: NOAA as 
late-carree" 
■veil, 1990). 
the thermal 
scan angle 
in continent 
profile of a 
At present time, no specific models can correct huge satellite data set from these phenomena. INTUITIV 
method is then proposed as an alternative procedure to reduce their influences. 
2.3.1. Cloud elimination. 
To filter cloud contaminated pixels, a temperature threshold technique is performed. For this purpose, 
NOAA/AVHRR channels 4 and 5 are used. For each pixel and each date, when radiative temperature is lower 
than 15° Celsius (GOES equivalent temperature), corresponding NDVI value is not considered. This limit was 
chosen assuming that, in tropical and sub-tropical areas, surface temperatures lower than 15°C only occur at 
cloud level. 
effects like 
However, 
iges in gain 
3 % for low 
lower than 
want NDVI 
calibration 
A threshold on red and near infrared reflectances is also applied. Most clouds highly reflect incident radiations in 
these two wavelengths whereas vegetation has very low reflectance in red channel. So, each pixel characterized 
by a red reflectance value higher than 0.3 and near infrared reflectance value higher than 0.5 is also excluded 
from the process. Then, each excluded NDVI data is replaced through a linear interpolation of non-excluded data 
surrounding it on temporal profile. 
It is arguable that for desert area, red and near infrared reflectances are also high. So, when reflectances remain 
high all year long in these two channels, pixel is classified as desert and the corresponding NDVI time-serie is 
not modified. 
We must notice that despite the cloud screening stage, some perturbations persist in time-series. One reason is 
bidirectionnal effects. Second reason is that semi-transparent clouds like cirrus or partial pixel cloud coverage 
can be easily missed by threshold methods while they affect the signal (Goward and Hope, 1989). 
earth to sun 
ances were 
2.3.2. Others effects. 
Bidirectional effects are extremely complex and, presently, models are in preliminary stage of development. 
However, some authors proposed satellite data filtering methods without a priori knowledge on influencing 
factors (Dijk et al ., 1987 ; Viovy, 1990). These procedures consist in NDVI profiles smoothing using 
: to changes 
;ar infrared 
al., 1990 ; 
ad target of 
a. We used 
alibrate our 
mathematical methods (mean of several values, Fourier transform) to eliminate high NDVI fluctuations. In our 
process, we used the BISE method (Viovy et al ., 1992) which was shown to retain more valuable elements of an 
NDVI profile than the MVC one. This method is based on three hypotheses : (1) cloudy sky and turbid 
atmosphere decrease NDVI values ; (2) data transmission errors can generate abnormally high NDVI increases 
or decreases ; (3) NDVI increases or decreases induced by vegetation changes, even if they can occur rapidly, 
always are followed by a stable period of several days or weeks. 
From the first value of a NDVI time-serie, the BISE algorithm retains the following value if it is higher than the 
; on red and 
Method for 
)r our data 
zone, water 
f data with 
lances data 
ita enclosed 
lion of 11 
en gathered 
ean values. 
al. (1976). 
al thickness 
ponds to an 
ler estimate 
reflectance 
;u, personal 
previous one. However, if NDVI increase is higher than 0.1, this value is excluded assuming that such an event 
only can be due to data error. For decreasing events occurring in the signal, the authors defined a period for 
which it is assumed that a regrowth of vegetation after a senescence is impossible within this time. This period 
was called "sliding period" by the authors. 
So, if a value is lower than the previous one, the BISE algorithm tests if there is no point within the 
"sliding-period", with a value 20 % higher than the difference between this low value and the previous high 
value. This threshold was set empirically by the authors assuming that two NDVI values are significantly 
different when they differ of more than 20 %. If no higher value is encountered, they consider that the decrease 
is not an artifact and the low value is retained. On the contrary, if a higher value exists, the algorithm shifts to 
this point and all intermediate data between this point and the previous retained high value are excluded 
When a value is retained, different tests, as previously described, are applied to the following point and so on to 
the end of the time-serie. In this algorithm too, non-accepted values are replaced by values issued from linear 
interpolation of surrounding retained points. 
In the BISE method, the length of the "sliding-period" was constant and fixed to 30 days. This value was 
determined empirically, observing improvement obtained on savanna NDVI profiles testing different lengths of 
"sliding-period". However, for specific regions such as rain forests, clouds may be present for more than one 
month and the selected length of "sliding-period" will not be sufficient to remove this source of noise. 
In the INTUITIV method, we propose an improvement of the above algorithm by creating an "adaptable length 
of sliding-period". The basic hypothesis of our process is to assume that the higher is the cloud occurrence for a 
considered area (i.e. one pixel) the longer must be the sliding-period. Practically, the first step of procedure 
consists in deriving a map of annual cloud occurrence index from thermal infrared, red and visible data. For a 
iration and 
of the earth 
rain forest, 
ng on scan 
; scattering 
increase of 
considered pixel, cloud occurrence index is determined by dividing the number of data rejected from the 
threshold method described for cloud elimination (§ 2.3.1.), by the total number of data in the time-serie (i.e. 52 
weeks). Then, the length of thè sliding-period is linearly related to the cloud occurrence value. The slope and 
intercept of the linear relationship were empirically fixed by observing NDVI temporal evolutions obtained for 
different relationships and for different kinds of locations (desert, savanna, rain forest). From these tests, limits 
were set in such a way that the length of the sliding period is 4 weeks when cloud occurrence index is 0 and 15 
weeks when cloud occurrence is 0.5. Profile resulting from INTUITIV filter is presented on figure 1 
95