LAUSR

Typical unsupervised classification of hyperspectral imagery (HSI) uses a Gaussian mixture model to determine intensity similarity of pixels. However, the existence of mixed pixels in HSI tends to reduce the effectiveness of the similarity measure and leads to large classification errors. Since a semantic class is always dominated by a particular endmember, a mixed pixel can be better classified by identifying the dominant endmember. By exploiting the spectral mixture model (SMM) that describes the endmember-abundance pattern of mixed pixels, the discriminative ability of HSI can be enhanced. A Bayesian classification approach is presented for spatial–spectral HSI classification, where the data likelihood is built upon the SMM, and the label prior is based on a Markov random field (MRF). The new approach has three key characteristics. First, instead of using intensity similarity, the new approach uses the abundance-endmember pattern of each pixel and classifies a pixel by its dominant endmember. Second, to integrate the SMM into a Bayesian framework, a data likelihood is designed based on the SMM to reflect the influence of the dominant endmember on the conditional distribution of the mixed pixel given the class label. Third, the resulting maximum a posteriori problem is solved by the expectation–maximization (EM) algorithm, in which the E-step adopts a graph-cut approach to estimate the class labels, and the M-step adopts a purified-means approach to estimate the endmembers. Experiments on both simulated and real HSIs demonstrate that the proposed method can exploit the spatial–spectral information of HSI to achieve high accuracy in unsupervised classification of HSI.