Bayesian network classifiers (BNCs) are now among the most used supervised probabilistic methods for remote sensing image classification. Our contribution lies in two principal points. First, the investigation of the applicability of Kruskal’s algorithm constructs the optimal tree structure of multinet Bayeisan network classifier (MBNC). Second, the focus on MBNC’s advantages is over other classical BNCs, such as naive Bayes classifier (NBC), tree augmented naive Bayes classifier (TANC), forest augmented naive Bayes classifier (FANC), and state-of-the art competitor classifiers such as maximum likelihood classifier (MLC) and support vector machine (SVM) classifier. While classical BNCs have a mere network for all predefined classes, MBNC has as many local Bayesian networks as the predefined classes. Hence, through a statistical evaluation and a visual inspection, our objective is to emphasize the contribution of MBNC to enhance the accuracy of urban land cover map obtained by classification of remotely sensed image. Performances of developed BNCs (NBC, TANC, FANC, and MBNC) are experimentally assessed using a multispectral image acquired on July 12, 2010, by Alsat-2A Algerian satellite. Based on a confusion matrix, overall accuracy, and Kappa statistic, results indicate that MBNC largely outperforms classical BNCs (NBC, TANC, and FANC) and probabilistic MLC, but performs slightly better than an SVM classifier. Due to its specific-class local network, MBNC gives a powerful tool for a better discrimination between different correlated spectral classes.
The aim of this work is to evaluate the performances of three Bayesian networks widely used for supervised image classification. The developed structures are constructed due to Kruskal algorithm which allows the determination of the maximum weight spanning tree by using the mutual information between the attributes. We started by the Bayesian naïve classifier (BNC), which assumes that there is no dependency, between the attributes to classify. In order to relax this strong assumption, we tested the tree augmented naïve Bayes classifier (TANC) where each feature has at most one variable as parent, and the forest augmented naïve Bayes classifier (FANC) where each attribute forms an arbitrary graph rather than just a tree. These classifiers are evaluated using a multispectral image and hyperspectral image in order to analyze the structure classifier complexity according to the number of attributes (04 and 10 spectral bands for the two images respectively). Obtained results are compared with state-of-the art competitor, namely, the SVM classifier. Classified images by TANC and FANC achieved higher accuracies than other classifiers including SVM. It is concluded that the choice of attributes dependencies significantly contributes to the discrimination of subjects on the ground. Thus, Bayesian networks appear as powerful tool for multispectral and hyperspectral image classification.