Pattern recognition in multiband imagery using stochastic expectation maximization

Hyperspectral sensors can facilitate automatic pattern recognition in cluttered imagery since man made objects often differ considerably from the natural background in absorbing and reflecting the radiation at various wavelengths i.e., the identification of the objects is based on spectral signature of the objects in the scene. In this paper, a unified approach for pattern recognition with known object signature is formulated by generating Gaussian mixture model to effectively utilize the underlying statistics of the data cube. To estimate the model parameters, enhanced version of the stochastic expectation maximization (SEM) algorithm is employed, which is also used successfully for image classification by reducing the unwanted information in the data cube. In the proposed scheme, at first we used the modified SEM to identify the different classes in the scene including the desired object class. Then, the Mahalanobis distance between the desired object signature and distributions of the mixture model is employed to detect the object class. Finally, the maximum a posteriori (MAP) probability for each pixel is estimated and Bayesian decision law is applied in order to isolate object pixels. The proposed algorithm has been tested using real life hyperspectral imagery and the results show that the algorithm shows robust performance in noisy environment.