| The expectation-maximization (EM) algorithm is widely used for parameter estimation for various pattern recognition applications. However, the convergence of the EM algorithm depends on the initialization of the algorithm. To avoid this problem and to guarantee the convergence of the algorithm, we propose herein an enhanced version of the EM algorithm by introducing a preprocessing step. We utilized an exponential function based Euclidian distance measure to rapidly separate the data cube into potential objects of interest class and background class. Then, the EM algorithm is employed to estimate the parameters of the two classes in order to detect the class of targets. The proposed EM algorithm needs a postprocessing step to classify the potential objects of interest class into the objects of interest class and the background class corresponding to posterior probabilities with user defined threshold. To avoid manual selection of the threshold, we employed stochastic version of the EM (SEM) algorithm, which does not depend on a threshold because the algorithm classifies the potential objects of interest class further into target and background classes. The proposed algorithm was tested using real life hyperspectral image datasets, and the results show low false alarm rate even with challenging scenarios. |
