Low level images representation in feature space performs poorly for classification with high accuracy since this level of representation is not able to project images into the discriminative feature space. In this work, we propose an efficient image representation model for classification. First we apply Hidden Markov Model (HMM) on ordered grids represented by different type of image descriptors in order to include causality of local properties existing in image for feature extraction and then we train up a separate classifier for each of these features sets. Finally we ensemble these classifiers efficiently in a way that they can cancel out each other errors for obtaining higher accuracy. This method is evaluated on 15 natural scene dataset. Experimental results show the superiority of the proposed method in comparison to some current existing methods