This paper describes a new algorithm for hyperspectral image unmixing. Mostof the unmixing algorithms proposed in the literature do not take into accountthe possible spatial correlations between the pixels. In this work, a Bayesianmodel is introduced to exploit these correlations. The image to be unmixed isassumed to be partitioned into regions (or classes) where the statisticalproperties of the abundance coefficients are homogeneous. A Markov random fieldis then proposed to model the spatial dependency of the pixels within anyclass. Conditionally upon a given class, each pixel is modeled by using theclassical linear mixing model with additive white Gaussian noise. This strategyis investigated the well known linear mixing model. For this model, theposterior distributions of the unknown parameters and hyperparameters allowones to infer the parameters of interest. These parameters include theabundances for each pixel, the means and variances of the abundances for eachclass, as well as a classification map indicating the classes of all pixels inthe image. To overcome the complexity of the posterior distribution ofinterest, we consider Markov chain Monte Carlo methods that generate samplesdistributed according to the posterior of interest. The generated samples arethen used for parameter and hyperparameter estimation. The accuracy of theproposed algorithms is illustrated on synthetic and real data.
展开▼
