Developments for vegetation fluorescence retrieval from spaceborne high-resolution spectrometry in the O2-A and O2-B absorption bands

Authors


Abstract

[1] Solar-induced chlorophyll fluorescence is a weak electromagnetic signal emitted in the red and far-red spectral regions by vegetation chlorophyll under excitation by solar radiation. Chlorophyll fluorescence has been demonstrated to be a close proxy to vegetation physiological functioning. The basis for fluorescence retrieval from passive space measurements is the exploitation of the O2-A and O2-B atmospheric absorption features to isolate the fluorescence signal from the solar radiation reflected by the surface and the atmosphere. High spectral resolution measurements and a precise modeling of the atmospheric radiative transfer in the visible and near-infrared regions are mandatory. Recent developments for fluorescence retrieval from passive high spectral resolution spaceborne measurements are presented in this work, which has been performed in preparation of the FLuorescence EXplorer (FLEX) mission, which is currently under development by the European Space Agency. A large data set of FLEX-like measurements has been simulated for the purpose of methodology development and testing. Issues related to vegetation chlorophyll fluorescence retrieval from space, a description of the proposed methodology, initial results from simulated test cases, and general guidelines for the specification of fluorescence retrieval instruments are presented and discussed in this work.

1. Introduction

[2] Solar-induced chlorophyll fluorescence (Fs) is emitted by vegetation chlorophyll a under excitation by solar radiation. The fluorescence emission occurs in two broad peaks in the red (685 nm) and far-red (740 nm) regions of the spectrum. A number of laboratory and field experiments have demonstrated that chlorophyll fluorescence is directly linked to the instantaneous plant photosynthesis [e.g., Papageorgiou and Govindjee, 2004; Baker, 2008], opposite to traditional reflectance-based vegetation parameters which are only indicators of the potential photosynthetic activity of the plant.

[3] The low intensity of the fluorescence signal with respect to the solar radiation reflected by vegetation in the same spectral region makes the estimation of fluorescence from remote measurements a challenging problem. Entcheva Campbell et al. [2008] estimated steady state Fs to be between 1.5 and 3.4 mW m−2 sr−1 nm−1 at 685 nm and between 2.4 and 5.4 mW m−2 sr−1 nm−1 at 740 nm for different plant species. Those numbers represent a fraction of the radiation reflected by vegetation between 8.7% and 21.9% at 685 nm and between 2.0% and 5.2 % at 740 nm. Similar numbers were reported by other studies [Corp et al., 2006; Amorós-López et al., 2008; Zarco-Tejada et al., 2009]. Therefore, the main challenge in fluorescence retrieval from remote measurements is the isolation of the fluorescence signal from the radiance arriving at the sensor, which in the red and far-red regions is mainly due to the solar radiation reflected by the surface under clear atmosphere conditions.

[4] Measurements of backscattered sunlight in atmospheric absorption features overlapping the fluorescence emission can serve to this purpose. The different atmospheric optical paths crossed by the fluorescence and reflected signals provide the information required to decouple these two components from the radiance measured at the top-of-atmosphere (TOA) level. The atmospheric O2-A and O2-B absorption features in 760.5 nm and 687.5 nm, respectively, can be used for fluorescence retrieval, as they provide a good sampling of the two peaks of the fluorescence emission, and O2 is a well-mixed gas in the atmosphere. Fraunhofer lines in the solar spectrum were also proposed for fluorescence retrieval in the past [Sioris et al., 2003], but are now discarded because of their low spectral overlap with the most intense emission in the chlorophyll a fluorescence spectrum; water vapor features are also discarded because of the high variability of water vapor in both the temporal and the spatial dimension. The suitability of the O2 absorption features for fluorescence retrieval is illustrated in Figure 1. A real top-of-canopy fluorescence spectrum is superposed to a TOA radiance spectrum simulated for a green vegetation target. The location of the O2-A and O2-B absorption features is depicted.

Figure 1.

Real top-of-canopy (TOC) chlorophyll fluorescence spectrum superposed to a top-of-atmosphere (TOA) radiance spectrum simulated from a green vegetation target. The O2-A and O2-B absorption features are marked with rectangles. Illumination and observation zenith angles in the simulation were 30° and 0°, respectively, and the spectral resolution 1 cm−1.

[5] As reported in Meroni et al. [2009], most of the methods in the literature for Fs detection from either ground-based, airborne, or spaceborne instruments are based on the simplistic Fraunhofer Line Depth (FLD) principle [Plascyk, 1975] applied to the O2-A absorption feature. The FLD method is a suite of 2 or 3 band differential absorption technique in which the in-filling of the absorption feature by fluorescence is used to decouple the fluorescence and reflectance signals. A nonfluorescent target, either a reference panel or a bare soil surface, is used for the normalization of the nonmodeled atmospheric effects [Moya et al., 2004; Pérez-Priego et al., 2005; Alonso et al., 2008]. However, FLD-like approaches cannot be applied for Fs estimation in the O2-B feature, which does not present deep and broad absorption lines to set the measuring and reference channels required by the FLD technique, and can hardly be applied to spaceborne fluorescence retrieval, as reference targets (bare soil targets surrounding green vegetation surfaces) are seldom available from the satellite scale [Guanter et al., 2007].

[6] Spectral fitting methods have been proposed as an alternative to FLD-like methods for the retrieval of Fs from O2-A and O2-B measurements [Meroni and Colombo, 2006; Meroni et al., 2010; Mazzoni et al., 2008]. These methods exploit high spectral resolution measurements (considered to be of the order of 0.1 nm in this work) for the decoupling of reflectance and fluorescence by means of multichannel regressions. The higher amount of information contained in high-resolution data enables the better normalization of atmospheric effects, the characterization of the instrument spectral calibration, and to account for nonlinear spectral variations of the background vegetation reflectance, which are especially important in O2-B.

[7] Vegetation chlorophyll fluorescence retrieval from high-resolution spectrometry in the O2-A and O2-B atmospheric absorption features is the main objective of the FLuorescence EXplorer (FLEX) mission concept under development by the European Space Agency (ESA) [Drusch and FLEX Team, 2008]. FLEX, whose core instrument is the Fluorescence Imaging Spectrometer (FIMAS), is presently under evaluation for implementation as an in-orbit technology demonstrator which would fly in tandem with the Sentinel-3 system [Drinkwater and Rebhan, 2007] of the Global Monitoring for Environment and Security (GMES) program. The exploitation of Sentinel-3 Ocean Land Color Instrument (OLCI) and Sea and Land Surface Temperature Radiometer (SLSTR) data for the support of FIMAS measurements in cloud screening and the characterization of atmospheric conditions is considered as baseline for the latest FLEX concept. Sentinel-3 OLCI and SLSTR will continue ENVISAT's MEdium Resolution Imaging Spectrometer (MERIS) and Advanced Along-Track Scanning Radiometer (AATSR) multispectral and multiangular measurements. If finally selected for implementation and according to the most recent instrument concept, FIMAS would measure with a high signal-to-noise ratio (SNR) in two spectral windows of width between 20 and 30 nm centered at the O2-A and O2-B absorption features, with spectral sampling interval (SSI) and full width at half maximum (FWHM) between 0.1 and 0.3 nm and ground sampling distance (GSD) between 300 m and 1 km. FIMAS observations would be in quasi-nadir mode and would cover an at-ground area of about 100 km. A summary of some of the most important FIMAS requirements is presented in Table 1.

Table 1. Key Specifications of the FLEX/FIMAS Instrument as of December 2009a
 O2-AO2-B
  • a

    “O” and “T” labels refer to optimum and threshold values, respectively. SNR is given at the mission reference radiance level. Pol. Sensit. stands for sensitivity to polarization.

Spectral window[745, 775] nm (O)[672, 702] nm (O)
[750, 770] nm (T)[677, 697] nm (T)
SSI (∼FWHM)0.1 nm (O)–0.3 nm (T)0.1 nm (O)–0.3 nm (T)
SNR[300:1–1600:1] (O)[600:1–900:1] (O)
[150:1–800:1] (T)[200:1–400:1] (T)
Pol. Sensit.1% (O)–2% (T)1% (O)–2% (T)
GSD300 m (O)–1000 m (T)300 m (O)–1000 m (T)
Swath∼100 km∼100 km

[8] Developments for operational Fs retrieval from high-resolution measurements in O2-A and O2-B performed during the FLEX preparatory activities are presented hereinafter in this paper. In particular, an overview of a tentative Fs retrieval methodology implemented for FIMAS-like type of data is described, as well as the simulated data set used for algorithm development and validation purposes. First estimates of potential Fs retrieval error figures and guidelines for the definition of specifications for future spaceborne spectrometers for fluorescence retrieval are also presented and discussed.

2. FLEX/FIMAS Simulated Data Set

2.1. Atmospheric Forward Model

[9] Under the assumption that both surface reflectance and fluorescence emission are isotropic, the spectral TOA radiance signal LTOA can be formulated by a simple expression accounting for the interaction with the atmosphere of the radiation reflected and emitted by the surface,

equation image

where ρs is the surface reflectance; L0 is the atmospheric path radiance; μil is the cosine of the illumination zenith angle, measured between the solar ray and the surface normal; Edirμil and Edif are the direct and diffuse fluxes, respectively, arriving at the surface; S is the atmospheric spherical albedo, reflectance of the atmosphere for isotropic light entering it from the surface, and T is the total atmospheric transmittance (for diffuse plus direct radiation) in the observation direction. It must be remarked that directional effects in the vegetation reflectance and their potential interaction with the Fs signal and the atmospheric radiative transfer are neglected by this formulation. The atmospheric parameters required for the conversion from ρs and Fs to LTOA were generated in this work by the Matrix Operator MOdel (MOMO) [Fell and Fischer, 2001] and the MODTRAN4 [Berk et al., 2003] radiative transfer codes.

[10] As shown in this work, aerosol scattering and surface pressure are the two atmospheric parameters which may introduce the highest uncertainty in Fs retrieval under clear-sky conditions. Atmospheric aerosols can be characterized in passive remote sensing studies by the aerosol optical depth (AOD) at a given wavelength, the aerosol model, and the aerosol vertical profile. AOD is the column integral of the aerosol extinction coefficient, the aerosol model defines the spectral scattering and absorption properties of the aerosol mass and the aerosol vertical profile describes the vertical distribution of the aerosol extinction.

[11] Concerning surface pressure (SPR), it can be formulated as a major component given by the surface elevation plus a minor component dependent on the atmospheric fronts. The major component explains about 90% of the surface pressure at a given location and time, and can be derived from an accurate digital elevation model. The variable component, in turn, must be estimated according to the instantaneous atmospheric conditions [Lindstrot et al., 2009].

[12] AOD at 550 nm (AOD550), aerosol model, and SPR are considered free variables in the atmospheric forward model developed in this work for the simulation of FIMAS-like data. A look-up table containing sets of atmospheric optical parameters (atmospheric path radiance, at-ground direct and diffuse irradiance, spherical albedo, and atmospheric transmittance) as a function of AOD550, aerosol model and SPR has been compiled with MOMO and MODTRAN4 calculations. Breakpoints of this look-up table are AOD550 of {0.05, 0.2, 0.4}, combined with rural, maritime and urban models and SPR of {750, 940, 1030} hPa. It is acknowledged that such a look-up table structure is too sparse to enable accurate interpolation in the parameter space, but it is considered sufficient for the analysis presented in this work.

[13] The impact of AOD, aerosol model, aerosol height, and SPR on TOA radiance at O2-A and O2-B regions is compared with that of Fs in Figures 2 and 3, respectively. The normalized difference between the LTOA simulated for a given input configuration and the LTOA for the reference configuration (specified in the figure caption) is plotted as a function of wavelength. The aerosol, surface pressure and Fs input parameters are varied at a time with respect to the reference configuration.

Figure 2.

Sensitivity of the TOA radiance in the FIMAS O2-A spectral window to different atmospheric parameters. The reference configuration is for nadir observation, illumination zenith angle = 30°, midlatitude summer atmosphere, green vegetation reflectance spectrum, Fs = 0 mW m−2 sr−1 nm−1, surface pressure = 991 hPA, AOD550 = 0.2, rural aerosol model, aerosol layer between 2–9 km, spectral sampling interval = 0.1 nm.

Figure 3.

Same as Figure 2 but for the O2-B spectral window.

[14] A very different sensitivity of LTOA to the different parameters is observed in O2-A and O2-B. Overall, the impact of Fs on O2-A is relatively small with respect to the other parameters tested. It also presents a high spectral correlation with the other parameters, especially with surface pressure. The main spectral difference is on the left hand side of the O2-A feature, where variations in Fs transmit to LTOA and in the smaller spectral dependence of the sensitivity to Fs than to the other parameters. On the other hand, a significant correlation can be observed between surface pressure and the height of the aerosol layer and between the AOD and the aerosol model. In the second case, it can be argued that different combinations of AOD and aerosol model could lead to a similar impact on LTOA. Regarding the aerosol models, there seems to be a larger difference between the rural and the urban models than between the rural and the maritime models, despite only a small proportion of soot-like aerosol is added in the urban model to the water-soluble and dust-like components of the rural aerosol model.

[15] Concerning the O2-B window, the impact of Fs is spectrally uncorrelated to the other parameters due to particular spectral shape of the fluorescence emission in O2-B. Moreover, Fs has a stronger impact in O2-B than in O2-A due the much lower contribution of the reflected solar radiation to LTOA, which can be observed in Figure 1. The impact of surface pressure and the height of the aerosol layer is much smaller than in O2-A, whereas the variation of the AOD for the three aerosol models leads to a similar spectral pattern with a different intensity depending on the model. In view of these results, the better performance of Fs retrieval from O2-B than from O2-A can a priori be expected.

2.2. Atmospheric Parameters Not Considered in the Forward Model

[16] There are other potential error sources for Fs retrieval related to the atmosphere-surface radiative transfer formulation. These are not considered explicitly by the atmospheric forward model in this work, but either they are assumed to be provided by external sources or they have to be added to the Fs retrieval error budget.

2.2.1. Thin and Subpixel Clouds

[17] Nondetected thin and subpixel clouds would introduce large errors in the modeling of the O2-A and O2-B absorption features under the assumption of clear-sky conditions. However, given the spatial resolution of FLEX/FIMAS data (300–1000 m), it can be assumed that pure cloud-free pixels can be screened out from cloudy pixels even in the case of partially cloudy skies [Miller et al., 2007]. It is also assumed that the high spectral resolution of FIMAS data in O2-A and O2-B, alone and in synergy with Sentinel-3 OLCI and SLSTR measurements (which include channels in the visible, shortwave infrared and thermal infrared regions, and in particular two channels in 1375 and 1610 nm suited for cirrus cloud detection), are sufficient for accurate cloud screening including the optically thinnest cirrus clouds. No Fs retrieval would be attempted for pixels suspicious of being contaminated by clouds.

2.2.2. Coupling of Atmospheric Radiative Transfer and Directional Reflectance Effects

[18] Directional effects in vegetation reflectance may lead to the vegetation reflectance response to be different for the incoming diffuse and direct solar radiation fluxes. This effect is not considered by the Lambertian formulation in equation (1), and directional reflectance effects coupled to the incoming atmospheric fluxes would appear as perturbations in the O2-A and O2-B absorption features to be modeled. Such directional effects are expected in vertically heterogeneous vegetation covers and for canopies with a preferential leave orientation. No systematic analysis of the impact of directional reflectance on Fs retrieval is yet available in the literature, but the topic is considered for near-future research. It must be remarked that such formulation of directional reflectance effects is not considered in other retrieval approaches requiring a rigorous modeling of the O2-A absorption feature [e.g., Bösch et al., 2006].

2.2.3. Aerosol Vertical Profile

[19] The vertical location of the aerosol layer affects the balance between atmospheric scattering and absorption, as the incoming solar radiation being reflected by aerosols back to the TOA crosses a shorter atmospheric path when the aerosols are located at a higher altitude [Dubuisson et al., 2009]. The simulations presented in Figure 2 show that the uncertainty of the aerosol vertical profile may have a very large impact on Fs retrieval in O2-A. Since no information about the aerosol profile is going to be provided by Sentinel-3 instruments, this error source is so far to be included in the Fs retrieval error budget. This error will be quantified later in this work. The retrieval from FIMAS measurements of an effective surface pressure product to be used for subsequent Fs retrieval is considered a topic for further research. Such a pressure product would be biased by variations in the aerosol profile according to Figure 2, and could thus serve to partially compensate this uncertainty.

2.2.4. Temperature Vertical Profile

[20] The oxygen absorption bands are composed of individual absorption lines that are subject to pressure and temperature-dependent broadening processes. Even though the dominant process in the lower atmosphere is pressure broadening, an increase of temperature at constant pressure results in narrower, more intense absorption lines, and vice versa. A sensitivity analysis presented by Lindstrot et al. [2009] shows that shifting the temperature profile by 1K can result in a maximum error of 5 hPa in surface pressure retrieval from MERIS data in O2-A. This uncertainty would be smaller due to the compensation of errors if the sign of the uncertainty in the temperature profile changed with height, as expected in the real case. Time- and latitude-resolved climatology or near-real-time ancillary information of temperature profiles can be added to the Fs retrieval scheme to improve the retrieval accuracy.

2.2.5. Polarization of Radiation

[21] The spectral dependence of the degree of polarization of the solar radiation reflected by the atmosphere could lead to in-filling effects of the O2-A and O2-B absorption features [Natraj et al., 2007; Boesche et al., 2009], which might in turn cause errors in Fs retrieval. However, the demanding sensitivity to polarization requirement (1%–2%) in FIMAS is expected to minimize errors in FIMAS measurements caused by polarization.

2.2.6. Water Vapor

[22] Nonneglectable water vapor absorption lines overlap part of the O2-B absorption feature, roughly between 690–700 nm. However, the continuity in Sentinel-3 OLCI of the accurate columnar water vapor product provided by MERIS [Bennartz and Fischer, 2001] is assumed to be enough for the compensation of water vapor absorption within the O2-B feature.

2.2.7. Ring Effect and Dayglow Emission

[23] The filling-in of the O2-A and O2-B features by the Ring effect [Grainger and Ring, 1962] associated to rotational Raman scattering is expected to be small for nadir-looking observations with respect to the fluorescence signal in the red and far-red spectral regions [Sioris et al., 2003; Sioris and Evans, 2000]. The same is true for O2-A dayglow emissions in the upper atmosphere [Wallace and Hunten, 1968].

2.3. Input Vegetation Reflectance and Fs Emission Spectra

[24] The spectral shape and magnitude of vegetation reflectance in the O2-A and O2-B spectral regions may have a strong impact on Fs retrieval. In order to recreate the widest range of vegetation reflectance patterns, a number of top-of-canopy vegetation reflectance spectra has been used in the generation of the FIMAS-like database developed for this study. For this purpose, the FluorSAIL and FluorMODleaf codes [Pedrós et al., 2010], which simulate top-of-canopy and leaf radiative transfer, respectively, have been run under 20 combinations of chlorophyll content and leaf area index. This has led to a range of red-edge positions, reflectance levels and spectral shapes at O2-A and O2-B. In particular, values of chlorophyll a + b content of {25, 50, 75, 95} μg cm−2 and of leaf area index of {2, 3, 4, 5, 6} were combined in order to generate a spectral library of green vegetation reflectance spectra. The rest of the parameters driving the leaf and canopy models, such as the leaf internal structure parameter, the water equivalent thickness, the dry matter content or the leaf inclination distribution function were set to FluorSAIL and FluorMODleaf default values. Concerning fluorescence, a real sunflower Fs emission spectrum has been scaled to 5 different intensity levels (including 0-fluorescence) which cover the range between 0 and 4 mW m−2 sr−1 nm−1 at 760 nm. This intensity range is selected according to supporting laboratory and field-based studies referenced in the Introduction. Resulting reflectance and Fs spectra are displayed in Figure 4.

Figure 4.

Set of (top) reflectance and (bottom) fluorescence spectra used for FIMAS TOA radiance simulations. The O2-A and O2-B spectral windows covered by FIMAS are marked with rectangles.

2.4. FIMAS Simulations in the Spatial Domain

[25] A consistent performance analysis and error budget estimation can be performed when realistic spatial distributions of vegetation covers, atmospheric parameters and surface elevation are included in the assessment. According to the latest developments in the definition of the FLEX mission, a GSD of 300 m and a spatial swath of about 100 km can be assumed for FIMAS. For the simulation set-up, this spatial configuration can be reproduced from spatial subsets of ENVISAT MERIS full resolution images [Rast et al., 1999]. MERIS full resolution images present a swath of about 1150 km and a GSD of 300 m. In the normal operation mode, MERIS provides measurements in the 400–900 nm spectral range in 15 spectral channels with varying bandwidths ranging from 3.75 nm (at O2-A) to 20 nm, 10 nm being a typical value in most of the bands. Realistic spatial patterns and texture for FIMAS scenes simulation are retrieved from 400 × 400 pixel subsets extracted from 6 MERIS full resolution images. MERIS data were atmospherically corrected with the Self-Contained Atmospheric Parameters Estimation (SCAPE-M) atmospheric processor [Guanter et al., 2008], which generates reflectance and AOD550 maps from MERIS TOA radiance data. The six subsets used in this work are displayed in Figure 5. They were selected so that the maximum variability of spatial patterns and vegetation types was obtained.

Figure 5.

True color composites of the MERIS full-resolution subsets used to simulate FIMAS-like scenes. Each subset covers an approximate area of 120 km.

[26] The conversion from the MERIS multispectral type of data to the FIMAS high spectral resolution in the O2-A and O2-B spectral bands is achieved by means of spectral unmixing. A nonnegative least squares (NNLS) unmixing algorithm [Lawson and Hanson, 1974] was used together with the FluorSAIL spectral library in Figure 4 to calculate end-member abundances from the MERIS reflectance subsets. The resulting abundances were then combined with the spectral library resampled to the FIMAS spectral response to convert from the original MERIS data to FIMAS-like reflectance spectra.

[27] Real spatial patterns of AOD550 and SPR were also used in the simulation of the FIMAS-like TOA radiance scenes. AOD subsets were extracted from the AOD550 maps generated by SCAPE-M. SPR was calculated from the Global Earth Topography And Sea Surface Elevation at 30 arc sec resolution (GETASSE30) digital elevation model which was co-registered to each of the MERIS images. In addition, horizontal distributions of AOD550 and SPR uncertainties were simulated for the Fs retrieval step. These error maps were not correlated to the input maps, but were generated from independent AOD550 and SPR spatial patterns. An example of AOD550 and SPR error maps for Fs retrieval is displayed in Figure 6. These error maps are intended to recreate uncertainties in AOD and SPR as provided by Sentinel-3 measurements. Pairs of input and error maps for AOD550 and SPR were generated for each of the test sites simulated from MERIS data. Concerning the aerosol model, constant abundances of aerosol types were used for the entire simulated area. Input relative abundances of the rural, urban and maritime models in the simulations were, respectively, 50%, 20% and 30%, while the corresponding values assumed in the retrieval step were 70%, 0% and 30%.

Figure 6.

Example of a priori aerosol optical depth at 550 nm and surface pressure error maps used with the subset 1 for Fs retrieval in the end-to-end simulation process.

[28] Input Fs maps to be used in the simulation were derived by means of an empirical linear relationship between Fs and the Normalized Difference Vegetation Index (NDVI) [Tucker, 1979] calculated from the MERIS reflectance subsets. Maximum Fs was 4 mW m−2 sr−1 nm−1 in both O2-A and O2-B.

[29] FIMAS-like TOA radiance scenes were finally produced by applying equation (1) to each pixel, the per-pixel atmospheric parameters being generated through interpolation from the atmospheric look-up table.

2.5. FIMAS Instrument Model

[30] The spectral convolution of the resulting TOA radiance data to the FIMAS spectral response and the simulation of instrumental noise are the last step in the forward simulation process. Only a very basic FIMAS instrument model could be specified at this point. In order to cover all the possible FIMAS spectral configurations under consideration, the FIMAS SSI was varied between 0.1 and 0.3 nm, and the spectral windows between 20 and 30 nm (Table 1). A Gaussian spectral response function and a fixed ratio between FWHM and SSI of 1.2 was selected for all the simulations. Uncertainties in the knowledge of the instrument spectral response were also simulated as spectral shift and channel broadening. Errors in spectral wavelength position and width up to ±0.2 and ±0.1 spectral pixels, respectively, were simulated in the spectral convolution step.

[31] Instrumental noise, in turn, was simulated on TOA radiance as Gaussian noise in accordance with the FIMAS SNR specifications presented in Table 1. The dependence of signal-to-noise ratio with spectral bandwidth was also simulated by means of an empirical relationship derived from a preliminary instrument design.

3. Fs Retrieval Algorithm

3.1. Fs Retrieval From FIMAS-Like High-Resolution Measurements

[32] Approaches for Fs retrieval from spaceborne FIMAS type of data can be classified in terms of how they deal with the uncertainty in the atmospheric conditions. On the one hand, the atmospheric state can be characterized prior to Fs retrieval by means of either ancillary data sources or FIMAS measurements concurrent to Fs retrieval. On the other hand, a second type of approach would assume that FIMAS high-resolution measurements contain sufficient information for the consistent retrieval of aerosol parameters, surface pressure, surface reflectance, and Fs by means of a multiparameter retrieval scheme. The optimal estimation technique [Rodgers, 2000] developed for the retrieval of atmospheric profiles and trace gases could be adapted to this purpose.

[33] However, given the fact that FIMAS would fly in tandem with the Sentinel-3 OLCI and SLSTR instruments capable of providing information for cloud screening and for the accurate retrieval of aerosol parameters and surface pressure, only approaches of the first type focusing on fluorescence and reflectance decoupling have been considered in this work. Uncertainties in AOD and SPR retrieval and modeling errors associated to those parameters are considered additional error sources to be propagated along Fs retrieval.

3.2. Initial Data Processing for Fs Retrieval

[34] Assuming Sentinel-3 OLCI and SLSTR flying in tandem with FLEX/FIMAS, a series of processing steps must precede the actual Fs retrieval.

[35] 1. First is the cloud screening step. Pixels suspicious of being affected by any degree of cloud contamination will not be considered for Fs retrieval. The synergy between Sentinel-3 and FIMAS data would provide visible and infrared multispectral measurements (near, shortwave and thermal) and high resolution in O2-A and O2-B, which are considered sufficient for accurate cloud screening [Gómez-Chova et al., 2009]. The detection of semitransparent high-altitude cirrus clouds would be possible by means of collocated Sentinel-3 SLSTR measurements in the so-called cirrus bands at 1375 and 1610 nm water vapor bands.

[36] 2. A second step is FIMAS spectral characterization. The accurate knowledge of the instrument spectral response is known to be essential for the reliable Fs retrieval. For this reason, a processing step devoted to the scene-based spectral characterization of FIMAS must follow cloud screening. The effects of spectral shift, spectral broadening, and array compression or stretching are decorrelated from those of the environmental parameters under consideration. This enables to perform spectral characterization as an independent processing step. An algorithm for the spectral characterization of FIMAS has already been developed. It estimates spectral channel position, bandwidth and a parameter accounting for spectral stretching or compression from O2-A and O2-B spectra.

[37] 3. A third step is aerosol retrieval. The synergy between Sentinel-3 OLCI's visible and near-infrared spectral coverage and SLSTR's dual view is expected to provide sufficient information for accurate aerosol retrieval. OLCI presents two channels in the shortest visible wavelengths (410 nm and 440 nm), whereas SLSTR will measure in nadir and backward (55° views. Good performance of existing AOD retrieval methods from OLCI and SLSTR predecessors MERIS and AATSR are in the literature [e.g., Grey et al., 2006; Vidot et al., 2008; Guanter et al., 2008; Kokhanovsky et al., 2007]. The exploitation of measurements in the blue wavelengths, where the aerosol contribution is normally highest, the dual-view capability to discriminate between surface and atmospheric signals, and the fact that fluorescent green vegetation targets have a very dark reflectance response in the visible wavelength range are expected to enable AOD retrieval to the accuracy levels required by Fs retrieval.

[38] 4. In the surface pressure retrieval step, existing algorithms for surface pressure retrieval from MERIS data will be adapted to Sentinel-3 OLCI data [Lindstrot et al., 2009]. These algorithms, complemented with a high precision digital elevation model and collocated near-real-time pressure measurements (e.g., from the European Centre for Medium-Range Weather Forecasts) can provide a very accurate SPR product. The high spectral resolution of FIMAS measurements in O2-A could also be applied independently for surface pressure retrieval. In this case, surface pressure retrieval could be performed before cloud screening so that the pressure product could be used for cloud detection purposes. As discussed previously, the retrieval of surface pressure from FIMAS measurements would compensate the error introduced by the uncertainty of the aerosol vertical profile.

[39] Once these preprocessing steps have been applied, the at-sensor atmospheric parameters in equation (1) can be calculated for clear-sky pixels. Fs retrieval becomes then a problem of decoupling reflectance and fluorescence from TOA measurements.

3.3. Decoupling of Fluorescence and Reflectance From TOA Radiance Spectra

[40] The baseline method proposed for Fs and ρs decoupling is based on a spectral fitting method (SFM), and is applicable to both O2-A and O2-B absorption features. However, an evolution of the FLD method [Plascyk, 1975] designed for the particular properties of the O2-A absorption feature (deep and broad absorption and spectrally linear vegetation reflectance response) and adapted to the satellite case for its use with FLEX data (FLD-S) [Guanter et al., 2007] has also been tested.

3.3.1. Fs Retrieval Based on Spectral Fitting (SFM)

[41] The decoupling of Fs and ρs terms from high resolution TOA radiance spectra can be performed by means of spectral fitting techniques based on linear regression. equation (1) can be linearized in terms of ρs and Fs as

equation image

where ρs,apj is the apparent reflectance (reflectance including the fluorescence contribution) at the FIMAS' jth channel, and

equation image

where 〈 〉j refers to the spectral convolution operation at channel j.

[42] In the proposed approach, surface reflectance within the FIMAS spectral windows is modeled as a linear combination of end-members,

equation image

where ai are the end-member abundance coefficients and ρemj the reflectance end-members. Four end-members, three green vegetation and one bare soil reflectance spectra generated with FluorMODleaf and FluorSAIL, have been found to be sufficient for the reproduction of most of the spectral shapes and red-edge positions in the reflectance library in Figure 4 (top), and also of the mixed vegetation pixels generated in the scene-based simulation data set. The selected end-members are plotted in Figure 7. Previous developments represented ρs by a n-order polynomial, but this approach showed a limited capability to account for real vegetation patterns, especially in the O2-B band, and a much higher sensitivity to instrumental noise. The extension of the set of end-members to include regionally representative vegetation reflectance patterns can be considered for future work.

Figure 7.

Reflectance end-members used in the decoupling of Fs from ρs.

[43] The fluorescence contribution is modeled as a fixed spectral pattern f(λ) derived from models modulated by an intensity coefficient F0,

equation image

This simple formulation of Fs is considered sufficient for fluorescence retrieval according to the low spectral variability of the fluorescence spectrum within the O2-A and O2-B spectral windows.

[44] Using equations (4) and (5) in equation (2), and assuming that the atmospheric variables are provided by external data sources, reflectance and fluorescence contributions to TOA radiance ({ai, F0}) can be estimated by least squares optimization of the ∼200300 FIMAS spectral channels.

3.3.2. Fs Retrieval Based on the FLD Method for Space Measurements in O2-A (FLD-S)

[45] As an alternative to the SFM, a modified version of the FLD method has also been implemented for fluorescence retrieval from FIMAS measurements in the O2-A band. This FLD-like approach takes advantage of FIMAS' high spectral sampling measurements to compensate for nonlinear spectral variations of reflectance and fluorescence. Fluorescence in the O2-A spectral region is retrieved by the solution of the system of two equations expressing the TOA radiance inside and outside the O2-A band:

equation image

where the subscripts i, o refer to in- and out-of-band finite spectral intervals, i defined to be at the bottom of the O2-A absorption feature (defined at 760.5 nm) and o at the continuum region (757.0 nm). Macrochannels of ±0.5 nm in the bottom of the O2-A feature and of ±1 nm in the borders are generated by channel binning in order to reduce retrieval errors from instrumental noise and spectral calibration issues.

[46] To constrain the system, correction coefficients are defined in order to relate ρs and Fs within the two spectral windows,

equation image

The A coefficient in equation (7) is derived from apparent reflectance ρs,ap, assuming that ρs and ρs,ap have the same spectral derivative around 760 nm.

equation image

Apparent reflectance is calculated from the inversion of equation (1) for Fs = 0. The assumption in equation (8) is one of the weakest points in FLD-like methods, as the spectral derivative of intrinsic and apparent reflectance can be very different for large Fs values [Alonso et al., 2008]. This is refined by an iterative procedure which updates A once a first estimation of Fs is performed. The apparent reflectance inside the absorption band, which is necessary for the calculation of the A coefficient, is calculated by means of polynomial interpolation from the continuum region in order to account for nonlinear trends in the reflectance spectral pattern. On the other hand, a fixed value of 0.125 has been selected for B from models, which is justified by the low spectral variation of the fluorescence emission in this region of the spectrum [Alonso et al., 2008].

[47] From equations (6) and (7), Fs in the O2-A band is given by

equation image

where

equation image

3.4. Use of Reference Targets to Constrain Fs Retrieval

[48] Previous analysis have shown that Fs retrieval in both O2-A and O2-B regions can be biased by a number of environmental and instrumental factors. In principle, errors in AOD retrieval, aerosol model and vertical profile, surface pressure, polarization, temperature profiles or radiometric or spectral errors would affect Fs retrieval. However, it can be a priori expected that the resulting errors in Fs may appear systematically in all the pixels with similar environmental and instrumental conditions.

[49] Nonfluorescent surfaces can be used to improve Fs retrieval by the normalization of those systematic errors. These surfaces could be used for the estimation of the zero error in Fs retrieval over potentially fluorescent targets, or to constrain the estimation of atmospheric parameters by forcing reference targets to give Fs = 0. Suitable reference targets must have a reflectance response comparable to that of vegetation, and should be located close enough to vegetation pixels to validate the assumption that at least the same values of surface pressure and aerosol conditions apply to both reference targets and green vegetation pixels.

[50] In order to ensure that errors in Fs are systematic to both vegetation and bare soil pixels, the initial Fs retrieval over vegetation and soil surfaces must be performed using a retrieval approach with the minimum sensitivity to the background reflectance. This would be the FLD-S method from the approaches discussed previously, as end-member-based spectral fitting approaches are optimized for application to green vegetation targets that can be represented by the end-members in the spectral library.

[51] It must be remarked that the need for such reference targets in Fs retrieval would handicap FLEX spatial coverage, as Fs retrieval could only be performed over areas where a sufficient number of these surfaces is available. This would discard most of the dense green forest covers in the planet. The best GSD would be required in FIMAS in order to identify the purest vegetation-free pixels.

4. Results

4.1. Spectrum-Based Sensitivity Analysis

[52] The sensitivity of Fs retrieval to different spectral configurations in FIMAS and to the uncertainty in aerosol parameters and surface pressure has been tested in the first place by means of a spectrum-based simulated database. FIMAS-like TOA radiance data in O2-A and O2-B were generated for varying aerosol and surface pressure values, the rest of the input parameters being those described as the reference simulation setup in Figure 2. The width of the spectral window and the SSI have been varied in the simulations according to the variation ranges given in Table 1. For each spectral configuration, which is associated to a given SNR figure, normally distributed random errors in AOD550, SPR and abundances of the rural, urban and maritime aerosol models have been simulated. Up to 1000 independent cases have been generated for each set of parameters, varying each at a time or all at the same time. Mean and maximum absolute errors in AOD550 were 0.035 and 0.075, respectively, and 2.5 hPa and 5 hPa in surface pressure. Fs is calculated from each of the 1000 cases by means of the SFM (O2-A and O2-B) and the FLD-S (only in O2-A) retrieval schemes presented in section 3.

[53] Root mean square error (RMSE) in Fs retrieval associated to instrumental noise and to the uncertainty in AOD550, surface pressure and aerosol model for the SFM and FLD-S retrieval methods and spectral windows of 20 nm and 30 nm is presented in Figure 8. It can be observed that Fs the estimated total RMSE is larger in O2-A than in O2-B. This is explained by the relatively low contribution of Fs to TOA radiance with respect to reflectance and atmospheric distortion in O2-A, as shown in Figures 23. Errors in O2-A Fs retrievals with the SFM method are driven by instrumental noise and surface pressure, while less sensitivity to noise is found in the FLD-S method due to the spectral binning performed to generate the i,o macrochannels. The total error is higher for FLD-S than for SFM, which confirms the results reported by Meroni et al. [2010] about the better performance of SFMs for fluorescence and reflectance decoupling. Concerning Fs retrieval in O2-B, the impact of both instrumental noise and atmospheric parameters remains very small due to the relatively high contribution of Fs to TOA radiance. It is also observed that the error decreases with the increase of the spectral window both in O2-A and in O2-B due to the better discrimination between reflectance and fluorescence with the wider spectral range.

Figure 8.

Root mean square error (RMSE) in Fs retrieval associated to instrumental noise and to the uncertainty in AOD550, surface pressure, and aerosol model for the SFM and FLD-S retrieval methods and spectral windows of 20 and 30 nm. The label ALL refers to the combination of the four error sources.

[54] In order to complete the sensitivity analysis presented before, the error in Fs retrieval caused by the uncertainty in the aerosol vertical profile has been assessed. FIMAS-like TOA radiance data have been simulated for a range of aerosol vertical distributions, each shifted in height by steps of 1 km. Fs retrieval is performed under the assumption of the aerosol layer being located in the range 2–9 km. The systematic absolute error estimated from this simulations is shown in Figure 9. The different vertical axes for O2-A and O2-B must be noted. It is observed that the impact of the aerosol vertical profile on O2-B retrievals is negligible, while it is very important in the case of O2-A for the two retrieval methods and spectral windows. Errors in O2-A due to the uncertainty of the aerosol vertical profile are of up to 150% for the simulated cases.

Figure 9.

Bias in Fs retrieval from the uncertainty of the height of the aerosol layer for the SFM and FLD-S retrieval methods and spectral windows of 20 and 30 nm.

[55] It must be remarked that the absolute errors in Figure 8 and Figure 9 are not fully representative of the real case, as they are referred to a pure green vegetation reflectance pattern. It is expected that those errors increase for mixed vegetation and soil pixels, as corresponds to the real case. With this consideration, RMSE in O2-A from both the SFM and FLD-S methods is reckoned as too large for the subsequent exploitation of the fluorescence signal if no Fs retrieval normalization is performed. This conclusion reinforces the need to use of reference targets to constrain Fs retrieval in O2-A.

4.2. Scene-Based Sensitivity Analysis

[56] The scene-based simulated data set can be used to provide more information about the feasibility of using reference targets for Fs retrieval and about the impact of different vegetation types and distributions.

[57] An example of scene-based Fs retrieval in O2-A after normalization with reference surfaces for the Northern Spain site #6 in Figure 5f is displayed in Figure 10. Maps of input Fs, retrieved Fs and retrieved Fs after normalization by reference targets are displayed. Fs is retrieved with the FLD-S approach, as the end-member-based SFM might lead to biases in Fs retrieval due to the wrong representation of soil spectral reflectance by the selected vegetation end-members. The normalization by reference surfaces was performed by subtraction of a Fs correction map derived from 2-D interpolation from all the bare soil areas in the image. These were defined as those with 0.0 < NDVI < 0.2. An overall overestimation of fluorescence of about 0.8 mW m−2 sr−1 nm−1 when no normalization by reference targets is performed is observed in Figure 10b. This error is mostly due to the uncertainty in the input atmospheric parameters, which in this simulation are those depicted in Figure 6, plus a constant deviation in the aerosol model from the 50%, 20% and 30% abundances of the rural, urban and maritime models, respectively, to the 70%, 0% and 30% proportions used in the retrieval step. The improvement from normalization by the nonzero Fs retrieved over bare soils can be noticed in Figure 10c. It can be observed that input and retrieved Fs levels are very similar within the entire area. It must be remarked that the area in Figure 5f can be considered the best case scenario for the application of this normalization method, as a number of bare soil surfaces are distributed uniformly over the image.

Figure 10.

Input and retrieved Fs maps in O2-A with the FLD-S approach before and after the normalization by reference soil targets.

[58] Results of Fs retrieval from three of the simulated test sites in Figure 5(#1, #2, #3) and for the 30 nm spectral window configuration are presented in Figure 11. Fs retrieval in O2-A is performed by the FLD-S technique with normalization by reference targets, while the SFM is used for O2-B. Different distributions of green vegetation patterns, from very green and homogeneous vegetation bodies to very sparse vegetation pixels, and different AOD and SPR input error maps are present in the three sites. RMSE and the Pearson's correlation coefficient R2 are calculated between the input and the retrieved Fs maps. Mean relative errors in AOD550 and SPR for each site are included in the plots. Only green vegetation pixels, defined as those with NDVI > 0.7, are considered to generate these plots. The very different performance in O2-A and O2-B stated in Figures 89 can also be observed in the results in Figure 11. Very high linear correlations between the input and retrieved Fs values are derived for O2-B, for which a systematic overestimation of the retrieved values is also detected. Since the absolute errors in the input AOD550 and SPR can be indistinctly positive or negative for each site, such bias is assumed to be due to a bad reconstruction of the vegetation reflectance pattern by the SFM end-member technique applied to O2-B. It appears in all the 6 sites tested. The decoupling between Fs and ρs in O2-B improves as the Fs values increase. In the case of the O2-A band, the correlation between input and output Fs is smaller, although the retrieval bias is smaller than in O2-B due to the normalization by nonfluorescent targets.

Figure 11.

Results from Fs retrieval in O2-A and O2-B for a spectral window of 30 nm (745–775 and 672–702 nm, respectively) over three of the test sites in Figure 5. Colors from blue to red depict an increasing density of points for a given interval of Fs values.

[59] A summary of the results obtained from all 6 simulated sites in Figure 5 is depicted in Figure 12. RMSE and R2 calculated between the input and the retrieved Fs maps are plotted for the O2-A and O2-B regions and 20 nm and 30 nm spectral windows. From the analysis of the results obtained from all the sites for the 20 nm and 30 nm band configurations, it can be observed that Fs retrieval performance in O2-A is highly site-dependent. This is probably due to the different availability and quality of nonfluorescent reference targets from one site to another. Differences can also be explained the dominant vegetation reflectance pattern, proportion of green vegetation pixels, and AOD and SPR absolute values and error spatial patterns. In O2-A, the calculated RMSE varies from about 0.2 to 1.0 mW m−2 sr−1 nm−1, and the R2 from about 0.5 to 0.8. Considering RMSE above 0.5 mW m−2 sr−1 nm−1 as too large for the usability of the Fs signal, Fs retrieval would only be within the admissible accuracy range in three sites. It is also demonstrated that RMSE figures are improved by up to 0.1 mW m−2 sr−1 nm−1 if the 30 nm spectral window configuration is used. The improvement in R2 is of 0.1–0.2 depending on the site. Such a relatively small dependence of Fs retrieval accuracy in O2-A with the width of the spectral window are a priori expected, since Fs retrieval is performed with the FLD-S approach, which does not take particular advantage of the larger spectral window. This is different in O2-B. Important improvements in RMSE and R2 are found in the 30 nm spectral window case. RMSE is below 0.5 mW m−2 sr−1 nm−1 for the 30 nm band in all the test sites, and R2 increases from 0.85 to 0.95. RMSE grows to more than 0.8 mW m−2 sr−1 nm−1 in two sites in the 20 nm spectral window case. This may be explained by the existence of a given dominant vegetation spectral pattern in the site that can only be well retrieved by the SFM end-member-based inversion for the widest spectral window, which provides a better sampling of reflectance nonlinear trends along the vegetation red edge region.

Figure 12.

Root mean square error (RMSE) and Pearson's correlation coefficient R2 of Fs retrievals calculated from the six test sites in Figure 5.

5. Summary and Conclusions

[60] The measurement of top-of-canopy chlorophyll fluorescence from space is currently being subject of research in the frame of the preparatory activities of the ESA FLEX mission. The Fs retrieval rationale in FLEX is the measurement with high spectral resolution inside the O2-A and O2-B atmospheric absorption features overlapping the Fs emission so that the differential optical paths crossed by fluorescence and the reflected solar radiation can be used to decouple the two contributions from TOA radiance spectra.

[61] Recent developments for fluorescence retrieval from space measurements have been presented in this work. It describes a prototype processing chain for Fs retrieval from FLEX/FIMAS-like high spectral resolution data. It is assumed as a baseline for this processing chain that the ancillary information necessary to characterize the atmospheric state prior to Fs retrieval is provided by GMES Sentinel-3, which would fly in tandem with FLEX eventually. Spectral fitting and FLD-like techniques are proposed to decouple the fluorescence and reflectance signals from TOA measurements once the atmospheric optical parameters are provided by external means.

[62] This methodology has been tested on a simulated data set of FLEX/FIMAS spectra and images. Different green vegetation reflectance patterns and fluorescence levels have been propagated to the TOA radiance level by means of radiative transfer simulations. Realistic variation ranges for the width of the spectral window, spectral sampling and SNR specifications of FIMAS are considered to analyze the suitability of different instrumental configuration for Fs retrieval. Errors in the knowledge of AOD550, aerosol model, aerosol profile and surface pressure have been considered in the simulation process. Other atmospheric parameters which may also interact with the O2-A and O2-B bands have been discussed, but have not been explicitly considered in the simulation. Those are thin and subpixel clouds, directional reflectance, temperature vertical profiles, polarization, water vapor, and the Ring effect and dayglow emissions. It is assumed that these factors are either of second-order importance for Fs retrieval with respect to aerosols and surface pressure, or that can be accounted for by means of models or external data sources.

[63] Results obtained from Fs retrievals over the simulated data set indicate that Fs can be retrieved with a RMSE smaller than 0.5 mW m−2 sr−1 nm−1 from O2-B measurements in most of the cases, while such accuracy has only been achieved in 3 of the 6 sites in O2-A and by means of the normalization by reference targets. The highest Fs retrieval accuracy have been obtained for the best spectral sampling of 0.1 nm and the widest spectral window of 30 nm if the SFM is used, either in O2-A or O2-B, whereas no noticeable differences are found when the FLD-S method is chosen for O2-A. The better performance in O2-B can be explained by the relatively higher contribution of Fs to TOA radiance in the red region, where vegetation reflectance is very low. In terms of methods, the best performance in O2-B has been found for the end-member-based SFM, while the combination of the FLD-S method with the normalization of Fs by nonfluorescent reference targets has led to the best results in O2-A. However, the implications that the use of this normalization technique might have in FLEX in terms of spatial resolution and site coverage recommends that alternative approaches should be developed to reduce the high sensitivity of Fs retrieval in O2-A to errors in the knowledge of the atmospheric conditions.

[64] Future research in the field of Fs monitoring from spaceborne high-resolution measurements is to be planned depending on the evolution of the FLEX mission. The improvement of Fs retrieval by the coupled inversion of Fs, ρs, AOD and SPR from FIMAS spectra or by the simultaneous inversion of spectra from the O2-A and O2-B atmospheric features, and the analysis of directional effects on Fs retrieval are already under consideration.

Acknowledgments

[65] This work has been done in the frame of the projects Atmospheric Corrections for Fluorescence Signal Retrieval (ESA-ESTEC contract 20882/07/NL/LvH) and FLEX Performance Analysis and Requirements Consolidation Study (ESA-ESTEC contract 21264/07/NL/FF).

Ancillary