Application Research And Develop On Remote Sensing Image Classification Based On Artificial Neural Network

Remote sensing(RS) image classification is always a pivotal part of remotesensing study. How to improve the accuracy of RS interpretation is a urgent problemin RS application. In recent years, with the development of the theory about artificialneural network, the neural network technology is becoming increasingly an effectivemeans of classification processing of remote sensing images. Compared withconventional statistic classifier, the artificial neural network (ANN) has beendeveloped and applied to remote sensing data classification problem, which doesn'tneed suppose parameterized distribution of sample space in advance. ANN hascomplicated mapping capability. The back propagation neural network model (BPmodel) is often been used. This reasearch use BP model for the remote sensing classification. Based on theanalysis of the BP model, in order to increase the possibility of artificial neuralnetwork convergence and accelerate the training speed, some improved methods arepresented ,such as dynamic study ratio, dynamic K value of the sigmoid function andincreasing momentum. Useing these methods, we develop a relative software byVisual Basic, and develop a relative plug program in ENVI by IDL/ENVI. We explorea new practice approach based on ENVI by IDL (Interactive Data Language). Useingthe image of capital airplane as the main data source, the classification results byartificial neural network are compared with those gained by maximum likelihoodmethod. The analysis result shows that this method can distinguish the matter andfeatures that can not be distinguished by maximum likelihood method, it prove that itcan improve the classification precision. Furthermore, as to the mix pixels, weconstruct a BP neural network which the nodes of input layer are the bands of remotesensing and the nodes of output layer are percent of several kinds of object. This is apractice method in the integrative research of the classification and the mixed pixels. Furthermore, in this research, some classification datails are optimized, such asband combination, distilling training data, and so on. The classification flow isconsidered as a whole. It does not only emphasize the model or the thought. It provethat it provide the guarantee for the classification precision by applying the optimizedclassification flow.