, T. Le Toan, J. Moreno, L. Kergoat, and E.
[13] Y. Yao, Q. Liu, Q. Liu, and X. Li, “LAI retrieval and uncertainty evalu- Rubio, “Seasonal variations of leaf area index of agricultural fields re-
ations for typical row-planted crops at different growth stages,” Remote trieved from landsat data,” Remote Sens. Environ., vol. 112, no. 3, pp.
Sens. Environ., vol. 112, no. 1, pp. 94–106, 2008. 810–824, 2007.
[14] C. Atzberger, “Object-based retrieval of biophysical canopy variables [32] G. D’Urso, L. Dini, F. Vuolo, and L. Guanter, “Retrieval of vegeta-
using artificial neural nets and radiative transfer models,” Remote Sens. tion biophysical parameters by inverting hyperspectral, multiangular
Environ., vol. 93, pp. 53–67, 2004. CHRIS/PROBA data from SPARC 2003,” in Proc. 2nd CHRIS/Proba
[15] B. Combal, F. Baret, and M. Weiss, “Improving canopy variables esti- Workshop, Frascati, Italy, Apr. 2004.
mation from remote sensing data by exploiting ancillary information. [33] J. M. Chen, A. Govind, O. Sonnentag, Y. Q. Zhang, A. Barr, and B.
Case study on sugar beet canopies,” Agronomie, vol. 22, no. 2, pp. Amiro, “Leaf area index measurements at Fluxnet-Canada forest sites,”
205–215, 2002. Agr. Forest Meteorol., vol. 140, no. 1–4, pp. 257–268, 2006.
[16] S. Jacquemoud, F. Baret, B. Andrieu, F. M. Danson, and K. Jaggard, [34] P. Stenberg, M. Rautiainen, T. Manninen, P. Voipio, and H. Smolander,
“Extraction of vegetation biophysical parameters by inversion of the “Reduced simple ratio better than NDVI for estimating LAI in Finnish
PROSPECT+SAIL models on sugar beet canopy reflectance data. Ap- pine and spruce stands,” Silva Fennica, vol. 38, no. 1, pp. 3–14, 2004.
plication to TM and AVIRIS sensors,” Remote Sens. Environ., vol. 52, [35] W. Verhoef, “Light scattering by leaf layers with application to canopy
pp. 163–172, 1995. reflectance modeling: The SAIL model,” Remote Sens. Environ., vol.
[17] B. Combal, F. Baret, M. Weiss, A. Trubuil, D. Macé, and A. Prag- 16, pp. 125–141, 1984.
nére, “Retrieval of canopy biophysical variables from bidirectional [36] W. Verhoef, “Earth observation modeling based on layer scattering ma-
reflectance using prior information to solve the ill-posed inverse trices,” Remote Sens. Environ., vol. 17, pp. 165–178, 1985.
problem,” Remote Sens. Environ., vol. 84, pp. 1–15, 2003. [37] A. Kuusk, “The hot-spot effect in plant canopy reflectance. Photon-
[18] M. Yebra and E. Chuvieco, “Generation of a species-specific look-up vegetation interactions,” in Applications in Optical Remote Sensing and
table for fuel moisture content assessment,” IEEE J. Sel. Topics Earth Plant Ecology, R. B. Myneni and J. Ross, Eds. New York: Springer-
Obs. Remote Sens. (JSTARS), vol. 2, no. 1, pp. 21–26, 2009. Verlag, 1991, pp. 139–159.
464 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 4, NO. 2, JUNE 2011

[38] S. Jacquemoud and F. Baret, “PROSPECT: A model of leaf optical Katja Richter received the Ph.D. degree from the
properties spectra,” Remote Sens. Environ., vol. 34, pp. 75–91, 1990. University of Natural Resources and Life Sciences
[39] F. Baret, O. Hagolle, B. Geiger, P. Bicheron, B. Miras, M. Huc, B. Vienna, Austria, in 2009.
Berthelot, F. Nino, M. Weiss, O. Samain, J. L. Roujean, and M. Leroy, From 2007 to 2009 she was with the Department of
“LAI, fAPAR and fCover CYCLOPES global products derived from Agricultural Engineering and Agronomy, University
vegetation. Part 1: Principles of the algorithm,” Remote Sens. Environ., of Naples Federico II, Italy, within the PLEIADeS
vol. 110, no. 3, pp. 275–286, 2007. project, addressing the use of E.O. based technolo-
[40] B. Andrieu, F. Baret, S. Jacquemoud, T. Malthus, and M. Steven, gies for the determination of irrigation water require-
“Evaluation of an improved version of SAIL model for simulating ments. Currently, she is with the Department of Ge-
bidirectional reflectance of sugar beet canopies,” Remote Sens. Env- ography, Ludwig-Maximilians Universitaet Munich,
Germany, where she is working within the EnMAP
iron., vol. 60, no. 3, pp. 247–257, 1997.
Core Science Team (ECST) for agriculture. In this context she focuses on the
[41] B. Koetz, F. Baret, H. Poilvè, and J. Hill, “Use of coupled canopy
development of fast and efficient methods for the retrieval of biophysical vari-
structure dynamic and radiative transfer models to estimate biophysical
ables from the future German hyperspectral satellite mission.
canopy characteristics,” Remote Sens. Environ., vol. 95, pp. 115–124,
2005.
[42] S. Jacquemoud, W. Verhoef, F. Baret, P. J. Zarco-Tejada, G. Asner,
+
C. Francois, and S. L. Ustin, “PROSPECT SAIL: 15 Years of use for
Clement Atzberger received the Ph.D. degree on
land surface characterization,” in Proc. IEEE IGARSS, Denver, CO, Jul.
crop growth modeling and remote sensing data
2006, vol. 1, pp. 1992–1995.
assimilation from Trier University, Germany, in
[43] S. Duthoit, V. Demarez, J. P. Gastellu-Etchegorry, E. Martin, and J.
1997.
L. Roujean, “Assessing the effects of the clumping phenomenon on
He is currently full Professor and head of the Insti-
BRDF of a maize crop based on 3-D numerical scenes using DART tute for Surveying, Remote Sensing and Land Infor-
model,” Agr. Forest Meteorol., vol. 148, pp. 1341–1352, 2008. mation (IVFL) at the University of Natural Resources
[44] N. S. Goel and T. Grier, “Estimation of canopy parameters for inhomo- and Life Sciences (BOKU), Vienna, Austria. After
geneous vegetation canopies from reflectance data: III. Trim: A model two years as Assistant Professor at ITC, Enschede,
for radiative transfer in heterogeneous three-dimensional canopies,” The Netherlands, and further two years in private in-
Remote Sens. Environ., vol. 25, no. 3, pp. 255–293, 1988. dustry (GeoSys SA, Toulouse, France), from 2007 to
[45] H. Gerighausen, E. Borg, C. Wloczyk, B. Fichtelmann, A. Günther, H. 2010 he was with the Joint Research Centre (JRC) of the European Commission,
H. Vajen, M. Rosenberg, M. Schulz, and H. G. Engler, “DEMMIN—A Ispra, Italy. His main expertise is in the following fields: coarse resolution time
test site for the validation of remote sensing data products. General series analysis (including noise removal and retrieval of phenological markers),
description and application during AgriSAR 2006,” in Proc. AGRISAR radiative transfer modeling (forward and inverse) in agriculture and forestry,
and EAGLE Campaigns Final Workshop, Noordwijk, The Netherlands, crop growth modeling and data assimilation, and imaging spectroscopy.
2007, p. 9.
[46] C. Atzberger, T. Jarmer, M. Schlerf, B. Koetz, and W. Werner, “Re-
trieval of wheat bio-physical attributes from hyperspectral data and
SAILH+PROSPECT radiative transfer model,” in Proc. 3rd EARSeL Francesco Vuolo received the Ph.D. degree in man-
Workshop on Imaging Spectroscopy, Herrsching, Germany, May 2003. agement of agricultural and forestry resources from
[47] K. Richter, P. Rischbeck, J. Eitzinger, W. Schneider, F. Suppan, and P. the University of Naples Federico II, Italy.
Weihs, “Plant growth monitoring and potential drought risk assessment Since 2006 he has been remote sensing specialist
by means of earth observation data,” Int. J. Remote Sens., vol. 29, no. at the spin-off company Ariespace addressing the
17–18, pp. 4943–4960, 2008. use of multi-spectral and high-resolution E.O. data
[48] F. Deng, J. M. Chen, S. Plummer, M. Chen, and J. Pisek, “Algorithm for vegetation monitoring, evapotranspiration, and
for global leaf area index retrieval using satellite imagery,” IEEE Trans. irrigation water requirements estimation. Currently,
Geosci. Remote Sens., vol. 44, no. 8, pp. 2219–2229, 2006. he is with the School of Geography,e University
[49] F. Canisius, R. Fernandes, and J. Chen, “Comparison and evaluation of Southampton, U.K., where he is working on the
of medium resolution imaging spectrometer leaf area index products development of vegetation chlorophyll algorithms
across a range of land use,” Remote Sens. Environ., vol. 114, no. 5, pp. for the future ESA Sentinel-2 and Sentinel-3 missions. He has been involved
950–960, 2010. in field campaigns for the validation of land-surface algorithms at fine and
medium resolution spatial scales.
[50] D. J. Major, G. B. Schaalje, C. Wiegand, and B. L. Blad, “Accuracy
and sensitivity analyses of sail model-predicted reflectance of maize,”
Remote Sens. Environ., vol. 41, no. 1, pp. 61–70, 1992.
[51] B. Koetz, M. Kneubühler, S. Huber, J. Schopfer, and F. Baret, “LAI
estimation based on multi-temporal CHRIS/PROBA data and radiative Guido D’Urso is full Professor of agricultural hy-
transfer modelling,” in Proc. ENVISAT Symp., Noordwijk, The Nether- draulics, irrigation management and remote sensing
lands, 2007, Montreux 2007, ESA Publications Division. at the University of Naples Federico II, Italy. He
[52] K. F. Huemmrich, “The GeoSail model: A simple addition to the SAIL received the Ph.D. degree from the Agricultural
University of Wageningen, The Netherlands. He has
model to describe discontinuous canopy reflectance,” Remote Sens. En-
coordinated several international research projects in
viron., vol. 75, no. 3, pp. 423–431, 2001.
the field of remote sensing for water management,
[53] C. Atzberger and K. Richter, “Geostatistical regularization of inverse as documented by more than 80 publications on
models for the retrieval of vegetation biophysical variables,” in Proc scientific journal, congress proceedings and special-
SPIE 2009 Conf., 2009, p. 747810. ized books. His research activities are focused on
[54] C. Lauvernet, F. Baret, L. Hascoët, S. Buis, and F. X. Le Dimet, “Multi- three main areas: development of Earth Observation
temporal-patch ensemble inversion of coupled surface-atmosphere ra- interpretation techniques for water management and land surface processes;
diative transfer models for land surface characterization,” Remote Sens. distributed agro-hydrological models for water management and irrigation; and
Environ., vol. 112, pp. 851–861, 2008. in-situ and remote active microwave sensing of soil water content.

You might also like