Research On The Application Of Deep Learning In SAR Image Classification

As a kind of active high-resolution microwave imaging radar,Synthetic aperture radar(S AR)has the ability to observe through clouds and rain with all-day and all-weather.As the key link of SAR image interpretation technology,SAR image classification has been paid more and more attention,and has a wide application prospect in the fields of urban planning,military reconnaissance and emergency disaster.Deep learning is a hot topic in the field of machine learning,which has been applied to image classification and recognition,compared with shallow learning which can extract the deep features from data.Therefore,it is of great theoretical significance and practical value to study the classification of SAR images based on deep learning.For the SAR image classification is severely affected by speckle noise,this paper analyzes the imaging mechanism of SAR image and the basic theory of deep learning deeply,and studies the key technology of applying deep learning to SAR image classification.The main research work of this paper is as follows:(1)A SAR image classification method based on GLCM-GMRF texture feature and depth confidence network(DBN)is proposed.This paper introduce the texture feature of the image as a priori information,reflect the spatial relationship between pixels and the unique characteristics of different types.This method uses gray level co-occurrence matrix(GLCM)to extract the spatial gray correlation feature,and it further uses gaussian markov random fields(GMRF)to construct the statistical dependence of neighborhood pixels,and the extracted GLCM-GMRF texture features are input to DBN along with the intensity vectors.Radarsat-2 data is used to validate the classification performance of the proposed method,compared with traditional classification methods based on DBN and SVM,the proposed method is more effective for noise suppression and get better classification results.(2)A SAR image classification method based on deep multi-feature fusion is proposed.Firstly,the method extracts the image features by convolution neural network(cnn),then extracts the GLCM,GMRF and Gabor texture features of the image,and then normalized the extracted features and combines them into eigenvectors.Finally,the feature vectors are fed into the DBN to achieve the deep fusion feature,and the classification results are obtained by the classifier.Radarsat-2 data is used to validate the classification performance of the proposed method,compared with the methods based on CNN and SVM,the proposed method can achieve more accurate classification of SAR images by the deep fusion feature.