Graph-Based Hyperspectral Image Classification Using Outliers Detection Based on Spatial Information and Estimating of the Number of GMM Mixtures

In this article is tried to improve the hyper spectral image classification by expectation maximization (EM) algorithm with proposed approaches for estimate the number of the mixture for image classes, covariance matrix correction and outlier detection. by decrease the number of the mixtures of an hyper spectral image, the time of the algorithm has decreased and by covariance matrix correction the accuracy and the validation of the classification could increased for the classes that have a few training sample. Because the EM algorithm is an iterative algorithm, if in one of the step occur an error in classification, this error inter in the next steps and decrease the accuracy and the validation of the classification. In this article this problem has addressed by outliers identification in each step and remove them for parameter estimation in the next step. Therefore, it is prevented from error propagation and increases the accuracy and the credit of the classification.