267 
- lignin : 306 + 696 £n22 - 1037.^^ + SO.^ 2212 
P 168 O P 168 O P 168 O 
- cell W : 88 + 312 lnp 17 * 22 - 93 Inp 2I * 00 
- nitrogen : 36.4 + 1359 lnp 21 * 80 - 1275 lnp 2230 
r = 0.89 ; F(Fisher)>31 
r = 0.83 ; F(Fisher)>29 
r = 0.97 ; F(Fisher)>385 
The validation of the above mentioned predictive relationships requires the computation of a parameter such as 
the standard error of prediction (SEP), which indicates to what extent the values of "unknowns" can be 
predicted (Card et al., 1988). Unfortunately, the number of samples that were collected in the field was not 
sufficiently large to allow us to have both a reliable prediction set and a reliable calibration set. 
IV. Validation of Predictive Equations with Airborne Measurements 
1.) Direct approach 
AVIRIS images were georeferenced with a local data base (limits of parcels, ages of trees, mean dbh,...). With 
simple crossing of different information maps spectral reflectance means and standard deviations per parcel 
were computed. This was done for the 13 parcels where samples were collected, both for atmospherically 
corrected and non corrected reflectances. For statistical purposes the number of sample parcels should largely 
exceed 13. This small number is undoubtedly a limiting factor. Correlations were computed between raw 
spectral reflectances (p) and reflectance ratios (p/p 16g0 , ar >d P/P 2230 )’ anc * chemical concentrations. 
- Correlations were low for most wavelengths. Surprisingly, best results were obtained for lignin, and not for 
nitrogen, as could have been expected from laboratory measurements. 
- Reflectance ratios did not always yield better results than single reflectances, with the exception of cellulose 
(most wavelengths). In the case of lignin, reflectance ratios systematically yielded worse results. 
- Atmospheric corrections did not significantly change, improve or worsen results. 
Laboratory-derived predictive relationships were applied to AVIRIS-derived canopy reflectances. Correlations 
between the predicted and the actual chemical concentrations of needles were relatively large (nitrogen, 0.74; 
cellulose, 0.79) except for lignin (0.55). Results were not improved when actual concentrations were weighted 
with local biomass density. This is surely due to the fact that reflectance by vegetation covers depends not only 
on the optical properties of individual elements, but also on their quantity, spatial distribution and orientation. 
This partly explains why in the laboratory nitrogen and lignin concentrations in dried and ground needles were 
rather well correlated with spectral reflectances, whereas at the canopy level this was not the case. It seems 
necessary to consider the optical properties of individual needles and not the optical properties of the whole 
cover. Therefore, an attempt has been made to take into account the above mentioned factors. 
2.) Indirect approach 
Two multi-layer reflectance models that explicitly relate canopy reflectance to its architecture and optical 
parameters were inverted: the SAIL model and the JPF model (Zagolski, 1994), which takes into account the 
non lambertian effect of each layer. Here, the forest was considered to be made of 3 layers: an homogeneous 
layer of pine needles, an homogeneous layer of fems and the soil layer. The inversion, i.e. the determination of 
foliar MIR reflectances, relied on the LAI computation with the inversion of the model with VIS reflectances. 
Predictive equations were applied to the reflectance of pine needles obtained through the inversion of the SAIL 
and JPF models. The correlations between these predicted concentrations and the actual concentrations were 
computed. Correlations were definitely better for lignin (74%) and slightly poorer for nitrogen (70%) and 
cellulose (69%) than those obtained with predicted concentrations that were derived from canopy reflectances. 
It can be assumed that results can be improved if canopy structure effects on reflectances are better taken into 
account. This apparoach is being investigated by the authors (Pinel, 1993) with the development of a 
geometric reflectance model of a forest cover. Another major source of improvement is linked to the 
correction of atmospheric effects, both in the MIR region, for the direct use of spectral bands within the 
predictive equations, and in the VS/NIR region for the estimate of the LAI. Indeed, errors on the LAI estimates 
lead to errors on the foliar reflectance estimates. This is all the more important that the aerosol distribution is 
heterogenous throughout the image. As an example, a common 10% relative error on the VIS reflectance of 
dense vegetation, due to atmospheric effects, may result in a LAI relative error larger than 10%, which may