[14] In the present paper, we are dealing with SeaWiFS derived products remapped to a sinusoidal projection at a spatial resolution of 2.17 × 2.17 km^{2}, such that the analysis of FAPAR time series can be conducted, without performing a spatial aggregation, over pixels identified as the nearest to the nominal location of the measurement site. It also means that this resolution is slightly too coarse to fully ensure that the groundbased, eventually domainaveraged, measurements result from a complete spatial sampling at that same resolution. Over these selected pixels, time composite algorithms can be applied to eliminate outliers and to limit the impact of uncertainties inherent in the algorithm (e.g., remaining biases due to changing Sun and view geometries and/or unforeseen atmospheric conditions), intermittent presence of subpixel clouds, or any other undesired events such as occasional water or snow during the composite period. For the present comparison exercise, daily and 10day composite values will be used. In the next section, a number of issues concerning the difficulties in assessing domainaveraged FAPAR values are discussed and associated with different categories of radiation transfer regimes, themselves related to different types of vegetation canopies.
4.1. GroundBased Estimations of DomainAveraged Values
[15] Domainaveraged FAPAR (the absorbed flux fraction estimated over the photosynthetically active radiation (PAR) spectral region) is neither measured nor simulated directly as such: Its estimation results from the closure of an energy balance equation in the vegetation layer at the spatial resolution of the domain. Under direct illumination with cosine zenith angle μ_{0}, it is equal to
where R(z_{toc}, μ_{0}) represents the albedo at the top of the canopy z_{toc} and T(z_{bgd}, μ_{0}) represents the total transmission (both the direct and diffuse components) at the bottom of the vegetation layer z_{bgd}. R_{bgd} denotes the albedo of the background underneath the vegetation layer, and H(μ_{0}) corresponds to the net effect (positive or negative) due to the horizontal fluxes through the lateral boundaries of the domain. Note that T(z_{bgd}, μ_{0}) includes the contributions due to scattering processes in the vegetation layer as well as those due the multiple interactions between the vegetation layer and the background [Pinty et al., 2005]. All normalized radiant fluxes listed in (1) are domainaveraged quantities.
[16] Equation (1) asserts that the net horizontal flux, in addition to three domainaveraged fractions of vertical fluxes, needs to be known in order to estimate FAPAR. The technical difficulties associated with the in situ assessment of the vertical fluxes depend on the canopy attributes, including average height and spatial heterogeneity. For short vegetation canopies, on the one hand, the measurements of the fluxes impinging on the background and arising from the background are rather difficult; on the other hand, the albedo of the vegetation canopy and its spatial variability can be assessed relatively easily. By contrast, in the case of tall vegetation, it is the estimate of spatially representative albedo values of the vegetation canopy that is rather difficult, and especially the documentation of the spatial variability of this normalized flux. Depending on the size of the footprint of the measuring device, high enough towers sometimes offer appropriate support. However, it remains hard to guarantee that this device samples statistically all scales of variability exhibited by the vegetation layer, such that its measurements can be analyzed to estimate representative domainaveraged values to be used in (1). The downward transmitted flux and its spatial variability at specific scales and resolutions is technically easier to estimate below tall canopies [e.g., Brown et al., 1994; Comeau et al., 1998]. When this variability is large (e.g., large values of the variance (or higher moments) with respect to the mean), the estimation of representative domainaveraged fluxes is even more complicated [e.g., Nicotra et al., 1999]. In these circumstances, it is difficult to ensure that the different sampling techniques used at the top and at the bottom of the vegetation layer lead to statistically consistent domainaveraged flux values needed for closing (1).
[17] When estimating FAPAR from medium or low spatial resolution sensors, the contribution due to the net radiant horizontal fluxes is negligible at any level with respect to the vertical fluxes [see Widlowski et al., 2006]. For all practical purposes, the inverse algorithms implemented to interpret medium and lowresolution remote sensing measurements can thus safely assume that H(μ_{0}) 0. By the same token, these results also suggest that at local resolutions of less than a few meters, such as those involved when performing in situ measurements, the contribution from H(μ_{0}) (which is rarely or never measured in situ) can be quite significant, especially for tall and open canopies. This contribution has to be compared with the one due to the horizontal variability of the vertical fluxes. One of the challenges when performing ground validation experiments (with the condition H(μ_{0}) 0) is to ensure that (1) the field measurements adequately sample the internal variability of the domain at the appropriate frequency (in the case of vegetation layers; this implies a sampling step corresponding to the typical interleaf/shoot distance); and (2) the sampled area is large enough so that the associated domainaveraged values are independent of the exact location of this sampled area within the vegetation system (i.e., deriving domainaveraged values that are representative of the system). For example, tall canopies with low leaf densities generally exhibit significant 3D structure and hence a hierarchy of gaps where the internal variability of the leaf distribution function is quite high [e.g., Parker, 1995]. In such conditions, the horizontal extent of the domain sampled in the field must be several times the typical height of the canopies to reach conditions where the relative contribution of H(μ_{0}) with respect to the vertical fluxes is negligible.
[18] The joint estimation of the appropriate domainaveraged vertical and horizontal flux quantities entering (1) in the PAR spectral region is therefore crucially dependent on the sampling frequency and the size of the sampled domain as well as the architectural attributes of the vegetation layer and the nature of the illuminating sources both at the upper (downward direct and diffuse sky radiation) and lower (upward radiation field scattered by the background) boundaries. For all practical purposes, the spatial variability of the leaf density prevailing inside the vegetation layer over the domain is one of the most delicate problems to be addressed both technically and statistically for accurately closing (1) on the basis of the individual flux contributions. As a consequence, one can anticipate that the accuracy level that can be reached when computing (1) genuinely depends on the level of statistical variability of the vegetation layer and the means to estimate it.
[19] The impacts of different types of internal variability of the extinction coefficient together with the resolution of the sampled domain on the radiation transfer regime for clouds was analyzed by Davis and Marshak [2004]. On the basis of theoretical arguments, these authors thoroughly established the conditions where 3D effects are anticipated to play a major role in the establishment of the radiation transfer regime. Their results can be extrapolated to the case of land surfaces to help us to associate the main radiative transfer regimes with the statistical properties of the leaf extinction coefficient inside the spatial domain of investigation: (1) a “fast” variability regime in the case of statistically homogeneous, Poissonlike, distributions of the leaf density, (2) a “slow” variability regime where, in fact, the leaf density distribution is close enough to being homogeneous only locally such that localscale average flux values are meaningful and (3) a “resonant” regime in other cases where the spatial complexity is such that a typical photon beam will sample different types of vegetation structures between entering and escaping the canopy, variability thus controls the domainaverage fluxes. Such regimes correspond to an intrinsic radiative property of the canopies themselves as they interact with the flow of solar radiation. We will however exploit this identification scheme from the practical standpoint of a given spatial resolution.
[20] Canopies and resolutions favoring regimes 1 and 2 are simpler to deal with than those exhibiting regime 3 in the context of both performing in situ flux measurements and modeling the radiation transfer phenomena. Indeed, these former regimes basically call for the use of onedimensional (1D) theory to be applied either over the full domain (regime 1) or only “locally” and then extended over the full domain using linear mixing techniques (regime 2). In the case of regime 3, canopy structure is a major component of the problem and the 3D heterogeneity controls a significant part of the domainaveraged flux values that are themselves neither easily estimated in the field nor modeled accurately, unless the statistics of the canopy elements are well known.
[21] For all practical purposes, short vegetation and tall but dense canopy layers tend to exhibit only small characteristic scales, close to the typical interleaf/shoot distance. If only one type of such vegetation exists in the studied domain for FAPAR, regime 1 prevails. When multiple types of vegetation/land cover exist inside the domain, a typical case for medium to low spatial resolution sensors, any vegetation storey must be sampled such that the average “local” flux values can be estimated in a statistically meaningful manner. The domainaveraged values are then generated by a straightforward spatial aggregation of the vegetation types, weighted by the fractional area inside the domain. Regime 2 may thus dominate, especially if the conditions for regime 1 are fulfilled “locally” over each individual land cover patch. One may logically expect that tall (favoring long horizontal displacement of radiation with respect to its origin of entry) and clustered (creating a hierarchy of small scale gaps) but not dense (permitting radiation to travel with small extinction probability) vegetation canopies exhibit the type of variability categorized in regime 3. The latter occurrence calls for a wellconceived measuring protocol [Widlowski et al., 2006].
[22] In actual conditions, all sorts of combinations of such regimes coexist at medium and low spatial resolutions. Vegetation canopies are also composed of woody elements for which both statistical and radiative properties significantly differ from those of the leaves. Given the many caveats and difficulties associated with the accurate estimation of domainaveraged fluxes, the next section discusses some approximations that can be made to simplify the problem of comparing FAPAR values retrieved from remote sensing against those deduced from ground estimations of measurable fluxes and major canopy attributes. Such simplifications are needed to better understand the interpretation of the ground data sets selected for comparison, as well as to acknowledge up front some of their limitations.
4.2. Intercepted Radiation: A Proxy for FAPAR
[23] Since the fraction of absorbed flux is a function of the directionality of the illumination source, an equation similar to (1) holds for diffuse sky illumination and, to simplify the problem further, one can assume that FAPAR estimates from ground flux measurements, FAPAR(μ_{0}), are approximated by
where the weights, f^{dir} and f^{diff}, sum up to 1 and correspond to the fractions of direct and diffuse to total downward flux density, respectively. Equation (2) is based on the reasonable assumption that the FAPAR, as would be measured in the field under actual conditions of illumination, can be approximated by a linear combination of two FAPAR contributions: a directionalhemispherical flux, FAPAR(μ_{0}), associated with a purely collimated incident intensity field, and a bihemispherical flux, , associated with a purely isotropic incident intensity field. Such an approximation is largely used for approximating the albedo of land surfaces under ambient conditions [Lewis and Barnsley, 1994; Pinty et al., 2005]. Equation (2) provides a means to estimate FAPAR under arbitrary sky conditions, provided that the vertical and horizontal fluxes in (1) can be estimated under direct and diffuse illumination separately.
[24] Since the single scattering regime (identified by a superscript 1 s) dominates the absorption process by vegetation in the PAR spectral region [see Pinty et al., 2006] it is possible to rewrite (2) as follows:
where the fluxes associated with the incoming diffuse sky illumination are identified by an overbar and where FIPAR(μ_{0}) denotes the fraction of direct radiation intercepted by the leaves only; that is,
where the socalled uncollided transmission, T_{veg}^{UnColl}, denotes the direct transmission of solar radiation that has traveled downward through the canopy gaps only and thus has not suffered from any collision with canopy elements. A similar equation can be written with respect to the diffuse sky illumination.
[25] In the limit of single scattering regime and under direct illumination, can be accurately approximated by [see Pinty et al., 2006]
is thus a function expressing the balance between the contributions due to the fraction of upscattered flux at the top of the canopy, that is the albedo of the canopy, R_{veg}^{Coll1 s}, (negative contribution), and the fraction of upscattered flux at the bottom of the canopy (positive contribution). The latter is logically given by the product of the source term at the background level illuminating the vegetation canopy from below, R_{bgd}T_{veg}^{UnColl}(μ_{0}), and the intercepted fraction over the entire upward hemisphere, (1 − ). A similar equation is obtained with respect to the diffuse sky illumination, by replacing the directionalhemispherical with their corresponding bihemispherical quantities in (5).
[26] As discussed in section 4.1, the estimate of radiant fluxes in both structurally heterogeneous and homogeneous vegetation canopies is rendered somewhat complex by the type of boundary conditions to be considered: direct and diffuse radiation at the top of the canopy and a reflecting background at the bottom. Note that this situation has been significantly simplified by assuming an isotropic diffuse sky illumination yielding the linear parameterization proposed in (2) for estimating the radiant fluxes. It can be further simplified if a Lambertian background is considered thus creating an isotropic source of radiation at the bottom of the canopy.
[27] Equation (5) highlights the role of the lower boundary condition which acts as an additional source of radiation to the vegetation layers and thus contributes positively to the absorption process. Its impact on the fraction of absorbed flux, has been estimated numerically for various background brightness conditions underlying a selection of canopy scenarios in the PAR spectral domain. The 3D heterogeneous canopy scenarios used below are the same as those described by Pinty et al. [2006, Tables 1 and 2]. They correspond to sparse, medium and dense forest canopies with allometric domainaveraged LAI values, 〈LAI〉, equal to 1.24, 2.0 and 4.82, respectively.
[28] The values of are displayed in Figure 4 as a function of the background brightness, R_{bgd}. The leaf absorption efficiency is such that for dense canopy conditions, the function values remain close to zero for almost any background condition. Since the decrease in 〈LAI〉 is accompanied by a higher probability for the solar radiation to reach the background, the effect of the latter increases up to a maximum reached for intermediate 〈LAI〉 values. Indeed, 〈LAI〉 values that are too low do not allow significant absorption to occur over the domain. Results from Figure 4 also confirm our expectations that tall, clustered and lowdensity vegetation layers, featured in the medium density canopy condition in Figure 4, constitute difficult cases. They can, indeed, yield values in the range +0.05 to +0.1 for bright enough background reflectance values, that is, larger than about 0.3 (a scenario occurring when the background is covered by snow, for instance). By contrast, these values are quite close to zero for environmental conditions where the vegetation layer is bounded by a vegetated understorey. Under diffuse sky illumination, the respective contributions to the fraction of absorbed flux due to the top of the canopy and background albedos are balanced for R_{bgd} values equal to 0.03, 0.04 and 0.46, for the sparse, medium and dense canopy conditions considered in these examples [Pinty et al., 2006, equation (36)].
[29] These simulation results suggest that for typical conditions of illumination and background reflectance, the main contribution to FAPAR is given by the intercepted fraction, with respect to both the direct and diffuse illumination. It was shown by Pinty et al. [2006, Figure 8] that the contribution to FIPAR due to diffuse sky illumination hardly exceeds +0.03 for typical clear sky conditions. Thus, together with results from Figure 4, one may anticipate that uncertainties due to the contributions from both and can be neglected in the overall uncertainty budget of FAPAR under typical conditions.
[30] The uncertainty of ground estimations of FAPAR, identified with a Δ symbol, can be approximated by
Equation (6) assumes that the relative uncertainty due to the assessment of the fraction of direct versus diffuse radiation is negligible as compared to other sources previously discussed. Thus, according to (4), (6) becomes
Equations (4) and (7) express the fact that measuring the probability distribution function of the gaps, i.e., T_{veg}^{UnColl}, over the domain constitutes a realistic and technically relatively simple approach for assessing both domainaveraged FAPAR and associated ΔFAPAR values. Such measurements are indeed feasible thanks to optical field instruments including hemispherical photographs [Rich, 1990], the Tracking Radiation and Architecture of Canopies (TRAC) instrument [Chen and Cihlar, 1995b], the Burr Brown Data Acquisition System (BBDAS) [Lang and Yueqin, 1986] and the LAI2000 Plant Canopy Analyzer (LICOR, Lincoln, Nebraska) [Gower and Norman, 1991]. Note that the sampling design with any of these instruments will probably be the limiting factor in determining accurate T_{veg}^{UnColl} values over the domain.
[31] Since these instruments are used to measure the total intercepted radiation, their results include a slight contribution from multiple scattering and a contribution from all woody elements, trunk and stems, present within the field of view of the instruments which are looking upward from below the canopy. If the latter contribution is not removed, a very difficult task to perform accurately, these measurements may thus translate into an overestimate of the FIPAR contribution due to leaves only [Serrano et al., 2000]. In order to translate quantitatively, that is, in terms of Δ T_{veg}^{UnColl} values, the relative contribution from woody elements, the domainaveraged direct transmissions were estimated successively with and without a simplified trunk and branching system [Widlowski et al., 2003] for the 3D heterogeneous canopy scenarios used in section 4.2. For the sparse, medium and dense scenarios, the FIPAR differences are found to be about 0.02 (0.05), 0.03 (0.07), 0.05 (0.02) for a Sun zenith angle of 30° (60°). As one may anticipate given the exponential or power law decay for extinction, the relative contribution due to the woody elements may be especially significant for medium density canopy conditions. A similar range of uncertainty equivalent to Δ T_{veg}^{UnColl} was reported for FIPAR by J. M. Chen et al. [1997], when interpreting measurements of gap fractions from in situ optical devices.
[32] The time delays, equivalent to Sun zenith angle differences, between the acquisition of fluxes on the ground and the remote sensing data increase the difficulty when conducting a comparison procedure. Indeed, FIPAR, a good proxy for FAPAR, is dependent on the Sun angle, but, unfortunately, this dependency is a function of the vegetation type, the ambient atmospheric conditions, the day of the year and time of the day, as well as the latitude of the sampled domain. FAPAR thus increases with Sun zenith angle at a rate that changes with 〈LAI〉 and canopy architecture; these changes are limited for dense canopies (high 〈LAI〉 values) but can be quite significant for medium dense conditions: For example, WalterShea et al. [1992] reported diurnal changes in instantaneous FAPAR values of about 0.2 from measurements collected at the First ISLSCP Field Experiment (FIFE) site. For high Sun zenith angles, for example, larger than 60°, the contribution to FAPAR associated with the diffuse sky illumination dampens the increase rate [Pinty et al., 2006, Figure 9]. Therefore canopies that are not subject to seasonal variations in 〈LAI〉 should logically exhibit a seasonal trend in timebased FAPAR, i.e., higher values in winter than in summer seasons, owing to their dependency with respect to the Sun zenith angle. In the present exercise, this issue will not be quantitatively addressed per se, that is, in terms of Δ T_{veg}^{UnColl}, partly because the needed information is not always available, especially for largescale (both spatial and temporal) investigations.
4.3. GroundBased FAPAR Data Sets
[33] To our knowledge, there is no complete data set that permits addressing all caveats discussed in the previous sections, assembling all the needed vertical and horizontal fluxes separately for the direct and diffuse radiation, measured with the appropriate sampling step and at a spatial resolution compatible with the SeaWiFS products, for the same ambient conditions as those prevailing during the acquisition of the remote sensing data. The previous discussion indicates, however, that such an extremely complex set of measurements may not be needed if we are to validate the SeaWiFS FAPAR JRC products within a ±0.1 uncertainty level. Indeed, we have shown that on the basis of model simulations of realistic vegetation canopy scenarios, the compensation between different contributions is such that approximating FAPAR by FIPAR constitutes a first good step in the comparison process.
[34] In the following exercise, we will thus rely only on a limited number of proxy data sets that are available, excluding highlatitude sites for which too few reliable SeaWiFS FAPAR JRC products are available owing to the occurrence of both large Sun zenith angle (larger than assigned in the algorithm training data set) and subpixel snow, water and cloud contamination. Those selected here include either measurements of local and domainaveraged gap fractions and 〈LAI〉, or combinations of these measurements, and span a wide range of vegetation canopy types which therefore can also be roughly categorized according to their expected or most probable radiation transfer regimes (as deduced from the Davis and Marshak [2004] analysis). The latter categorization is based on qualitative knowledge and description of the field sites and not on the detailed analysis of the leaf density distribution function over the domain as should be done, ideally. As suggested from discussions in section 4.1, it first seems appropriate to classify the field sites according to the domainaveraged heights and densities of the prevailing vegetation. This to some extent postulates that these two metrics are inherently linked to their radiation regimes, and serves as the basis for designing Table 1. This table lists a series of sites and associated references that will be used below to evaluate the SeaWiFS FAPAR JRC product. A summary of the different approaches adopted to estimate FAPAR values is given in Table 2. The detailed characteristics of the field sites, the size of the sampled domain and the full descriptions of the measuring protocols are available in the publications referred to in Table 1.
Table 1. Categorization of Field Validation Sites According to Their Anticipated Radiation Transfer RegimesAnticipated Radiation Regime^{a}  Field Site  Description 


Regime 1 “fast variability” short and homogeneous vegetation over 1–2 km  Dahra^{b}  semiarid grass savannah 
Tessekre^{b}  semiarid grass savannah 
SEVI^{c}  desert grassland 
Regime 2 “slow variability” mixed vegetation with different land cover types  AGRO^{c}  corn and soybean 
HARV^{c}  conifer/broadleaf forest 
De Inslag^{d}  conifer/broadleaf/shrub forests 
KONZ^{c}  grassland/shrubland/cropland 
Regime 3 “resonant variability” intermediate height and lowdensity vegetation  METL^{c}  dry needleleaf forest 
Mongu^{e}  shrubland/woodland 
Table 2. Outline of the Methodology Adopted for Estimating FAPAR at the Field Validation Sites^{a}Field Site Identification  Summary of the Approach for DomainAveraged FAPAR Estimations 


Dahra  based on BBL's law with measurements of the LAD function; FAPAR(μ_{0}) derived from the balance between the vertical fluxes; 〈LAI〉 derived from PCALICOR 
Tessekre  based on BBL's law with measurements of the LAD function; FAPAR(μ_{0}) derived from the balance between the vertical fluxes; 〈LAI〉 derived from PCALICOR 
SEVI  based on BBL's law with an extinction coefficient equal to 0.5; 〈LAI〉 derived from specific leaf area data and harvested above ground biomass; advanced procedure to account for spatiotemporal changes of local LAI^{b} 
AGRO  based on BBL's law with an extinction coefficient equal to 0.5; 〈LAI〉 from leaf area per plant area and plant density; advanced procedure to account for spatiotemporal changes of local LAI^{b} 
HARV  based on BBL's law with an extinction coefficient equal to 0.58; 〈LAI〉 derived from optical PCALICOR data; advanced procedure to account for spatiotemporal changes of local LAI^{b} 
De Inslag  based on full 1D radiation transfer models; 〈LAI〉 derived from optical PCALICOR data; timedependent linear mixing procedure weighted by species composition 
KONZ  based on BBL's law with an extinction coefficient equal to 0.5; 〈LAI〉 derived from optical PCALICOR data; advanced procedure to account for spatiotemporal changes of local LAI^{b} 
METL  based on BBL's law with an extinction coefficient equal to 0.5; 〈LAI〉 derived from optical PCALICOR data; advanced procedure to account for spatiotemporal changes of local LAI^{b} 
Mongu  based on FIPAR estimated from TRAC data; slight contamination by the woody canopy elements; correction to account for the contribution (see equation (5)) 
[35] Figure 5 shows the time series of the SeaWiFS FAPAR products together with the groundbased estimations available from the five sites, namely, Dahra (15°22′N; 15°26′W), Dahra North (15°24′N; 15°26′W), Tessekre North (15°53′N; 15°3′W), Tessekre South (15°49′N; 15°3′W), and SEVI (34°2′N; 106°42′W), all associated with radiation transfer regime 1, corresponding to the socalled “fast variability” category. The baseline FAPAR value for these sites is close to zero and signatures of the different vegetation phenological cycles (both for the growing and decaying periods) are remarkably well identified by both remote sensing and groundbased estimations. Moreover, the amplitudes, both maxima and minima, are in very good agreement with each other although the remote sensing retrievals tend to slightly underestimate the groundbased values over the site of Dahra during the peak season for 2001 (Figure 5, top left). Incidentally, the landscape at this latter site exhibits some significant spatial heterogeneity at mesoscale which was not sampled by the in situ measurements (and thus was not accounted for in the FAPAR groundbased estimations) but which was probably captured at the resolution available from the SeaWiFS FAPAR products.
[36] Results over vegetation conditions belonging to the “slow variability” category, that is radiation transfer regime 2, are displayed in Figure 6. In the particular case of De Inslag (51°18′N; 4°31′E) (Figure 5, top left), since the ground measurements were collected during year 1997, we have plotted the SeaWiFS FAPAR JRC products for the end of 1997 together with those estimated for 1998, thus assuming that no exceptional event occurred during this period over this site. The amplitudes, that is, the summer maxima and winter minima estimated from both remote sensing and ground measurements, are in very good agreement. The inspection of additional daily instantaneous SeaWiFS products suggests that the yeartoyear variations appear to be limited and remain mostly confined within the range of the expected uncertainty of ±0.1, except during the spring period. The groundbased estimated FAPAR values over the agricultural field site identified as AGRO (40°0′N; 88°17′W) follow a welldefined time trajectory that is correctly tracked by the SeaWiFS FAPAR JRC products (Figure 6, top right). We can, however, notice that, on average, the FAPAR maxima and minima from the latter data set tend to be biased high. The third comparison performed with regime 2 canopy conditions, is conducted at the Harvard site (identified as HARV (42°32′N; 72°10′W)), which is a mixture of conifer and hardwood forests. Results from both data sets (Figure 6, bottom left) compare very well with each other for the first 6 months of the year, which includes the growing period. The SeaWiFS FAPAR JRC products then show systematically lower values (about 0.1) than the groundbased estimations during the summer season where vegetation gets very dense over the site. The largest discrepancies, however, occur during the senescent period where a time delay of about 1 month is observed between the FAPAR signatures given by the two data sets. The agreement becomes very good again during the winter season, where the FAPAR values are mostly driven by the relative contribution of the vegetation activity of the coniferous patches [Aber et al., 1996]. Both remote sensing and groundbased estimations of FAPAR over the Konza tallgrass prairie site (identified as KONZ [39°4′N; 96°33′W]) indicate the occurrence of a wellmarked vegetation seasonal cycle (Figure 6, bottom right). These two estimations are well correlated along the cycle over this site covered by mixed grassland/shrub land and cropland, although the SeaWiFS FAPAR JRC products are slightly biased low. Such a bias occurring during the period of senescence may be a consequence of using total (in groundbased estimations) instead of green (as assumed in the retrieval algorithm) 〈LAI〉 values when assessing the FAPAR values. The differences in both 〈LAI〉 values could indeed be somewhat significant, as reported for instance, by T. H. Chen et al. [1997].
[37] The high patchiness of the medium resolution domains, investigated in the case of vegetation canopies associated with regime 2, and the uncertainties inherent in the geographical colocation of the field sites and the remote sensing products, decrease the probability of comparing the radiant fluxes precisely enough over the same domains. This is a reasonable argument to be invoked when dealing with a quite spatially and temporarily changing environment sampled with a medium resolution space sensor. However, given the quite complex procedures involved in the assessment of the groundbased FAPAR values, various other factors, including the Sun zenith angle effects, could also explain some of the observed biases and discrepancies when occasionally larger than ±0.1.
[38] The results of this comparison exercise for vegetation conditions associated with regimes 1 and 2 are summarized in Figure 7 (based on 10day composite values). It shows that, generally, the FAPAR values retrieved from a medium resolution space sensor are well within the specified uncertainty range of ±0.1, when directly compared to a series of available groundbased proxy data sets for FAPAR. Given the many sources of sometimes uncontrolled uncertainties and errors that are contaminating this comparison exercise (some of them yielding deviations as large as 0.4), we can consider that Figure 7 displays quite encouraging results for both the remote sensing and groundbased estimations.
[39] The comparison results of groundbased and SeaWiFS retrieved FAPAR over the METL site (44°26′N; 121°34′W), associated with regime 3 are shown in Figure 8 (top). The two main interesting features are that (1) both sources of information indicate the absence of a strong seasonal cycle, as could be expected over this ponderosa pine conifer forest, and (2) the discrepancy in the FAPAR amplitudes between the two data sets is extremely high (about a factor of 2). Interestingly, this is a typical class of canopies deviating significantly from the 1D statistically homogeneous situation. In that instance, the classical BeerBouguerLambert law of exponential attenuation applies only if the 3D radiative effects are adequately parameterized. As shown by Pinty et al. [2006], all radiant fluxes including FAPAR can be retrieved accurately if effective instead of true domainaveraged state variables are adopted. To do so, Pinty et al. [2006] have proposed inferring a structure factor, ζ(μ_{0}), defined with respect to the domainaveraged transmission factor, T_{veg}^{UnColl}(μ_{0})〉,
with
where (μ_{0}) denotes the effective LAI, a domainaveraged quantity which differs from the true or allometric 〈LAI〉 by a factor embedding a number of effects associated with the 3D heterogeneous structure of the vegetation. Chen and Cihlar [1995a] have reached a similar conclusion regarding the estimate of the direct transmitted flux, and they adopted a factor defined explicitly as a function of parameters representing the foliage clumping and the woodrelated contribution to the extinction of solar radiation.
[40] The ζ(μ_{0}) factor in (9) thus forces the direct radiation to be intercepted, on average over the domain, with statistical rules identical to those prevailing in the case of 1D homogenous canopies, assuming a spherical leaf angle distribution function which implies a leaf extinction coefficient equal to 0.5. This factor is thus equal to unity for 1D homogeneous canopies only. Adopting a value equal to unity for estimating the FAPAR from in situ information over the METL site [Turner et al., 2005] corresponds to an assumption that may therefore significantly contribute to the discrepancy between the groundbased and satellite retrieved FAPAR values over this site. The values of the ζ(μ_{0}) structure factor that would provide the optimal fit between the groundbased and the satellitederived FAPAR value range between 0.5 and 0.6. Incidentally, such values are similar to those simulated with a 3D model for a mediumdensity conifer forest having an allometric 〈LAI〉 equal to 2.0 [Pinty et al., 2006, Figure 2]. It is also noteworthy that such values for ζ(μ_{0}) are consistent with those measured in situ during the international Boreal EcosystemAtmosphere Study (BOREAS) [Chen and Cihlar, 1995a, p. 782]. Indeed, these authors reported an effective LAI value of about 1.5 over a young jack pine forest for which allometric 〈LAI〉 is equal to 2.8. In turn, this reduction in LAI values entering the 1D extinction scheme significantly affects the groundbased estimated FIPAR. Results for this simple analysis therefore suggest that the SeaWiFS FAPAR JRC products could actually well be in the range to be expected for this type of 3D structured lowdensity conditions. Clearly, similar effects could also partly explain some of the discrepancies observed over other sites where some 3D heterogeneity exist. The specific case of METL is especially interesting since, by contrast to other sites, these 3D induced effects are not smeared out by strong seasonal cycles.
[41] The additional groundbased data set associated with regime 3, identified in Table 1 as Mongu (15°26′S; 23°15′E), derives from a careful collection and analysis of the canopy gap fraction using the TRAC instrument over two consecutive years in a mixed shrubland/woodland environment. Figure 8 (bottom) thus shows the time series of the SeaWiFS FAPAR JRC products for years 2000 to 2002 together with the interpreted measurements (in terms of FAPAR spatial averages and associated standard deviations) collected by the TRAC instrument over three transects of 750 m at a spatial resolution of about 1.7 cm. These data include a numerically small (as confirmed by our model simulations shown in Figure 4) correction to account for the contributions. During the two wet seasons, that is, approximately from September to January, the agreement between the SeaWiFS and groundbased estimations is very good. By contrast, SeaWiFS FAPAR JRC products are systematically biased low, by about 0.1 on average, during the two dry seasons, although the uncertainty ranges of both estimations do overlap and the correlation between the two estimations always remains quite high. The groundbased estimations report FAPAR values for 2001 that slightly exceed those from the previous year, leading to an enhancement of the observed bias. The discrepancies are also generally higher in the case of transect C, corresponding to denser canopy conditions, than with transects A and B. On the positive side, one may keep in mind that the remaining contamination of the FIPAR measurements by the woody (nongreen) elements of the canopy favors the occurrence of such a bias with respect to the SeaWiFS FAPAR values. This feature is expected to be enhanced during the dry season when the relative contribution to the extinction process by the leaves only is decreasing, especially with such a lowdensity canopy (the 〈LAI〉 varies approximately in the range [1–1.5] during the dry seasons [Privette et al., 2004]). Much more advanced investigations including, for instance, the Sun angle effects and the 3D modeling of such vegetation scenarios, must be carried out to better understand the observed discrepancies and thus reduce the specified uncertainty range on the SeaWiFS FAPAR JRC products.