187 
Conversely to the good performances above observed about the spectra reconstruction 
capabilities, all of the retrieved values of the four canopy structural variables \LAI, 6 h s, TV] were poorly 
estimated. In many cases, the estimated values are stacked to one of the 2 bounds imposed (Figure 3). On the 
other hand, variables describing the biochemical composition of the leaves are retrieved with a much better 
accuracy as shown in Figure 4 and table 1. This suggests unstable inversion processes for the structural 
variables. Otterman (1990) explained part of these deficiencies when canopy variables appear as a product in 
the mathematical expression of the model. That generally prevents them from being individually inferred. 
Although structural variables did not appear as a simple product in the formulation of the canopy reflectance 
model, they appear grouped in expressions that defined the bidirectionnal gap fractions observed in canopies. In 
the same way, Jacquemoud (1993) showed from numerical experiments that the variables LAI and 0, were 
dependent through the inversion process when using only the spectral variation as the source of information. 
Additional constraints or information must be introduced to stabilize the inversion process. This could be 
achieved by introducing complementary observations under several view and sun geometrical configuration. 
This could also be achieved by assigning fixed values to some of the variables that are not changing 
drammaticaly from one plot to an other, and that has little influence on canopy reflectance spectral variation. 
We choosed this second solution. In the following section,we will investigate the performances of the retrieval 
of a selection of the 6 canopy variables. 
3.1.2. Retrieval of 3 canopy biophysical variables: ¡LAI, Cat, C^J. From simulation studies 
with the PROSPECT+SAIL model, the leaf structure parameter N showed a limited influence on canopy 
reflectance (Jacquemoud, 1993): Roughly, N drives the balance between reflectance and transmittance of the 
leaf. A change in N induces only little variation of the single scattering albedo which is one of the most 
effective variable that governs the spectral variation of canopy reflectance. In these experiments, N was 
estimated from inversion of the PROSPECT model using individual leaf reflectance or transmittance spectra 
measurements. Results show that N varies from 1.00 to 1.38 with an average value N=1.23. The hot-spot 
parameter is the ratio between the average size of the leaves and canopy height. It follows that throughout 
canopy development, the increase of canopy height follows tightly the increase of the average leaf size. 
Consequently, the hot-spot parameter can be assumed to be constant We assigned the hot-spot parameter to its 
mean value: 5 = 0 . 33 . This is in the same range as what was observed by Looyen et al. (1991) for sugarbeet crops 
( 5 = 0 . 5 ). In the same way as for the two previous structural variables, we assigned to the leaf angle inclination 
the value proposed by Baret et al. (1993): The value of 0,=28.6° was found to describe the variation of gap 
fractions with the zenithal direction and leaf area index when assuming a random distribution of the leaves. 
Gap fraction is one of the main characteristics that drive radiative transfer in canopies. Further, because the 
assumption made about the randomness of the leaf position and their azimuthal distributions were identical for 
the SAIL model and the simple Poisson model used to describe gap fractions, this average angle of 0,=28.6° 
was thought to provide good results. 
We investigated eventually the performances of the model inversion when retrieving one 
structural parameter ( LAI) and the 2 biochemical composition variables ( C ab and C w ). Among structural 
variables, LAI expresses the widest variations from plot to plot and is the most effective variable that drive 
canopy radiometric response. The other 3 structural variables were assigned to the values described previously: 
[6 ; =28.6°, 5=0.33, N=1.23]. 
For all the plots, the inversion process ended regularly. Results show that the reconstruction 
performances decreased (mi5e=0.0330).as compared to the previous inversion with the 6 biophysical 
parameters to be retrieved. The biases increase (figure 2b), particularly in the red edge and in the water 
absorption domain (middle infrared). However, the mise is only slightly wavelength dependant, presummably 
in connection with the noise associated to the mesurements. 
The structural variable LAI is now better estimated, except for the white backgrounds (table 
1). The scattering around the 1:1 line (Figure 4) increases with the leaf area index, with some underestimation 
of the actual values except for most of the white backgrounds. The problems observed with the white soils could 
be explained mainly by two facts: (i) The contribution of the soil background to canopy reflectance signal is 
quite strong as compared to the black or natural soil backgrounds. An error on the background reflectance 
values could induce large changes in canopy reflectance and in consequence in the retrieved values of canopy 
biophysical characteristics. Although directional properties of the white fabric exhibits a quasi lambertian 
behavior, errors in soil background reflectance was still possible. When installing the white fabric under the 
canopy it was very difficult to maintain it perfectly flat and horizontal. This could create deviations from the 
nominal value used that was measured when it was perfectly flat and horizontal. In the same way, the contrast 
between soil background and leaf reflectances is maximum for the white fabric. An error in the structural 
variables that were assumed perfectly known will lead to a significant change in canopy reflectance and thus in