Traditional image processing techniques have been proven to be inadequate for urban land-cover mapping using very high resolution (VHR) remotely sensed imagery. Abundant features such as texture, shape, and structural information can be extracted from high-resolution images, which make it possible to distinguish land covers more effectively. However, the multisource characteristics of VHR images place significant demands on the classification method in terms of both efficiency and effectiveness. The most often used method is vector stacking fusion, in which a single classifier is trained over the whole feature space; statistical differences and separability complementarities among different features are rarely considered. Hence, appropriate feature fusion and classification of multisource features become the key issues in the field of urban land-cover mapping. A novel decision fusion method based on a Bayesian network is proposed to handle the multisource features of VHR images which provide redundant or complementary results. Subclassifiers are constructed separately based on multiple feature sets and then embedded into the naive Bayesian network classifier (NBC). The final results are obtained by fusing all the subclassifiers into the NBC framework. Experiments on aerial and QuickBird images demonstrated that the performance of the proposed method is greatly improved compared with vector stacking methods, and significantly improved compared with the multiple-classifier systems and multiple kernels learning support vector machine. Moreover, the proposed method has advantages in feature fusion of VHR images in urban land-cover mapping.
In the traditional BNC model, the relationship between the attributes are the same for all the instances of the class
variable C. BMN classifier is a generalized form of BNC, in the sense that it allows different relationships among
attributes for every values of the class variable, and provides a unique net structure for every object class. This paper
proposes Bayesian Multi-nets (BMN) Models based on the analysis of conditional mutual information(CMI) between
image features of different objects classes, and constructs BMN classifier for remote sensing images on the basis of
experiment. Classification accuracy of single objects in BMN classifier outperforms that of traditional BN, proves the
latent value of the proposed models in the classification of remote sensing images.
SVM (Support Vector Machine) is a new kind of machine learning method , it can solve classification and regression
problems very successfully and accomplish classification with small sample incident perfectly. In this paper, the NPA is
proposed to compute the optimization problem to achieve the classification for hyperspectral remote sensing (RS) image
by "1 VS m" strategy and radial basis kernel function. Besides, a new method, the dual-binary tree + SVM algorithm is
proposed, to solve the mutil-class, high-dimensional(HD) problems of hyperspectral RS image. In the end, the test is
carried on the OMIS image. The comparative results of this algorithm with other methods are given, which shows that
our algorithm is very competitive, particularly for the small samples and non-equilibrium surface features. Both the
accuracy and speed of classification are improved greatly.