Reaserch On Fine Classification Of Polarimetric SAR Images Based On Convolution Neural Network

Polarimetric synthetic aperture radar(Pol SAR)can obtain multi-channel polarization images.It can improve the ability of target detection,recognition and classification.In recent years,the Pol SAR plays a more and more important role in the application of target detection and the terrain classification.The classification of Pol SAR images is to classify the obtained polarization data and get different information,which make the decision of the computer close to the real object category.In recent years,there are a lot of Pol SAR image classification methods,but these traditional classification methods have a common character,that is,the feature extraction is usually carried out,that means more human intervention process has been added.Convolutional neural network(CNN)is a new type of neural network based on multi-layer supervised learning.It has been widely used in image processing field in recent years.It can transform the original image data into more abstract expression through many simple nonlinear models.Compared with traditional classification algorithm,convolutional neural network do feature extraction without too much manual participation.Based on the basic theories,this paper applies the convolution neural network classification method to the classification of Pol SAR image,and realizes the fine classification of high precision ground objects with Pol SAR images.In this paper,through the study of Pol SAR based feature extraction method and convolution neural network the basic idea,structure and working principle,and combined with the training process of convolution neural network,the main steps of applying convolution neural network to the classification of Pol SAR image objects are proposed,and the design of convolution neural network for Pol SAR image classification and the optimization of classification methods are carried out.Besides,the classification based on the Pol SAR image data and the selected characteristics are classified respectively.In addition,in order to make full use of the multichannel characteristics of Pol SAR data,based on the traditional 2D convolution neural network,3D convolution neural network is introduced.And aiming at the small sample size problem of some objects,the method of virtual sample expansion is proposed to improve the result.In this paper,three groups of Pol SAR images with different resolutions are used to carry out the experiments.The results show that convolution neural network can significantly improve the classification effect of Pol SAR images.Moreover,3D convolution neural network have better classification performance than 2D convolution neural network by making full use of the information of Pol SAR data.In addition,compared with the extracted features,convolution neural network can make full use of Pol SAR raw data for feature extraction and classification,and the classification results are more accurate and efficient.The sample expansion method can improve the classification effect of the objects with insufficient samples.Finally,comparing with the results of BP neural network classification,the superiority of convolution neural network to fine classification of Pol SAR images is verified.