96 
3. Results 
3. J. INTUITIV time-serie smoothing 
On figure 2 are presented examples of NDVI time-series obtained using INTUITIV method over Africa area. We 
can observe that they are in good agreement with the theoretical expected profiles based on present knowledge 
concerning these areas. For instance, in the Nile Delta, a double annual cycle is observed which undoubtedly 
corresponds to the two marked seasons of crop production. Because determined cloud occurrence is low in this 
area, the length of the sliding-period is short and the observed NDVI decrease between the two maximum is 
retained. On contrary, when considering rain forest pixel example, the decrease occurring in the middle of the 
year on the non-filtered profile is suppressed by the INTUITIV method. Obtained profile is characterized by high 
and relatively constant NDVI values which is the response expected for a dense evergreen tree canopy. 
Concomitant to these results, we must note that the entire satellite data processing (calibration, atmospheric 
correction, cloud elimination, INTUITIV method) leads to NDVI values varying between 0.8 and 0.9 for closed 
canopies whereas NDVI issued from raw data are never greater than 0.45. In fact, Verhoef (1984) had estimated, 
using a canopy reflectance model (SAIL), that maximum NDVI values that one can actually observe over a 
canopy never exceed 0.9. This fact tends to quantitatively confirm improvements brought by the applied entire 
data processing. 
If we consider NDVI profile such as the one of sudanian savanna, we can see that a start, an end and a length of 
cycle can be determined. They are in good agreement with vegetation cycle evolution usually encountered in this 
area. This qualitative result is confirmed by the figure 3 where lengths of NDVI cycles are reported for West 
Africa. In this region, air mass fluxes and weak reliefs lead to mean annual rainfall that increase from north to 
south with isohyets distribution in longitudinal stripes. In regions where annual rainfall is lower than 1000 
mm/year, vegetation dynamics largely reflect the spatial and time distribution of rainfall. We estimated lengths 
of "vegetation cycles" that regularly increase from North to South, varying from 1 to 4 months for sahelian 
vegetation, 4 to 7 months for sudanian vegetation, 7 to 11 months for guinean vegetation and finally 12 months 
for dense evergreen forests. 
When considering signal amplitudes, results are in good agreement with what is expected. Maximum NDVI 
observed along temporal profiles increase from North to South. In the North of West Africa, vegetation 
formations are rarely closed when reaching maximum development stage. Maximum NDVI obtained for these 
areas never exceed 0.7. Next, the more we move to the south, the more temporary or permanently closed 
vegetated formations are encountered and the more saturating NDVI values are similarly detected along 
temporal profiles. 
4. Conclusion 
In our study, we tried to take into account present knowledge concerning different perturbation factors affecting 
red and near infrared satellite measurements and concerning ground-based observations of vegetation phenology. 
With regards to the results, we can conclude that the INTUITIV method in addition to others data correction 
steps really improve satellite data quality. The proposed data processing method seems to be a good alternative 
solution before development of specific models to discard influencing factors such as bidirectional effects. 
INTUITIV method is shown to provide better results than the MVC or BISE method do. Nevertheless, due to 
cloud threshold characteristics, INTUITIV method is only efficient for satellite data concerning tropical and 
sub-tropical regions. In fact, in cold or temperate regions, temperature at ground surface can frequently be equal 
or lower than the one observed for some cloud surfaces. So it is necessary to find another kind of filtering 
process for these areas. 
Another insufficiency of the data processing is that some steps such as : calibration drift correction, NDVI 
profile smoothing using the "adaptative sliding-period" and cloud occurrence mapping, forbid data correction in 
real time. Consequently, satellite data utilization for real time applications or models is prevented. 
Despite these annoyances, the proposed method allows the development of different studies : 
- interannual comparisons of results derived from NDVI values. This fact is the consequence of temporal drift 
correction step ; 
- vegetation classification . Due to reduction of data error sources, it is expected that filtered data better 
discriminate and reflect radiometric signal of entities located at ground surface. Thusly, Gond et al. (1993) have 
produced an African vegetation map derived from automatic classification of filtered NDVI profiles. This map, 
when compared to the AETFAT vegetation map (White, 1983) and to ground-based observations showed 
substantial better agreement that the one obtained with raw NDVI data ; 
- generation of model inputs ; 
Filtered and smoothed NDVI profiles allow to mathematically determine informations such as : the beginning of 
the increase or the decrease of the signal, the maximum and the amplitude of NDVI time-series. Assuming that 
temporal NDVI profile truly corresponds to phenological cycle, values like the beginning, the end, the length of 
the vegetation cycle can be estimated for each pixel. These informations are of first interest as inputs for crop 
production or net primary productivity models. These values usually are only available for few sites by operating