In this work, we developed Blind Source Separation methods (BSS) for hyperspectral images, concerning two applications : urban remote sensing and astrophysics. The first part of this work concerned spectral unmixing for urban images, with the aim of finding, by an unsupervised method, the materials present in the scene, by extracting their spectra and their proportions. Most existing methods rely on a linear model, which is not valid in urban environments because of 3D structures. Therefore, the first step was to derive a mixing model adapted to urban environments, starting from physical equations based on radiative transfer theory. The derived linear-quadratic model, and possible hypotheses on the mixing coefficients, are justified by results obtained with simulated realistic images. We then proposed, for the unmixing, BSS methods based on NMF (Non-negative Matrix Factorization). These methods are based on gradient computation taking into account the quadratic terms. The first method uses a gradient descent algorithm with a constant step, from which we then derived a Newton version. The last proposed method is a multiplicative NMF algorithm. These methods give better performance than a linear method from the literature. Concerning astrophysics, we developed BSS methods for dense field images of the MUSE instrument. Due to the PSF (Point Spread Function) effect, information contained in the pixels can result from contributions of many stars. Hence, there is a need for BSS, to extract from these signals that are mixtures, the star spectra which are our "sources". The mixing model is linear but spectrally non-invariant. We proposed a BSS method based on positivity. This approach uses the parametric model of MUSE FSF (Field Spread Function). The implemented method is iterative and alternates spectra estimation using least squares (with positivity constraint) and FSF parameter estimation by a projected gradient descent algorithm. The proposed method yields good performance with simulated MUSE images.
