24 
note that decreases in NDV1 relating to changes in vegetation status, whilst possibly sudden, will persist for a 
number of days, during which re-growth will be relatively slow. From a given date of the time series, the 
authors search forward and accept the following date, within the next 30 days, if the variations of NDVI 
between both dates lie in an acceptable range. Dijk et al. (1987) reduce the radiometric disturbances in the 
GVI data by smoothing the vegetation index time profiles by statistical filters disregarding the cause of the 
disturbances. Gutman (1991) develops an empirical viewing angle correction of AVHRR data, which is valid, 
however, only for a given range of Sun angles and for a particular surface type (Kansas prairie). Paltridge and 
Mitchell (1990) operate rough atmospheric, Sun and view angle corrections on a set of AVHRR data on an 
other particular surface type (grasslands at Victoria, Australia). Kimes and Deering (1992) propose to 
reconstruct hemispherical reflectance (albedo) using multiple off-nadir view angles measurements taken in a 
single azimuth plane ('strings of data'), such as those from the AVHRR, and a knowledge-based expert system 
which extrapolates the measurements to all uncovered azimuth and view angles. 
At the exception of the work of Paltridge and Mitchell (1990), none of the mentioned methods 
corrects the data simultaneously from Sun and view angles. Consequently, the reflectances obtained by the 
compositing procedure can not easily be compared in space and time, and this limits their potential use. 
3 - APPLICATION OF A SURFACE BIDIRECTIONAL REFLECTANCE MODEL ON AN 
ATMOSPHERICALLY CORRECTED DATA SET 
We describe here first the principle of methods based on the application of surface bidirectional reflectance 
models on data sets of atmospherically corrected reflectances, and then briefly review some recent work 
applying such methods, for vegetation monitoring on the one hand, and surface albedo extraction on the other 
hand. A short discussion on atmospheric corrections follows. 
3.1 - Principle 
Unlike most methods described in § 2, there is here an attempt to separate on a physical basis the respective 
contributions of the atmosphere and of the surface in the observed reflectance. This implies that instrumental 
calibration should be of high accuracy and that atmospheric corrections should be performed as accurately as 
possible. Accurate and fast atmospheric correction algorithms exist, such as the 5S (Tanr6 et al., 1990) or the 
SMAC codes (Rahman and Dedieu, 1994). The accuracy of atmospheric corrections depends on that of input 
data descriptive of the atmosphere, such as the aerosols, water vapor, and ozone column densities, and aerosol 
phase functions, rather than on the accuracy of the algorithms themselves (see § 3.4). 
Assume that a series of atmospherically corrected reflectances {R;, i = 1, n)has been acquired, in a 
given spectral band, during a given period called period of composition (say, 1 month), each measurement 
being associated with sun and view zenith angles 8 J( and 9 VI - , and azimuth angles <(>,• between Sun and view 
directions. A least-squares regression is then applied between the (R^, i = 1, n) and a series of values {R^j, i 
= 1, n), where R^ (0 f ,0 v ,<5)is an analytical model of bidirectional reflectance of the surface, which reads 
^mod (9» >®v > < t ) ) = /(9f >9y >§'^0 >^1 (1) 
and where 
•^modi = ^mod (®ji >®vi > < t ) i )■ (2) 
In Eq. (1), f is an analytical function of the geometric angles and of a number of parameters kQ,k\,...,k m . The 
parameters {kj , j = 0, m}, are retrieved as a result of the regression between model and observations. The 
process may be repeated by displacing the period of composition by a certain amount, which may be called the 
sampling period of the compositing process (say, 15 days for example). As a result the reflectance time series 
is replaced by time series of the parameters kj , with a degraded temporal resolution equal to the sampling 
period. The sampling period can not be taken much smaller than the period of composition, since the 
regression in the period of composition acts more or less as a low pass filter which removes high frequency 
temporal fluctuations of reflectance. Not much information is available at the sampling frequency when the 
sampling period is chosen too small. 
In turn, tii 
according to Eqs. 
residue 5, given by 
5 
is sufficiently smal 
parameters are usee 
R 
The corrected refle 
and sun directions 
direct albedo (refle 
diffuse albedo (dire 
32 - Vegetation m 
This technique has 
1989, on 7 test sites 
Leroy and Roujean 
equal to 30 days an 
Holben et al. (1990 
(comparison of rai 
predetermined valu< 
optical depth is der 
the test sites, and th 
from a meteorologi 
provides on a 30 kn 
The choser 
Roujean et al. (1992 
where fj and f 2 are j 
has been designed si 
surface reflectance 
reflectance mention 
Lambertian charact 
representations of t 
canopies and bare s 
dependence of the s 
made to reduce the 
parameters (ko, ki, 1 
The results 
sites in the visible ai 
(ii) modelled surfac 
modelled reflectanci 
the observed surface 
due to directional ef 
appropriate to descr 
aspect and filter out 
other sites is that 
especially when the 
since ko (a surface re