In this thesis, we focused on hyperspectral images processing and more precisely, about their limited spatial resolution. This limitation affects directly the classification process of this kind of images due to the fact that is more difficult to estimate the proportion and the area occupied by a pure material in an observed pixel. To overcome this limitation, two approaches are commonly used. The first one is known as the spectral unmixing, which consists in unmixing an observed pixel in order to estimate the pure material spectra and their associated abundances. The spectral unmixing is based on methods developed in the fields of Blind Source Separation. The second approach consists in fusing a hyperspectral image with a multispectral or ...