Shape-texture features for the VHSR satellite images classification using the MLP neural net.

With the emergence of commercial satellites with on-board sensors characterized by Very High Spatial Resolution (VHSR), identification and localization of topographic features are becoming conceivable. The VHSR sensors provide images with significant amount of geometrical details that yield new kinds of information, such as shape information. On the other hand, the classification of this heterogeneous information set requires also the establishment of new techniques different from those used for low and medium spatial resolution data classification. To this end, the Multilayer Perceptron (MLP) neural network has been investigated to classify this heterogeneous set of information using a combination of different features extracted from objects, namely Quaternion Zernike moments as shape descriptor and Haralick's features as textural descriptor. The proposed approach was tested using a sub-scene of Quickbird image datasets of Algiers (northern Algeria). The results indicate a mean accuracy value of 79.99 percent using only the shape feature information, 74.89 percent by applying the textural features information and 86.23 percent by combination of shape descriptor and texture descriptor. The results of the proposed method with MLP classifier are also compared with those results obtained by k-NN and SVM classifiers, a fact which confirms the effectiveness of the suggested approach.