Terrain Classification For PolSAR Image Classification Based On Deep Curvelet-Residual Neural Network

Polarized synthetic aperture radar has become one of the indispensable objects in the development of synthetic aperture radar at home and abroad.In the interpretation of SAR images,the classification of polarimetric SAR image features is a crucial research direction.The target polarization information acquired by a resolution unit determines the category to which the pixel belongs.Compared with shallow networks,deep network models have prominent advantages in the representation of complex high-variation functions,network computational complexity,and effective information acquisition.Therefore,in this paper,deep-learning methods are used to classify polarimetric SAR imagery.Thesis mainly includes the following aspects:1.A method of classification of polarized SAR image features based on deep residual network is proposed.Polarimetric SAR imagery feature classification is a key step in image processing and requires determining the type of each pixel in the imagery.Effective feature learning is the basis for solving classification tasks.Therefore,in order to learn more abundant features,this paper first uses polarization,rotation,and other means for polarimetric SAR images to amplify the data,and then classifies them through a deep residual network.The data amplification enriches the characteristic expression form of polarimetric SAR images,and provides sufficient amount of data for the deep residual network to exploit higher-level abstract features of polarimetric SAR images.The method uses the deep residual network to build a short-circuit structure to obtain a higher nonlinear representation learning ability,effectively learns shallow and deep features,and selects 5 to 8 percent of labeled samples of polar SAR images as training samples.The use of supervised learning to achieve the classification of polarographic SAR images is compared with the typical convolutional network model of the moment,and this method obtains better classification results.2.A polarization SAR image feature classification method based on depth curve wave-residual network is proposed.Polarimetric SAR images are rich in texture edge features.Curvelet transform is performed on them.Threshold operation can reduce the interference of noise on the classification process,and multi-scale information such as direction,angle and space can be extracted to ensure the complete extraction of deep residual network.With multi-scale information,this method can better identify the edge information in the image.Compared with other convolutional network models,it improves the accuracy of the classification of polarimetric SAR images.3.A method of classification of polarographic SAR image features based on depth-expanding convolution and residual network is proposed.In this paper,we use the idea of feature combination to input polarimetric SAR images after Pauli decomposition and after Curvelet transform into different channels of two-channel combined depth residual network.On the one hand,we extract the physical scattering characteristics of polarimetric SAR images.On the other hand,multi-scale information such as the direction,angle,and scale of polarimetric SAR images is input into the model,and noise is effectively suppressed.In addition,an extended convolution operation is introduced in the first residual module of this method to increase the receptive field.Without losing the effect of data information,ensure that each convolution output contains a large range of information for polarimetric SAR images,achieving macroscopic and global results in microscopic operations,reducing the loss of features,and comparing with the above methods to further optimize The result of the classification of polarized SAR imagery features has improved the classification accuracy.