Research On Applications Of Bayesian Network Model In The Classification Of Remote Sensing Images

The technology of remote sensing image interpretation (or image classification) is along with the birth of remote sensor technology, and the data from the sensor must be conducted a series of processing and analysis to obtain useful information and knowledge. Comparing to the development speed of remote sensor technology, the method extracting information from remote sensing image is relatively slow much more. Image interpretation is a scientific problem, and the accuracy, reliability and stability of image interpretation cannot meet the actual requirements of production so far. The automation degree of image interpretation is not high, and is still in the semi-automatic level. The research around the image interpretation is very active for the past several decades, and there have been many classification methods among which the supervised method including maximum likelihood method, artificial neural networks, and support vector machines, etc., but there are certain problems or limitations in these methods. The research of some new methods, such as genetic algorithms, artificial immune algorithm, ant colony algorithm, have also made progress, but are still staying in the theoretical research and experimental stage. So we need introduce some new ideas and methods in the field of artificial intelligence into the image interpretation constantly in order to improve the intelligence and automation level of image interpretation.Bayesian network is a new achievements in the study of Bayesian statistical methods in recent years, is the product of combination of graph theory and probability theory, and is a powerful tool for uncertainty reasoning. Bayesian networks are successful in many fields such as medical diagnosis, natural language understanding, fault detection etc, but are relatively few studied in the processing and analysis of remote sensing images The primary intention of this paper includes two aspects. One problem is considering how to apply Bayesian networks into remote sensing image interpretation. Specifically, integrating priori knowledge and information of samples, and building Bayesian network classification models adapting to different types of remote sensing data in view of the characteristics of high-dimensional features and multi-source data of remote sensing data. Another problem is proposing new models based on the comparison and evaluation of existing methods and models.Naive Bayesian network is a classic Bayesian network, is the foundation and core of this paper also. The research of this paper starts around the naive Bayesian network, includes three aspects. To begin with, the improvement of naive Bayesian network--constrained Bayesian network. Then, the extention of naive Bayesian network--hierarchical naive Bayesian network. Finally, Gaussian mixture model and it’s combination with naive Bayesian network--GMM based Unsupervised Classifier, and GMM based Naive Bayesian Classifier. The main content of this paper includes following three aspects:(1) The improvement of naive Bayesian network--constrained Bayesian network. There are strong correlations between adjacent bands of remote sensing data. basing on this characteristic, this paper exploits a Bayesian network with constraints--constrained Bayesian networks, to construct the classifier. That is, feature node is constrained as the child nodes of class node, and the connections between child nodes can be vary. Studys how to construct four classifiers (SNBC, TAN, BAN, BMN) through structure learning from remote sensing data(Landsat ETM+ images and OMIS images), and discuss their performance. TAN, BAN and SNBC are improved form of NBC, BMN is an extended form of BAN.(2) Gaussian mixture model and its combination with naive Bayesian network. The extention of naive Bayesian network--hierarchical naive Bayesian network. Gaussian mixture model is one kind of graph model, is a special case of Bayesian network model. Remote sensing data are usually assumed to obey a single Gaussian distribution, but there are some problems with this assumption in actual application. Due to the complexity and randomness of imaging process in remote sensing, the overall distribution of one group of remote sensing data can be seen as the mixture of multiple sub-Gaussian distributions. Thus, an instructed unsupervised classification methods for remote sensing image based on Gaussian mixture model--GMM-UC, and the model which combines Gaussian mixture model and naive Bayesian network-GMM-NBC model, are proposed in this paper. GMM-UC is based on the theory of finite mixture model, in which the remote sensing data is "mixed" by a finite number of sub Gaussian distributions to a certain percentage. Through the improved EM algorithm, GMM-UC automatically determines the number of sub-Gaussians and their parameters, and then restores all land objects (one sub-Gaussian is corresponding with one class of land object). GMM-NBC embeds Gaussian mixture model (multi-dimensional) into naive Bayesian network, and the extended naive Bayesian networks classifier is built.(3) The extention of naive Bayesian network-hierarchical naive Bayesian network. This paper generalizes the naive Bayesian network, extends the sub-nodes type of naive Bayesian network, that is, the sub-nodes of NBC are not limited in bands or texture features, but the combination of bands or features, or the intermediate classification results of a group features. HNBC model embeds Gaussian mixture model (multi-dimensional) into naive Bayesian network on the basis of high-dimensional features’s organization. The organization of high-dimensional features is the key issue of building HNBC model, which involves the grouping and hierarchy of features. The problems to be resolved indlude:groups features in order to meet the assumption of conditional independence in NBC, and combines of the parameter expression model- GMM with NBC.