The Design And Implementation Of Remote Sensing(RS) Image Classification System Based On Artificial Neural Networks(ANN)

Remote sensing image classification is always the important content of remote sensing research field. How to solve multi-class recognition of remote sensing images under the condition of ensuring higher classification accuracy is a key problem for the remote sensing images research. The development of neural network technology provides a new method to solve this problem. The neural network has an ability to learn, robustness and without providing a hypothesis for probability model, as a result of that, the neural network is suitable for dealing with various problems of spatial pattern recognition. Consequently, the neural network technology is being more and more widely used in the research of remote sensing image classification.Back propagation (BP) neural network model and Radial basis function (RBF) neural network model are often mentioned in the application and research of neural network, compared with other neural network algorithms. According to the research condition that using BP and RBF network for classification of remote sensing image in our country, there are two mainly technical directions in this field. One way is that making use of the language provided by professional software of remote sensing processing for the secondary development of neural network algorithm, the other way is by combining Matlab neural network toolbox with professional software of remote sensing processing to classify. However, both of these two kinds of method have its disadvantages. The former is difficult for abecedarian to implement this function, so it is bad for the beginner to study and research the performance of neural network. Although the second method is easy to grasp, it is inconvenience for people to manage the Remote sensing data.Above all, in allusion to shortcomings of the front two methods, the processes of using remote sensing image data to classify were simulated. The separating degree of feature vectors, linear relationship and Probability density were combined to structure16groups of different simulation data. This paper researched the input and output characteristics of BP and RBF network and analyzed characteristic vector of different conditions influence on the classification performance of the two networks. The experiment also explained why two network and fuzzy logic can be used at the same time, which provides experiment basis for the subsequent construction of Matlab classification system. In the second place, on the basis of the research about the input and output characteristics of the two networks, this paper realized supervision classification of the remote sensing image on Matlab platform with the artificial neural network toolbox. The classification system included the calculation of features component, the constitution of feature vectors, the selection of multiple polygon training area, the construction of the neural network classifier, the processing of Classification results and the evaluation of classification accuracy. The great advantages of the programming of system is adopting the modularization thought, consequently, classifier module, feature extraction module and classification accuracy evaluation module can be replaced according to the actual needs of classification. At the last, the paper used the classification system developed by author to compare the performance of BP and RBF network classification whose research object is the TM images and its spectrum characteristic data. These experiments prove this system which can be used for classification of hyperspectral image data is effective.In addition, besides discussing the effect of the classification influenced by the artificial neural network algorithm parameters, this paper put feature selection and the performance of classifier together to discuss. Adding some auxiliary band features to the feature vectors which were comprised of targeted characteristics to stabilize the performance of classifier.Finally, this paper analyzed the probability density of the date caused by the original data which was put into the two networks and the classification effect of the two networks. Analysis of the probability density of the sorted date can initially determined whether the gray image that is the output of the network can use the fuzzy reasoning to get ideal classification results, which reduce workload to some extent.