BAYESIAN SUBSPACE ESTIMATION USING SPARSE PROMOTING PRIOR

Clément Elvira,P. Chainais,N. Dobigeon

Abstract:Hyperspectral sensors record the light intensity beyond the visible spectra in hundreds of narrow contiguous bands. Images are characterized by a high spectral resolution but a low spatial precision due to sensors constraints. A crucial step called unmixing consists of decomposing each pixel as a combination of pure spectra, called endmembers. Endmembers act as fingerprints, improving the ability to analyse a scene. Under reasonable assumptions, pixels are expected to live in a lower dimensional subspace whose dimension is intimately linked to the number of endmembers. The identification of this subspace yields gains in computational time, complexity and in data storage. However, determining both the relevant subspace dimension (e.g. the number of endmembers) and a suitable representation is a difficult problem. Existing methods are often parametric, such as thresholding the eigenvalues of Principal Component Analysis (PCA), and using eigenvectors as a subspace base (see [1] for a review). My Ph.D. aims at exploring Bayesian nonparametric inference to tackle these tasks. We have already proposed a Bayesian formulation of anti-sparse coding [2, 3], primary motivated by the search of representations for endmembers. Anti-sparse representations aim at spreading the energy over all components uniformly. In Data analysis, many methods exist to extract the underlying subspace. Model selection methods include criteria that quantify compromises between reconstruction and complexity (e.g. AIC or BIC). PCA also implicitly permits a dimension reduction by projecting observations onto a subset of orthonormal vectors. A probabilistic formulation of PCA has been proposed through factor analysis [4]. Existing extensions rely for instance on Laplace [5] or variational [6] approximations of the posterior distribution. In this work, we investigate the use of Bayesian nonparametric inference associated to directional statistics to explore the set of subspaces. The goal is to avoid an arbitrary thresholding of eigenvalues as often done for PCA. We derive an empirical posterior distribution of bases of the latent subspace, where coefficients, e.g. projections, have been marginalized out. We proposed to use MCMC methods to sample according to this posterior and approximate estima-