The Key Technology Research On Three Important Neural Networks In Remote Sensing Image Classification

Remote sensing image classification is an important part of remote sensing information processing.The choice of classifier directly determines the accuracy of image classification.The traditional classifier based on statistical analysis has error classification.It has been unable to meet the current classification requirements of high spatial resolution,and high spectral resolution remote sensing image.In addition,it is usually ineffective for increase of classification through improving statistical classification classifier.Therefore,a new classifier is urgently needed in order to meet require of the high classification accuracy.As a kind of flexible and intelligent classifier,the artificial neural network has become a hot research content in the field of remote sensing image classification in recent years,because of its good self-learning,self-organization and adaptive ability.In order to make better application of artificial neural network in remote sensing image classification,based on the existing neural network research,in-depth study of network algorithm and performance.Optimization of artificial neural network,has become an important research direction of artificial neural network.There are many kinds of artificial neural networks presently.This paper chooses BP neural network,wavelet neural network(WNN)and fuzzy adaptive resonance theory neural network(Fuzzy ARTMAP)which are the most representative neural networks as the research object.The structure,algorithm and training process of three kinds of neural networks are introduced and discussed.According to the performance of three kinds of neural networks in remote sensing image classification,the key technology of neural network is studied.The BP neural network is optimized by the gradient type learning algorithm,and the FR conjugate gradient method with adaptive adjusting step size BP neural network is obtained.Wavelet neural network is studied in three aspects: the choice of wavelet basis function,the choice of error function and coding mode.ARTMAP Fuzzy neural network is studied on the selection of the vigilance parameter.The rule of selecting the best warning line parameters is obtained according to the features of remote sensing image.By using seven sets of experimental data,which include the hyperspectral image,multispectral image,high-resolution image and SAR image.The results show that the convergence rate is faster,and the classification accuracy is higher in the FR conjugate gradient method BP neural network,compared with the traditional steepest descent method BP neural network.Wavelet neural network composed of DOG wavelet function,NB entropy error function and moderate saturated value encoding mode is better than that of the traditional wavelet neural network in order to get the proper classification,and improve the classification accuracy.If the feature class is small,and the difference between different categories is large,the vigilance parameter should take smaller value.If the feature class is big,and the difference between different categories is small,the vigilance parameter should take bigger value.The classification effect of ARTMAP Fuzzy neural network is better than that of wavelet neural network,and the classification effect of wavelet neural network is better than that of BP neural network.