squared correlation coefficient, R 2 , larger than 0.8 for 50% of the area and smaller than 0.5 for about 20% of it 
at 550 and 850 nm. The results are somewhat less good at 650 nm, may be because the level of reflectance is 
lower at this wavelength and therefore more sensitive to noise. A map of the correlation coefficient (not shown) 
allows us to identify the pixels for which the directional model is not satisfactory (i.e. when the correlation is 
lower than 0.5). Those correspond to i) river and canal, ii) rice paddies, and iii) pixels with few measurements 
(NE and SW comers of the study area). River, canals and rice paddies show a large specular reflection signal 
that is not accounted for in the functions fi and f2. When the specular direction is disregarded, the model can 
reproduce the directional signature of the rice as it does for other canopies. On the other hand, the reflectance of 
water is mostly specular and its value in other directions is hardly distinguishable from noise. The pixels located 
in the NE and SW comers of the study area are relatively far from all flight lines. Thus, few observations 
correspond to these points and they are all acquired for large viewing angles. The measurements show little 
variation hardly distinguishable from noise, which yields a small correlation. 
The results presented in this section show our ability to derive from POLDER directional measurements a 
reflectance corrected for angular effects. This correction is of uppermost importance for a comparison of 
measurements acquired from different instruments with various geometry. Similarly, a “normalized” reflectance 
is useful if we want to derive empirical relationships that yield physical surface parameters from the 
reflectances. We acknowledge that we have not shown that our corrected reflectance is better than a simple 
averaging of the measurements. Although the procedure is reasonable, it may introduce additional noise which 
could actually deteriorate the data quality. We must therefore imagine a procedure to compare the usefulness of 
our corrected reflectance ko, with either a reflectance derived from the same set of data but with a simpler 
procedure (simple averaging of all directional measurements, for example), or the reflectance that would be 
acquired with a traditional scanner (only one directional measurement). The difficulty for such a validation is 
the lack of direct validation dataset: we do not have surface reflectance measurements for the normalized 
geometry, especially at the spatial scale used here. Thus, indirect methods are needed. 
One method to validate our normalisation procedure is to compare the ko obtained for a given surface target 
from various datasets of directional measurements. Our data is well suited for this procedure since we can 
compare, for instance, the outcome of various flight line measurements. We can also compare the kO obtained 
from all measurements in the forward scattering hemisphere, with those from the backscattering hemisphere. 
Another validation procedure could make use of the land use map which was elaborated by INRA, from 
surface survey, during the campaign. One can determine, for each of the surface type classes, the variability 
(standard deviation or another statistical indicator) of the kq, and that of the reflectance derived with other 
procedures. A small variability intra-class when compared to the distance between each class, indicates that 
surface types can be discriminated from the parameter. A smaller variability intra class indicates a better 
parameter, and the method which derives the ko can be quantitatively compared to other normalisation 
procedures. 
Finally, we can also make use of POLDER measurements acquired, over the same area, during other days. 
Similar flight plans have been used and the processing described above can be repeated. The spectral 
normalised reflectance and angular signatures obtained from the various flight measurements can be compared 
since we can assume little variability of the surface during the campaign time period. This validation procedure 
evaluates, in particular, the atmospheric correction since the various flights have been performed with various 
optical thicknesses. 
These validation procedures will be performed in the near future and their results will be presented in a 
forthcoming paper. 
5 DISCUSSION AND CONCLUSION 
The study presented in this paper shows that meaningful reflectance directional signatures are accessible from 
airborne POLDER measurements. A key point is the confidence we have in the measurements. Our data 
accuracy is limited mostly by two factors: i) spatial registration of the measurements and ii) atmospheric effects 
on the reflectance. 
Many effects contribute to the misregistration of our measurements, including the GPS uncertainty, a bad 
timing between the GPS and the measurements, and an uncertainty on the radiometer optical axis. A 
misregistration induces discontinuities in the derived reflectance angular signature that can easily be identified. 
This effect is particularly troublesome for very heterogeneous (at the measurement scale) landscapes such as 
those in the northern part of our study area. The quality of the surface targets directional signatures which we 
studied in detail, and the high correlation between the model and the measurements shows that the data are 
relatively well registered. We believe misregistration is responsible for the lower correlation found in the 
northern part of our study area. 
The measurements we used for the retrieval of surface directional signature have been selected for their 
relatively low atmospheric optical thickness. Atmospheric measurements have been acquired at the same time 
ns POLDER flight and we have therefore a high confidence in our atmospheric correction. Uncertainty remains 
since aerosol optical thickness measurements are sensitive to the total optical thickness whereas atmospheric 
correction needs that below the aircraft. Besides, atmospheric effects induce a directional signature that is 
705