A method to estimate the pure reflectance of landscape components in mixed pixels is presented in this paper. Based on the usual linear mixture model, this original method allowed us to restitute the temporal and spatial variability of the reflectance of the components of a NOAA-AVHRR scene. Ground cover proportions were obtained from a SPOT image classification degraded at NOAA resolution. The signal deconvolution was made using blocks of pixels to limit the errors due to the misregistration between NOAA and the ground cover images. The model was applied to NOAA-AVHRR simulations and to actual NOAA temporal series (1992) over the HAPEX-Sahel degree square, to deconvoluate the reflectances of the three landscape components (millet, fallow and plateau). The data analysis indicated a high sensitivity of the method to the misregistration and the need to define an optimal size of pixels which provide a good compromise between the misregistration errors and the spatial variability of the reflectance components. The deconvoluated reflectances were compared to airborne measurements acquired over two sites during the HAPEX-Sahel experiment. The quality of the results depends on the type of landscape. The model performs best on a landscape with small surface units, well spatially distributed.