SAR Remote Sensing Image Classification Based On Deep Learning

Synthetic aperture radar(SAR)is an active microwave remote sensing technique that can capture high-resolution images in all weather and all time.It has been widely applied in many fields such as military target detection,marine environmental monitoring,and urban coverage survey.SAR remote sensing image classification is a fundamental step for SAR image inter?pretation,which aims to classify the pixels with different categories based on the backscattering characteristics and other information.However,there are some problems extremely hinder the classification of SAR images,which are speckle noise contained in the images,ambiguous and unstable characteristics of targets in SAR data,and non-uniform characteristic of different re-gions in SAR images.Therefore,it is still an urgent issue to automatically achieve high precision SAR image classification.To realize the classification of SAR remote sensing image,combined with the characteristics of SAR images,this dissertation studies the feature extraction method and classification method based on deep learning neural network.The major contributions of this dissertation are indicated as follows:Firstly,inherent speckle noise in SAR images leads to the absence of effective feature rep-resentation.To deal with this issue,deep convolutional auto-encoders are proposed to extract the potential spatial features of SAR images.The deep network is composed of a convolutional layer based on traditional filters to obtain initial features,a scale transformation layer to remove speck-le noise,multi-layers based on sparse auto-encoders to enhance the discrimination of features,and a postprocessing layer to overcome the influence of speckle noise to the classification results.In order to improve classification accuracy with limited training samples,weighted fusion-based representation classifiers are proposed to balance the "competitive" and "collaborative" rep-resentation of the test sample on the training dictionary,which can further improve the SAR image classification performance.Secondly,characteristics of targets in SAR images are ambiguous and unstable.To over-come this problem,deep supervised and contractive neural network is proposed to capture the relevant information between features and labels,which is able to enhance the discrimination and robustness of feature representation.The deep model develops the supervised penalty with training labels to add high-level semantic information,which aims to overcome the ambiguous characteristic of targets;at the same time,the contractive restriction in the deep model is adopt?ed to enhance the locally invariant of the encoding representation,which prefers to overcome the unstable characteristic of targets.In order to extract abundant spatial features,a multiscale patch-based feature extraction model is developed to obtain low-level and mid-level features from SAR images.In order to further restrain the influence of speckle noise,a graph-cut-based spatial regularization model is proposed after classification to suppress misclassified pixels.Thirdly,non-uniform characteristic of different regions in SAR images leads to weak gen-eralization ability of traditional feature extraction models.To solve this issue,deep recurrent encoding neural networks are proposed to automatically extract features and conduct classifi-cation,which can enhance the generalization ability of feature extraction.The deep network consists of two types of deep modules:patch-based recurrent neural networks and nonnegative and Fisher constrained auto-encoders.Patch-based recurrent neural networks are developed to learn spatial dependencies of SAR images,which prefer to discover intrinsical correlations a-mong pixels and extract contextual features automatically.Nonnegative and Fisher constrained auto-encoders develop nonnegative constraint to improve the sparsity of features,and propose Fisher constraint to increase intraclass compactness of features,which can enhance the discrim-inative representation ability of features and improve the final classification accuracies.Finally,marine floating raft extraction based on SAR images is studied in this dissertation.In order to solve the issue of more serious speckle noise in marine SAR images,an iterative low-rank representation algorithm is developed for SAR image despeckling.It not only effectively removes speckle noise of marine SAR images,but also increases the visual contour informa-tion of floating raft.In order to conduct marine floating raft extraction,the three deep neural networks proposed in this dissertation are adopted to demonstrate the effectiveness in practical applications.In summary,through analyzing the characteristics of SAR images,this dissertation re-searches SAR image classification models based on deep learning neural network,which can extract abundant and effective features,and achieve high precision classification automatically.The proposed approaches are developed for marine floating raft extraction,and produce state-of-the-art results,which can afford effective methods for intelligent interpretation of SAR remote sensing images.