Résumé
Ecosystems classification is the process of allocating vegetation types into groups so that individuals in the same class are similar according to their physiological and phonological characteristics to another one. Over large areas, the only suitable technique to obtain frequent and repetitive data acquisitions over such large areas is the use of observations recorded by sensors of moderate resolution. In order to minimize the role of the analyst and to improve the accuracy of the results, innovative and efficient approaches for the classification of ecosystems continue to appear in the literature. This research developed and implemented a new hybrid unsupervised classification approach to derive ecosystems using multi-annual time series by combining hierarchical and partitioning clustering principles. The latter approach is applied on 8-years time series (2000-2007) of 10-day composite Normalized Difference Vegetation Index (NDVI) recorded by SPOT/VEGETATION. After the first segmentation of the mainland in ecoregions using the Fast Fourier Transform (FFT), successive k-nearest neighbor (k-NN) clustering enhance the discrimination of ecosystems and yields to the production of a new ecosystem map for the African continent. The nomenclature relied on the Land Cover Classification System (LCCS) of the Food and Agricultural Organization (FAO). On the basis of validated continental, a pixel-by-pixel analysis is conducted to assess the accuracy of the new classification. The hybrid clustering facilitates the identification/labeling process and the obtained results which should provide key information needed for management/monitoring of natural resources, biodiversity conservation and biogeochemical studies may also deserve vegetation cover modeling at regional and local scale